Flood Studies Update Technical Research Report Volume II ... › data › files › Technical... ·...
Transcript of Flood Studies Update Technical Research Report Volume II ... › data › files › Technical... ·...
Flood Studies Update
Technical Research Report
Volume I Rainfall Frequency
Volume II Flood Frequency Estimation
Volume III Hydrograph Analysis
Volume IV Physical Catchment Descriptors
Volume V River Basin Modelling
Volume VI Urbanised and Small Catchments
Volume II
Flood Frequency Estimation Conor Murphy, Conleth Cunnane, Samiran Das and Uzzal Mandal
Derived from Technical Research Reports by
NUI Galway and NUI Maynooth
Volume II Flood Frequency Estimation
ii
Abstract
Flood frequency estimation addresses the issue of flood risk assessment required in flood
zoning and spatial planning, and in the specification of flow values for the design of flood
alleviation and control works.
This volume describes flood frequency estimation research undertaken at NUI Galway and
NUI Maynooth and provides a statistical summary of flood flows in Ireland taken from the
data archives of the OPW, EPA and ESB. The research is based on analysis of annual
maximum flow records at some 200 river flow gauging sites. The records analysed range
from eight to 55 years in length. More detailed research has been carried out on subsets of
110 gauged sites for which the quality of recorded flows is judged the most reliable.
Most floods occur during the winter half of the year, with some notable floods also occurring
during summer, especially in August. Because of Ireland’s humid climate, the year-to-year
variation of flood flow values (indexed by the coefficient of variation) is typically quite small
by international standards. The so-called skewness of the flood series is also modest. While
no single statistical distribution can be considered “best” at all locations, it has been found
that the EV1 (Gumbel) and lognormal distributions provide a reasonable model for the
majority of stations.
Guidance is provided on the estimation of the design flood of required annual exceedance
probability at both gauged and ungauged locations and on how to express the uncertainty in
the resulting estimates. The use of growth curves or growth factors based on data pooled
from a group of sites is generally advocated, with the use of suitable 3-parameter
distributions recommended in many applications.
A procedure is presented for estimating the so-called index flood (QMED) at an ungauged
site. QMED estimation is especially important in Ireland because flood growth rates are
generally mild. The recommended procedure for QMED estimation at ungauged sites is to
transfer information from a nearby site, ideally one upstream or downstream of the site of
interest. The user must apply experience and technique to select the so-called pivotal gauged
catchment from which to make the transfer.
Examples presented show a range of difficult cases that can arise in practice. In some cases,
the user is advised to consider single-site estimation in addition to – or in combination with –
the generally recommended pooling method.
The volume concludes that flood estimation cannot be reduced to a formula-based procedure.
Individual analysts must make choices which reflect the circumstances of the problem, the
available flow data and their own knowledge and experience.
©Office of Public Works 2014
Volume II Flood Frequency Estimation
iii
Further information about the research
FSU Technical Research Reports (TRRs) are available in their original
form for researchers and practitioners who seek additional information
about a method. The original TRRs sometimes document exhaustive
application of a method to many catchments. In others, additional options
are reported.
Inevitably, the relevance of the original TRRs is influenced by OPW
decisions on which methods to implement, and how best to arrange and
support them. Readers who consult the original TRRs will notice editorial
re-arrangements and compressions, and occasional changes in notation and
terminology. These were judged necessary to enhance understanding and
use of the FSU methods amongst general practitioners. More significant
changes are labelled explicitly as editorial notes.
Volume II Flood Frequency Estimation
iv
Contents
Abstract ii
Contents iv
Notation xiv
Lower case symbols xiv
Upper case symbols xiv
Subscripts xv
Abbreviations and descriptor names xv
Glossary of terms xvii
1 Introduction and data 1
1.1 Why we estimate flood frequency 1
1.2 Flood data 2
1.2.1 Introduction 2
1.2.2 Annual maximum series 2
1.2.3 FSU flood datasets 2
1.2.4 Expressions of flood rarity 3
1.2.5 Index flood 3
1.3 Historical floods 4
1.4 Structure of volume 4
1.5 Material in common with rainfall frequency studies 5
1.5.1 Annual maximum and peaks-over-threshold series 5
1.5.2 Return period 6
1.5.3 Average recurrence interval 6
1.5.4 Langbein’s formula 7
2 Estimation of the index flood, QMED 8
2.1 Exploratory data analysis 8
2.1.1 Introduction to the datasets 8
2.1.2 Adjustments for period-of-record effects 10
2.1.3 Physical catchment descriptors 12
2.1.4 Rank correlations 14
2.1.5 Principal component analysis 15
2.1.6 Correlations and competing variables 16
2.1.7 Scatter-plots and summary information for selected PCDs 17
2.2 Rural-catchment model for estimating QMED from PCDs 18
2.2.1 Regression methods 18
2.2.2 Alternative methods 20
2.2.3 Selection of catchments for calibration and validation 20
2.2.4 Selecting PCDs 21
2.2.5 Choosing a model 22
2.2.6 Model performance 23
2.2.7 Checking for logical consistency 25
2.2.8 Checking and investigating the model residuals 25
2.2.9 Validation of model performance 28
2.3 Assessing model robustness 29
Volume II Flood Frequency Estimation
v
2.3.1 Model robustness 29
2.3.2 Bootstrap resampling 29
2.3.3 Model robustness to influential stations 30
2.3.4 Robustness to spatial drift in model coefficients 31
2.4 Investigating the effect of arterial drainage 32
2.4.1 Approach 32
2.4.2 Analysis of stations for which both pre and post-drainage records held 32
2.4.3 Modelling QMED on drained and undrained catchments separately 34
2.5 Adjusting the selected model for urbanisation 34
2.5.1 Introduction 34
2.5.2 Deviation of QMED on urbanised catchments from rural model 35
2.5.3 A special check 36
2.5.4 Approach to devising an urban adjustment 36
2.5.5 Exploratory data analysis 37
2.5.6 An urban adjustment model 37
2.5.7 Performance of urban adjustment model 38
2.5.8 Comparisons and contrasts with the FEH 39
2.6 Improving model performance by data transfer 39
2.6.1 Introduction 39
2.6.2 Review of techniques 39
2.6.3 Geostatistical mapping of residuals 40
2.6.4 Possible disadvantages of automated methods 42
2.6.5 Recommended procedure for data transfer 43
2.7 Worked example of QMED estimation at an ungauged site 43
2.7.1 Illustrative example 43
2.7.2 Merging data transfers from two sites 46
2.7.3 Geostatistical mapping method 46
3 Trend and randomness 47
3.1 Tests 47
3.1.1 Methods 47
3.1.2 Formats 48
3.2 Findings 48
3.3 Pragmatism 49
4 Descriptive statistics – and inferences therefrom 50
4.1 Descriptive statistics 50
4.1.1 Introduction 50
4.1.2 Summary statistics – the idea 50
4.1.3 Summary statistics based on moments and L-moments 50
4.1.4 Additional summary statistics 51
4.2 Summary statistics for Irish flood data 52
4.2.1 Summary statistics for 181 FSU stations 52
4.2.2 Variability and skewness of Irish flood data 58
4.2.3 Comparisons of CV with L-CV and of H-skew with L-skew 59
4.3 Geographical traits 61
4.4 Preliminary distribution choice from skewness v. record length plot 63
4.5 Preliminary distribution choice aided by L-moment ratio diagrams 66
4.5.1 L-moment ratio diagrams 66
Volume II Flood Frequency Estimation
vi
4.5.2 Implication for choice of 2-parameter distribution 66
4.5.3 Implication for choice of 3-parameter distribution 68
4.6 Exceptional recorded values and outliers 68
4.6.1 Qmax/Qmean ratios 68
4.6.2 High outliers 69
4.6.3 Low outliers 70
5 Probability plots … and inferences from them 71
5.1 Introduction 71
5.1.1 The idea of a probability plot 71
5.1.2 Synthetic examples 71
5.1.3 Standard plots for the FSU flood peak data 71
5.1.4 Approach taken to assessing plots 74
5.2 Linear patterns 74
5.3 Curve patterns 74
5.3.1 Assignment of patterns 74
5.3.2 Probability plots for 43 Grade A1 stations 75
5.3.3 Probability plots for 110 Grade A1+ A2 stations 76
5.3.4 Curve pattern in relation to skewness coefficient 77
5.4 Flood volumes associated with largest peaks on convex probability plots 79
5.4.1 Stations with a convex curve pattern 79
5.4.2 Hypothesis 79
5.4.3 Arrangement of study 79
5.4.4 Calculation of hydrograph volumes 80
5.4.5 Example 80
5.4.6 Summary of findings 82
5.4.7 A further check on the hypothesis 82
5.5 Flood seasonality 83
5.5.1 Seasonality of annual maximum floods 83
5.5.2 Seasonality of largest floods 84
5.5.3 Circular diagrams 84
5.6 Flood statistics on some rivers with multiple gauges 85
5.6.1 Down-river growth in QMED 85
5.6.2 Down-river variation in probability plots 85
6 Determining T-year flood magnitude QT by index flood method 87
6.1 Introduction 87
6.2 Regional flood frequency analysis and the index flood approach 88
6.2.1 Index flood approach 88
6.2.2 Type of region for pooling flood data 88
6.2.3 Choice of index flood 88
6.2.4 A 2-stage approach 89
7 Flood growth curve estimation 90
7.1 Introduction to xT 90
7.2 Single-site and pooled estimates of QT 90
7.2.1 Advantages and drawbacks of the two approaches 90
7.2.2 Choice of distribution 90
7.3 Pooling groups 91
7.3.1 The idea of pooling 91
Volume II Flood Frequency Estimation
vii
7.3.2 Construction of pooling groups 91
7.3.3 The 5T rule 92
7.3.4 Recommended metric for pooling-group construction 92
7.4 Growth curve estimation 93
7.4.1 Pooled L-moment ratios 93
7.4.2 Fitting growth curve distributions by the method of L-moments 93
7.4.3 Growth curves for particular distributions 94
8 Effect of catchment type and period of record on xT and QT 97
8.1 Data screening 97
8.1.1 Discordancy 97
8.1.2 Characteristics of the three discordant stations 99
8.1.3 Other heavily urbanised catchments 102
8.2 Effect of catchment type on pooled growth curve estimates 103
8.3 Temporal effect on pooled growth curve estimates 106
8.3.1 Periods of record considered 106
8.3.2 Summary results for 90 stations 106
8.3.3 Individual results for three decades (50 stations) 106
8.3.4 Individual results for five decades (26 stations) 108
8.3.5 Individual results for a different set of five decades 109
8.3.6 Outcome 109
8.4 Arterial drainage effect on pooled growth curve estimates 110
8.5 Implications for flood frequency estimation 111
8.5.1 Implications for pooling-group formation 111
8.5.2 Respecting recent flood data 112
9 Uncertainty estimation 112
9.1 Standard errors – an introduction 112
9.2 Standard error of QMED estimation from gauged flood data 113
9.3 A comparison with FEH methods 114
9.4 Standard error of QT in single-site estimation 115
9.4.1 Method 115
9.4.1 Relative standard error 115
9.4.2 Relative standard errors under the EV1 assumption 115
9.4.3 Relative standard errors under the GEV assumption 116
9.5 Standard error of pooled estimate of xT and of QT 117
9.5.1 Simulation method 117
9.5.2 Results based on EV1 simulations 118
9.5.3 Results based on GEV simulations with k = -0.1 119
9.5.4 Summary 120
9.6 Standard error of QT based on PCD estimate of QMED and pooled xT 121
10 Guidelines for determining QT 123
10.1 Introduction 123
10.1.1 Single-site or pooled analysis? 123
10.1.2 Probability plots 123
10.1.3 Factors to be borne in mind 123
10.2 Determining QT by single-site analysis 124
10.2.1 General guidance 124
10.2.2 Plotting positions 124
Volume II Flood Frequency Estimation
viii
10.2.3 Parameter estimation for EV1 distribution 125
10.2.4 Parameter estimation for LN2 distribution 126
10.3 Determining QT from pooled data 126
10.3.1 When flood data are available at the subject site 126
10.3.2 When the subject site is ungauged 127
10.4 Characteristic examples of probability plot behaviour 127
10.4.1 Good straight-line behaviour 127
10.4.2 Good straight-line behaviour but single-site and pooled disagree 128
10.4.3 Concave upwards behaviour with outlier 129
10.4.4 Convex behaviour 131
10.4.5 Unclear behaviour with extreme outlier 132
10.4.6 Irregular behaviour 133
10.5 Additional notes on the choice of distribution and method 134
10.5.1 Problems in the use of 3-parameter distributions for single-site analysis 134
10.5.2 Reconciling single-site and pooled analyses 135
10.5.3 Probability associated with a very large recorded flood 135
10.5.4 Flood growth curves with an upper bound 136
11 Data transfers revisited 137
11.1 Interim assessment of QMED data transfers 137
11.1.1 Subject sites used in the interim assessment 137
11.1.2 Methods used in the interim assessment 138
11.1.3 Results 139
11.1.4 Remarks 141
11.2 Further guidance on pivotal catchments and data transfers 141
12 Summary and conclusions 143
12.1 Data 143
12.2 Descriptive statistics 143
12.3 Seasonal analysis 143
12.4 Estimation of design flood 144
Acknowledgements 146
References 146
Appendices 151
Appendix A Review of stage-discharge relationships 151
A1 Terminology 151
A2 Review of stage-discharge relationships 151
A2.1 General form of stage-discharge relationships 151
A2.2 Analysis tools and background information 152
A2.3 Review procedure 152
A2.4 Gauging station surveys 153
A3 Gauging station classification 154
A3.1 Initial site classification 154
A3.2 FSU station classification 154
A3.3 Uncertainty analysis 155
A4 Production of annual maximum flood series 156
Appendix B Flood data exclusions 157
B1 Stations omitted from the QMED modelling research 157
Volume II Flood Frequency Estimation
ix
B2 Other notes on stations omitted from research 158
B3 General recommendation 158
Appendix C Notes on the regression methods used in Chapter 2 159
C1 Ordinary least-squares (OLS) regression 159
C2 Weighted least-squares (WLS) regression 159
C3 Generalised least-squares (GLS) regression 160
C4 Geographically-weighted regression (GWR) 161
Appendix D QMED models for drained/undrained catchments 164
D1 Partitioned approach 164
D2 Undrained catchments 164
D3 Drained catchments 165
D4 Choosing a general purpose model 166
Appendix E Trend analysis of AM flows in Irish rivers 168
E1 Introduction 168
E2 Purpose 168
E2.1 Importance of testing whether hydrological processes are stationary 168
E2.2 Types of change 169
E3 Procedure 169
E3.1 Steps in the analysis 169
E3.2 The idea of exploratory data analysis 170
E3.3 Hypothesis testing 170
E3.4 Use of p-values 171
E4 Statistical tests 171
E4.1 Parametric and non-parametric tests 171
E4.2 Null hypotheses 172
E4.3 Tests adopted 172
E4.4 Mann-Kendall test 173
E4.5 Spearman’s ρ test 174
E4.6 Mean-weighted linear regression test 175
E4.7 Mann-Whitney U test 175
E4.8 Turning points test (Kendall’s test) 176
E4.9 Rank difference test (Meacham test) 176
E5 Exploratory data analysis 176
E5.1 Selection of data 177
E5.2 Stations showing trend 177
E5.3 Stations with pre and post-drainage records 177
E6 Trend analysis and results 181
E6.1 Formats of flood data tested for trend 181
E6.2 Main test results 181
E6.3 Non-randomness 184
E6.4 Trend 184
E6.5 Discussion 186
E7 Summary 186
E8 Critical values for mean-weighted linear-regression test 187
Appendix F Additional summary statistics 189
Appendix G Sample probability plots and summary information 194
Volume II Flood Frequency Estimation
x
Appendix H Probability-plot linear scores and curve patterns 195
Appendix J Flood volumes in relation to convex EV1 plots 199
J1 Flood peaks and volumes for Station 07006 Moynalty at Fyanstown 199
J2 Flood peaks and volumes for Station 15003 Dinan at Dinan Bridge 201
J3 Flood peaks and volumes for Station 16008 Suir at New Bridge 203
J4 Flood peaks and volumes for Station 16009 Suir at Cahir Park 205
J5 Flood peaks and volumes for Station 24013 Deel at Rathkeale 207
J6 Flood peaks and volumes for Station 24082 Maigue at Islandmore 209
J7 Flood peaks and volumes for Station 25017 Shannon at Banagher 211
J8 Flood peaks and volumes for Station 25021 Little Brosna at Croghan 213
Appendix K Seasonal distribution of annual maximum floods 215
Appendix L Distance metrics for pooling-group construction 217
L1 Introduction 217
L2 Notation 217
L3 Selecting variables to define the distance metric 217
L4 Statistics to help in choosing a good distance metric 218
L5 Pooled uncertainty measure, PUM 218
L6 Application to flood data at 90 A1 + A2 stations 218
L7 Discussion 220
L8 Alternative weightings of the recommended distance metric 220
Maps
Map 2.1: Spatial distribution of 205 catchments 11 Map 2.2: Factorial error in QMED estimated by selected rural model 26 Map 2.3: IDW-interpolated residuals from rural ℓnQMED model 41 Map 4.1: Specific QMED (m
3s
-1/km
2) for 176 A1 + A2 + B stations 61
Map 4.2: CV at 110 A1 + A2 stations 62 Map 4.3: Hazen skewness at 110 A1 + A2 stations 63 Map 8.1: Location of the 85 stations within four geographical regions 104
Map C.1: Spatial variation in the FARL coefficient as interpolated from GWR 163
Figures
Figure 2.1: Number of stations in each quality category (full dataset of 205 stations) 9 Figure 2.2: QMED adjusted for period-of-record effects 14 Figure 2.3: Annual maximum flow series for Station 09035 Cammock at Killeen Road 14 Figure 2.4: Relationship between PCDs summarising extent of arterial drainage 17 Figure 2.5: Association of ℓnQMED and selected PCDs for 205 catchments 18 Figure 2.6: Improvement of r
2 for a model size of one to nine variables 23
Figure 2.7: Observed and modelled QMED for the 145 calibration stations 24 Figure 2.8: Diagnostic plots of ℓnQMED model performance 25 Figure 2.9: Residuals versus individual PCDs for rural ℓnQMED model 27 Figure 2.10: Observed and modelled QMED for the 25 validation stations 28 Figure 2.11: Normal quantile-quantile plots of the bootstrapped model coefficients 30 Figure 2.12: Influence of individual stations in determining model coefficients 31 Figure 2.13: QMED for pre and post-drainage records at 15 stations 33 Figures 2.14: QMED change following drainage (against ARTDRAIN and ARTDRAIN2) 33 Figure 2.15: QMED change following arterial drainage (against BFIsoil) 34 Figure 2.16: Performance of rural QMED model on urbanised catchments 35 Figure 2.17: Correlations between ℓnUAF and selected PCDs 37 Figure 2.18: Performance of UAF-adjustment to QMED 38
Volume II Flood Frequency Estimation
xi
Figure 4.1: Relationship between L-CV and CV for 110 A1 + A2 stations 60 Figure 4.2: Relationship between L-skew and H-skew for 110 A1 + A2 stations 60 Figure 4.3: Skewness at 110 A1 + A2 stations versus record length 64 Figure 4.4: L-skewness against record length for 110 A1 + A2 gauging station 65 Figure 4.5: L-moment ratio diagram for annual maximum floods 67 Figure 4.6: Comparison with equivalent samples drawn from particular distributions 67 Figure 5.1: Relative frequency of curve patterns – 43 Grade A1 stations 76 Figure 5.2: Relative frequency of curve patterns – 110 Grade A1 + A2 stations 76 Figure 5.3: Hydrographs of four largest AM events, Station 24082 Maigue at Islandmore 81 Figure 5.4: Hydrograph volumes of four largest AM events at 24082 Maigue at Islandmore 82 Figure 5.5: FAI v. Hazen skewness, labelled by EV1 curve pattern (110 A1 + A2 stations) 83 Figure 5.6: Seasonality and magnitudes of 6969 AM floods (202 A1 + A2 + B stations) 84 Figure 5.7: Variation of QMED down the Barrow, Suir and Suck 85 Figure 6.1: Probability density function and cumulative distribution function 87 Figure 8.1: AM flow series for three discordant sites (Stations 08005, 09002 and 09010) 101 Figure 8.2: Decadal estimates for 50 stations (arranged in station-number order) 107 Figure 8.3: Decadal estimates for 26 stations (arranged in station-number order) 108 Figure 8.4: Comparison of pre and post-drainage flood statistics at 16 stations 111 Figure 11.1: Performance at Station 16002 Suir at Beakstown (512 km
2) 139
Figure 11.2: Performance at Station 16008 Suir at Newbridge (1120 km2) 139
Figure 11.3: Performance at Station 18050 Blackwater at Duarrigle (244.6 km2) 140
Figure 11.4: Performance at Station 24008 Maigue at Castleroberts (805 km2) 140
Figure 11.5: Performance at Station 26002 Suck at Rookwood (626 km2) 140
Figure C.1: Fitted model for inter-site correlation 160
Figure D.1: Performance achieved with undrained model (on 95 undrained catchments) 165 Figure D.2: Performance achieved with drained model (on 50 drained catchments) 166
Figure E.1: Stations for which AM flow series shows significant trend 179 Figure E.2: AM flow series for stations with arterial drainage works during record 180 Figure E.3: AM flow series showing likely influence of arterial drainage 185
Figure J.1: Hydrograph volumes for Station 07006 Rank 1, 2 and 4 AM flood peaks 199 Figure J.2: Hydrograph volumes for Station 15003 Rank 1, 2 and 4 AM flood peaks 201 Figure J.3: Hydrograph volumes for four largest Station 16008 AM flood peaks 203 Figure J.4: Hydrograph volumes for four of five largest Station 16009 AM flood peaks 205 Figure J.5: Hydrograph volumes for four largest Station 24013 AM flood peaks 207 Figure J.6: Hydrograph volumes for four largest Station 24082 AM flood peaks 209 Figure J.7: Hydrograph volumes for four of five largest Station 25017 AM flood peaks 211 Figure J.8: Hydrograph volumes for four largest Station 25021 AM flood peaks 213
Histograms
Histogram 2.1: Length of available record for QMED estimation (205 stations) 9 Histogram 2.2: QMED values across full dataset of 205 stations 10 Histograms 2.3: Catchment sizes for calibration and validation datasets 21 Histograms 4.1: CV and L-CV at Grade A1 and A2 stations 59 Histograms 4.2: Hazen skewness and L-skewness at Grade A1 and A2 stations 59 Histogram 5.1: Seasonal occurrence of annual maximum floods 83 Histogram 5.2: Month corresponding to series maximum flow (202 A1 + A2 + B stations) 84
EV1 Probability Plots
EV1 Probability Plot 5.1: Nine random samples of size 25 drawn from EV1 72 EV1 Probability Plot 5.2: Nine random samples of size 50 drawn from EV1 73 EV1 Probability Plot 5.3: Station 24082 Maigue at Islandmore 80 EV1 Probability Plot 5.4: Multiple stations on the Rivers Barrow, Suir and Suck 86
Volume II Flood Frequency Estimation
xii
EV1 Probability Plots 8.1: Three discordant sites (Stations 08005, 09002 and 09010) 100 EV1 Probability Plots 8.2: Two further heavily urbanised catchments 102 EV1 Probability Plot 10.1: Single-site and pooled estimates, Station 25016 128 EV1 Probability Plot 10.2: Single-site and pooled estimates, Station 09001 129 EV1 Probability Plot 10.3: Single-site and pooled estimates, Station 09010 129 EV1 Probability Plot 10.4: Single-site and pooled estimates, Station 25002 131 EV1 Probability Plot 10.5: Single-site and pooled estimates, Station 08009 132 EV1 Probability Plot 10.6: Single-site and pooled estimates, Station 09002 133 EV1 Probability Plot 10.7: Single-site and pooled estimates, Station 26008 134
EV1 Probability Plot J.1: Station 07006 Moynalty at Fyanstown 199 EV1 Probability Plot J.2: Station 15003 Dinan at Dinan Bridge 201 EV1 Probability Plot J.3: Station 16008 Suir at New Bridge 203 EV1 Probability Plot J.4: Station 16009 Suir at Cahir Park 205 EV1 Probability Plot J.5: Station 24013 Deel at Rathkeale (post-drainage) 207 EV1 Probability Plot J.6: Station 24082 Maigue at Islandmore 209 EV1 Probability Plot J.7: Station 25017 Shannon at Banagher 211 EV1 Probability Plot J.8: Station 25021 Little Brosna at Croghan 213
Box-plots Box-plots 8.1: Pooled GEV estimates of x100 showing effect of (a) PEAT and (b) FARL 105 Box-plots 8.2: Pooled GEV estimates of x100 showing effect of (a) location and (b) AREA 105 Box-plots 8.3: Period-of-record and decadal effects on pooled flood growth by GEV 106 Box-plots 9.1: Relative SE of single-site quantile estimates – EV1 assumed 116 Box-plots 9.2: As above but with lines superposed to show theoretical values of relative SE 116 Box-plots 9.3: Relative SE of single-site quantile estimates – GEV assumed 117 Box-plots 9.4: As above but with lines superposed to show theoretical values of relative SE 117 Box-plots 9.5: Relative standard error in xT – pooled EV1 simulations 119 Box-plots 9.6: Relative standard error in QT – pooled EV1 simulations 119 Box-plots 9.7: Relative standard error in xT – pooled GEV simulations for k = -0.1 120 Box-plots 9.8: Relative standard error in QT – pooled GEV simulations for k = -0.1 120
Box-plots L.1: 100-year PUM values for eight formulations of distance metric dij 220
Boxes
Box 1.1: The risk equation 6 Box 2.1: Collinearity 17 Box 2.2: Interpretation of urban adjustment models for QMED 39 Box 2.3: Data transfer procedure when one of the catchments is urbanised 43 Box 4.1: Calculation of moments and their dimensionless ratios 51 Box 4.2: Calculation of L-moments and their dimensionless ratios 52 Box 4.3: Station 06030 Big at Ballygoly 53 Box 4.4: Station 08009 Ward at Balheary 68 Box 7.1: FEH pooling scheme 92 Box 7.2: Qualitative outline of simulation results for flood quantile estimate 95 Box 8.1: Box-plots 103 Box 8.2: Incorporation of historical flood data 109 Box 9.1: Comparison with uncertainty of FEH methods 114 Box 10.1: Plotting positions 125 Box 10.2: Combined use of single-site and pooled estimates 135
Tables
Table 2.1: QMED adjustments for period-of-record effects 13 Table 2.2: Rank correlations amongst QMED and 19 PCDs (205 gauged sites) 15
Volume II Flood Frequency Estimation
xiii
Table 2.3: Principal component loadings of PCDs on ℓnQMED for 190 rural stations 16 Table 2.4: Summary information for PCDs at 190 rural stations used in model-building 19 Table 2.5: PCDs in best three 1 to 9-variable models for ℓnQMED (with ℓnAREA forced) 22 Table 2.6: Performance diagnostics for the addition of each independent variable 22 Table 2.7: Coefficient and collinearity statistics for selected rural model for ℓnQMED 24 Table 2.8: Bootstrapped confidence intervals for model coefficients (BCa method) 29 Table 2.9: Factorial change in QMED following drainage 32 Table 2.10: Five catchments poorly predicted by the QMEDrural model 36 Table 2.11: PCDs for Suck at Rookwood worked example 44 Table 2.12: Data transfers for Suck at Rookwood worked example 45 Table 3.1: Number of cases (out of 94) in which the null hypothesis is rejected 48 Table 4.1: Principal summary statistics 53 Table 4.2: Average values of some statistics for gauging stations (by station grade) 58 Table 4.3: High outliers amongst 110 stations graded A1 + A2 69 Table 4.4: Low outliers amongst 110 stations graded A1 + A2 70 Table 5.1: Linear pattern statistics for 110 A1 + A2 stations 74 Table 5.2: EV1 plot curve patterns – 110 A1 + A2 stations ordered by Hazen skewness 77 Table 5.3: Stations for which flood volumes were specially investigated 80 Table 8.1: L-moment ratios, PCDs and station discordancy within set of 88 A1+A2 stations 97 Table 8.2: Stations showing large discordancy values (in pool of 88 A1 + A2 stations) 102 Table 9.1: Typical standard errors when estimating QMED from annual maxima 114 Table 9.2: Relative standard errors for growth factors xT and quantile estimates QT 121 Table 11.1: Rivers having three or more gauging stations for assessment 137
Table C.1: Test of significance of spatial variability in ℓnQMED model parameters 162
Table D.1: Coefficient and collinearity statistics for undrained ℓnQMED model 164 Table D.2: Coefficient and collinearity statistics for drained ℓnQMED model 166 Table D.3: Validation of partitioned and general models for ℓnQMED 167
Table E.1: Stations with significant trends 177 Table E.2: Stations with pre and post-drainage records 178 Table E.3: Test statistics for trend and change in AM flow series 181 Table E.4: Number of cases (out of 94) in which the null hypothesis is rejected 184 Table E.5: Stations for which the trend is judged most highly significant 185 Table E.6: Critical values of the test statistic bs 187
Table J.1: Basic information for Station 07006 Moynalty at Fyanstown 199 Table J.2: Basic information for Station 15003 Dinan at Dinan Bridge 201 Table J.3: Basic information for Station 16008 Suir at New Bridge 203 Table J.4: Basic information for Station 16009 Suir at Cahir Park 205 Table J.5: Basic information for Station 24013 Deel at Rathkeale (post-drainage) 207 Table J.6: Basic information for Station 24082 Maigue at Islandmore 209 Table J.7: Basic information for Station 25017 Shannon at Banagher 211 Table J.8: Basic information for Station 25021 Little Brosna at Croghan 213
Table K.1: Percentage of AM floods occurring in winter half-year (Oct-Mar) 215 Table K.2: Month of maximum recorded flood in AM series 216
Table L.1: Summary of AM flow dataset used in dij study 219 Table L.2: Mean values of PUM100, H1 and H2 for various pooling schemes 219
Volume II Flood Frequency Estimation
xiv
Notation
Lower case symbols
d Inter-catchment distance in km (often measured centroid-to-centroid)
dij Dissimilarity (or other) distance/measure/metric between sites i and j
f(x) Probability density function
g Coefficient of skewness
h Hardness (of data transfer)
k Shape parameter of GEV (or GLO) distribution
ℓn, ln Natural logarithm
m Number of stations (e.g. in a pooling group)
n Number of years of record, sample size
p Number of PCDs/variables in regression model
r Rank
r2 Coefficient of determination
t Test statistic
t2 Sample L-CV value
t3 Sample L-skewness value
t4 Sample L-kurtosis value
2t , 3t and 4t Pooled values of t2, t3 and t4
u Location parameter in extreme value distributions
Vector of L-moment ratios for station i ui
w Weighting term
xT T-year growth factor
y Reduced variate (e.g. for use with EV1/GEV distribution)
yT Value of y corresponding to T-year event
yi Value of y at which ith smallest AM to be plotted
yL Reduced variate for use with Logistic (i.e. LO or GLO) distribution
yN Reduced variate for use with Normal distribution
z Variate used in statistical exposition, z = ℓnX
α Scale parameter in extreme value distributions
β Scale parameter (growth curve)
β Test statistic
γ Euler’s constant (≈0.5772)
λ1 1st L-moment
λ2 2nd
L-moment
λ3 3rd
L-moment
μ Mean
μ3 3rd
central moment
μz Mean of z = ℓnX
ρ Spearman’s rank correlation coefficient
σ Standard deviation
σz Standard deviation of z = ℓnX
ω Weighting term
Upper case symbols
A Matrix of sums of square and cross-products of pooled L-moment ratios
A1 Gauging station grade A1
A2 Gauging station grade A2
B Gauging station grade B
Volume II Flood Frequency Estimation
xv
CV Coefficient of variation
Di (Hosking and Wallis) discordancy measure for station i
E Expected value
F(Q), F Cumulative distribution function
Fi Frequency at which ith smallest AM is to be plotted
F Test statistic
H0 Null hypothesis (in a specific statistical test)
H1 Heterogeneity measure based on L-CV
H2 Heterogeneity measure based on L-CV and L-skewness
H-skew Hazen-corrected skewness coefficient
L-CV Coefficient of L-CV
L-kurt Coefficient of L-kurtosis
L-skew Coefficient of L-skewness
M Number of repetitions in simulation (i.e. sampling) experiments
M100 1st PWM
M110 2nd
PWM
M120 3rd
PWM
M(ρ) Correlation function (used to adjust QMED)
N, n Number of years of record, sample size
N(0, γ2) Normally distributed with mean of zero and a variance of γ
2
Q Flood peak discharge (m3s
-1)
Q1, …, QN Flood series arranged in chronological order
Q(1), …, Q(N) Flood series arranged in rank order (smallest first)
QI Index flood
Qmax Maximum flood on record
Qmean, Q Mean of annual maximum flow series
QMED, Qmed Median of annual maximum flow series
Qmin Minimum AM flood on record
QT T-year flood
R2 Coefficient of determination
T Return period (years)
TAM Return period on annual maximum scale (years)
TPOT Return period on peaks-over-threshold scale (years)
U Test statistic
Var Variance
Matrix of independent variables X
Γ Gamma function
Standardised Normal distribution function
Subscripts
adj Adjusted (estimate)
d Donor (site)
o Overlap (period)
s Subject (site)
Abbreviations and descriptor names
AD, AD2 Abbreviations for ARTDRAIN and ARTDRAIN2
AdjFac Adjustment factor (in data transfer)
AEP Annual exceedance probability
ALLUV Proportional extent of floodplain alluvial deposit
ALTBAR Mean elevation of catchment (m)
AM Annual maximum
Volume II Flood Frequency Estimation
xvi
ANN Artificial neural network
AREA Catchment area (km2)
ARI Average recurrence interval
ARTDRAIN Proportion of catchment area mapped as benefitting from arterial drainage schemes
ARTDRAIN2 Proportion of river network length included in Arterial Drainage Schemes
BCa Bias-Corrected and accelerated (see Section 2.3.2)
BFI Baseflow index
BFIHOST Baseflow index derived from HOST soils data (a UK descriptor)
BFIsoil Soil baseflow index (estimate of BFI derived from soils, geology and climate data)
CDF Cumulative distribution function
CFRAM Catchment flood risk assessment and management
C.I. Confidence interval
Cov Covariance
CV Coefficient of variation
D Downwards concave i.e. convex (a curve pattern)
DMF Daily mean flow
DRAIND Drainage density (km/km2)
EDA Exploratory data analysis
EPA Environmental Protection Agency
ESB Electricity Supply Board
EV1 Extreme Value Type 1 = Gumbel (a 2-parameter distribution)
FAI Flood attenuation index
FARL Index of flood attenuation by reservoirs and lakes
FEH Flood Estimation Handbook
FLATWET PCD summarising proportion of time soils expected to be typically quite wet
FOREST Proportional extent of forest cover
FSE Factorial standard error
FSR Flood Studies Report
FSU Flood Studies Update
GEV Generalised Extreme Value (a 3-parameter distribution)
GLO Generalised Logistic (a 3-parameter distribution)
GLS Generalised least-squares (a regression method)
GWR Geographically weighted regression
HA Hydrometric Area (e.g. HA10 is Hydrometric Area 10)
HGF Highest gauged flow
H-skew Hazen-corrected skewness
IDW Inverse distance weighting
iid Independently and identically distributed
L Linear (a curve pattern)
L-CV Coefficient of L-variation (Hosking and Wallis, 1997)
LH Left hand
L-kurt L-kurtosis
LN2, LN 2-parameter lognormal (a distribution)
LN3 3-parameter lognormal (a distribution)
LO Logistic (a 2-parameter distribution)
L-skew L-skewness
MSL Mainstream length (km)
N.D. Normal distribution
NERC (UK) Natural Environment Research Council
NETLEN Total length of river network above gauge (km)
NUI National University of Ireland
O/L Outlet
OLS Ordinary least-squares (a regression method)
OPW Office of Public Works
PASTURE Proportional extent of catchment area classed as grassland/pasture/agriculture
Volume II Flood Frequency Estimation
xvii
PCA Principal components analysis
PCD Physical catchment descriptor (see Volume IV)
PDF Probability density function
PEAT Proportional extent of catchment area classified as peat bog
POT Peaks-over-threshold
PUM Pooled uncertainty measure (see Section L5 of Appendix L)
PWM Probability-weighted moment
RDS Royal Dublin Society
RH Right hand
RMSE Root mean square error
ROI Region of influence
S S-shaped (a curve pattern)
S1085 Slope of main stream excluding the bottom 10% and top 15% of its length (m/km)
SAAPE Standard-period average annual potential evapotranspiration (mm)
SAAR Standard-period average annual rainfall (mm)
SD Standard deviation
SE, se Standard error (of an estimate)
STMFRQ Number of segments in river network above gauge (differs from FSR definition)
TAYLSO Taylor-Schwartz measure of mainstream slope (m/km)
U Upwards concave (a curve pattern)
UAF Urban adjustment factor
UK United Kingdom
URBEXT Proportional extent of catchment area mapped as urbanised
Var Variance
VIF Variance inflation factor
WLS Weighted least-squares (a regression method)
WMO World Meteorological Organization
Glossary of terms
Term Meaning
Adjusted r2
Effective coefficient of determination r2, after allowing for degrees of freedom
consumed by estimating model parameters
Analogue catchment Catchment that is hydrologically similar to subject catchment in terms of key
PCDs but too distant to make it a natural choice as the pivotal catchment
Annual exceedance
probability AEP Probability of one or more exceedances in a year of a given extreme value
Annual maximum
flow series Time series comprising the largest flow in each year or water-year of record
At-site estimation
The FSU refers to the at-site QMED but not to at-site estimation of the flood
frequency curve; the term single-site analysis is preferred because it provides a
clear contrast with pooled analysis
At-site QMED Estimate of QMED made directly from flood peak data at the particular site
Average recurrence
interval ARI Average interval (often measured in years) between successive exceedances of
a given extreme value
Bias The amount by which a procedure typically overestimates the true quantity; if
the bias is small, the estimator is said to be unbiased, irrespective of scatter
Calibration Comparison of a model’s predictions with actual data, and adjustment of its
parameters to achieve a better fit with reality
Volume II Flood Frequency Estimation
xviii
Term Meaning
Coefficient of
determination r2
Proportion of variation accounted for by (e.g.) a regression model
Confidence interval Bounds within which a population parameter is estimated to lie with a stated
(usually %) confidence; used to indicate the reliability of an estimate
Cumulative
distribution function
CDF Probability F(x) of a value of a random variable X being less than or equal to x
Curvature parameter Parameter controlling the shape of a distribution; chiefly referred to in this
volume as the shape parameter
Donor catchment
Gauged catchment (usually one nearby) whose data are relevant to flood
estimation at the subject site; the pivotal catchment is the donor catchment that
the user judges to be most relevant to the specific flood estimation problem
Easting and Northing Coordinates of a location expressed as distance eastwards and distance
northwards from a fixed reference point
Essentially rural
(catchment)
Catchment for which the proportional extent mapped as urbanised (URBEXT)
is less than 0.015
Exploratory data
analysis EDA
An exploratory approach to analysing (often large) datasets to summarise their
main characteristics; EDA typically provides extensive visual summaries
Factorial error Ratio of estimated value to true value
Factorial standard
error FSE
The factorial standard error of X is the exponential of the standard error of
ℓnX; under certain assumptions, about 68% of estimates of X are expected to
lie within the factorial range 1/FSE to FSE of the true value
Flood-poor period Period in which floods are few and/or small
Flood rating curve Long-term relationship between river flow and water level; the stage-discharge
relationship; used to infer high flows from water-level measurements
Flood-rich period Period in which floods are many and/or large
Generalised least-
squares (GLS)
regression
Variation of ordinary least-squares regression or weighted least-squares
regression that can take account of inter-site dependence in observations of the
dependent variable (e.g. QMED); see Section C3 of Appendix C
Geographically
weighted regression
(GWR)
Technique that expands ordinary regression for use with spatial data; see
Section C4 of Appendix C
Geometric mean nth root of product of a sample of n values of a positive variable
Geostatistical
mapping
Method of mapping a variable (such as the residual error from a model) by
recognising its spatial structure; often done by deriving a semivariogram to
summarise how the variance between pairs of points changes with their
separation
Growth curve Model specifying the proportional increase of peak flow with return period
Growth factor Factor by which the index flood is multiplied to estimate the T-year flood
Hardness of a data
transfer
A “hard” data transfer assumes that the factorial error that a model is seen to
make at the pivotal site will be replicated at the subject site; a “soft” transfer
respects the model’s performance at the pivotal site in part rather than in full
Heterogeneity
In the context of a proposed pooled analysis, the degree of unacceptable
dissimilarity in single-site growth curves at stations in the pooling group,
taking due account of their record lengths
Volume II Flood Frequency Estimation
xix
Term Meaning
Heteroscedacity Uneven error variance
Hydrometry The science, technology and practice of water measurement
Independent and
identically
distributed (iid)
Observations are often assumed to be independent (in time and in space) and
identically distributed (i.e. sharing a common behaviour) for the purpose of
statistical inference; the assumption may not always be realistic but simplifies
the statistical analysis necessary to make practical applications
Index flood
A reference flood that can be relatively reliably estimated from gauged data;
the index flood adopted in the FSU is the median annual flood QMED; this is
the median of the annual maximum (AM) flow series.
Interpolation Any method of computing new data points from a set of existing data points
Interquartile range Measure of dispersion or scale defined as the difference between the 3
rd
quartile and the 1st quartile
Kriging An interpolation method based on a distance-weighted average of data at
neighbouring locations
L-CV An L-moment ratio; provides a dimensionless measure of the spread of the
distribution or the sample data
L-kurtosis An L-moment ratio; provides a dimensionless measure of the peakiness of the
distribution or the sample data
L-moments
Moments computed from linear combinations of the ordered sample values
that lead to summary statistics of (e.g.) variation and skewness; often more
efficient than ordinary moments in parameter estimation of distributions;
L-moments are intimately related to probability-weighted moments
L-moment ratio A dimensionless ratio of two L-moments; useful because values from different
catchments can be compared
L-skewness An L-moment ratio; provides a dimensionless measure of the asymmetry of the
distribution or the sample data
Least-squares
regression A method of fitting a model based on minimising the sum of squared residuals
Location parameter Parameter representing value subtracted from or added to a variable x to
translate the graph of its probability distribution along the x-axis
Median annual flood
QMED
QMED is the median of the annual maximum (AM) series. Half of AM floods
are larger than QMED and half are smaller; thus, the annual exceedance
probability associated with QMED is precisely 0.5; QMED is said to have a
return period of two years on the AM scale of frequency
Met Éireann Irish National Meteorological Service
Method of
L-moments
A method of calibrating the parameters of a distribution so that the lower-order
L-moments respect their sample values
Method of moments A method of calibrating the parameters of a distribution so that the lower-order
moments respect their sample values
Metadata Information about information; catalogue information
Multiple linear
regression Linear regression using two or more independent variables
Non-parametric test
A test that does not require an assumption about the probability distribution
underlying the variable being studied (also known as a distribution-free
method); typically, non-parametric tests use ordinal information e.g. the
position of the values in an ordered sample
Volume II Flood Frequency Estimation
xx
Term Meaning
Non-stationarity Non-stationary effects; see stationary
Ordered sample Data sample {x(1), x(2), …, x(n)} in which elements have been reordered so that
x(1) ≤ x(2) ≤ … ≤ x(n).
Ordinary least-
squares (OLS)
regression
The classical least-squares regression approach in which observations are
treated as being equally reliable and mutually independent (see Section C1 of
Appendix C)
Outlier An observation that lies an abnormal distance from other values in a supposed
random sample from a population
Parametric test A test requiring an assumption about the probability distribution underlying
the variable being studied
Peaks-over-threshold
(POT) series
Time series of independent events exceeding a given threshold; series
comprises magnitudes (e.g. peak exceedance in m3s
-1) and their dates of
occurrence; abstraction requires criteria for judging whether successive peak
exceedances are mutually independent
Pivotal catchment Gauged catchment judged by the user to be most relevant to the specific flood
estimation problem
Pooled analysis
Combined analysis of standardised AM flow data from a group of gauged
catchments deemed hydrologically similar (or otherwise relevant) to growth
curve estimation at the subject site
Pooling group Set of gauged catchments thought to be hydrologically similar to the subject
catchment
Probability density
function PDF
For a continuous random variable x, the PDF specifies the relative frequency
or probability of occurrence of x over all subsets of its range of values
Probability-weighted
moments PWMs
Certain weighted linear functions of the ordered sample data that statistical
theory shows as useful and efficient for parameter estimation of probability
distributions; L-moments are a development of PWM theory
Quantiles
Values taken at regular intervals from the cumulative distribution function of a
continuous random variable; where the CDF is broken into four parts, the
quantiles are known as quartiles; if the CDF is broken into 100 parts, the
quantiles are known as percentiles
Quartiles
For an ordered sample, the quartiles are the three points that divide the dataset
into four equal groups, each group comprising a quarter of the data; the second
quartile is the middle observation i.e. the median of the data; the lower quartile
is the middle value between the smallest observation and the median, while the
third quartile is the middle value between the median and the highest
observation; if the quartile (or some other desired quantile) does not
correspond to an observation, it is usual to interpolate between successive
sample values (if n is odd, the median corresponds to the middle-ranking value
x(i) where i = (n+1)/2; if n is even, the median is taken as the average of x(j) and
x(j+1) where j = n/2)
Region of influence
ROI
The region of influence approach selects gauged catchments that are thought
hydrologically similar to the subject catchment; stations are recruited to the
pooling group according to their nearness to the subject catchment in a
measurement system or metric that represents catchment dissimilarity
Residual Observed value minus the value estimated by a model
Return period T
Average number of years between years with floods exceeding a certain value.
T is the inverse of the annual exceedance probability; thus, a 50-year return
period corresponds to an AEP of 0.02.
Volume II Flood Frequency Estimation
xxi
Term Meaning
Scale parameter Parameter controlling the spread of a distribution
Shape parameter Parameter controlling the shape of a distribution
Single-site analysis Analysis of flood peak data at a particular gauged site; sometimes referred to
as at-site estimation
Skewness A measure of the departure from symmetry of a distribution
Soft data transfer See hardness of a data transfer
Standard-period
average annual
rainfall SAAR
Standard-period average annual rainfall, i.e. annual average rainfall evaluated
across a WMO standard period; in FSU usage, SAAR relates to 1961-90.
Standard deviation Measure of spread of values about their mean
Standard error Estimated standard deviation of a sample statistic such as the mean, i.e. the
standard deviation of the sampling distribution of the statistic
Standard error of
estimate Standard error of the variable being estimated
Stationary
A stochastic process is said to be stationary if its probability distribution is
independent of time; if a series of AM flows can be assumed stationary, the
values can be analysed together without the need to consider their sequence;
the assumption is upset by non-stationarity such as long-term trend, step-
change or long-term cyclical variation
Step-change An abrupt change e.g. in the frequency or magnitude of flood occurrences
Stepwise regression A method of multiple regression which begins by selecting the best one-
variable model and then adds further variables one by one
Subject catchment Catchment for which the flood estimate is required
Summary statistic
A statistical measure (e.g. the mean or standard deviation) used to summarise a
set of observations or results; the measure may be relatively intricate or relate
to a formal statistical test of significance; a summary statistic summarises an
aspect of the data or analysis in one number
Trend A data series shows trend if, on average, the series is progressively increasing
or decreasing
Turning point A data point defined by a triple of consecutive values xi-1, xi, xi+1 such that
xi-1 < xi > xi+1 or xi-1 > xi < xi+1; corresponds to a local maximum or minimum
Variance inflation
factor (VIF)
Reciprocal of the tolerance, which in turn denotes the proportion of the
variance in a given regressor (i.e. one of the PCDs) that cannot be explained by
the other regressors
Water-year Hydrological year beginning 1 October and ending 30 September
Weighted least-
squares (WLS)
regression
A variation of ordinary least-squares regression that can take account of the
variable record lengths supporting observations of the dependent variable (e.g.
QMED); see Section C2 of Appendix C
Volume II Flood Frequency Estimation
xxii
(This page is intentionally blank.)
Volume II Flood Frequency Estimation
1
1 Introduction and data
1.1 Why we estimate flood frequency
Flood frequency analysis is concerned with the assessment of flood magnitudes of stated
frequency (or degree of rarity) for use as input into the process of flood risk assessment and
management. Flood risk assessment is needed in the design of flood relief and protection
works and in the assessment of the safety of existing and planned infrastructure. This
includes domestic properties, commercial and industrial buildings, bridges, roads and
railways, and critical infrastructure such as hospitals, electrical stations, gas stations and
water works.
No development can be guaranteed immune from flooding during its projected life. Flood-
proofing every structure would be prohibitively costly. As a result, developments for which
the consequences and costs are modest may be required to tolerate occasional flooding.
Assessment of the residual risk requires estimation of the probability of occurrence of the
flood magnitude that would inundate or damage the structure or infrastructure. It is these
probabilities that are estimated with the help of flood frequency analysis.
The Flood Studies Update (FSU) arises in the context of the multifaceted approach to flood
risk assessment and management outlined by the Government’s Flood Policy Review Group:
an expert group appointed by the Minister of State. Their report (OPW, 2004) recommended
that a two-pronged approach to flood management be pursued with a greater level of
importance attributed to non-structural flood relief measures supported, where necessary, by
traditional structural flood relief measures. In order to make this possible, the report
recommended that research be undertaken in various sectors to develop a strategic
information base. The update of the Flood Studies Report for Ireland was one such body of
research identified.
A pivotal element identified was the need for improved hydrological information to support
effective decision-making, and a proposal was included for the Flood Studies Update (FSU)
programme. The programme has components related to Rainfall Frequency (Volume I),
Flood Frequency Estimation (this volume), Hydrograph Analysis (Volume III), River Basin
Modelling (Volume V), and Urbanised and Small Catchments (Volume VI). Additional
important features are the application of Geographic Information Systems to automate the
calculation of Physical Catchment Descriptors from digitally mapped data (Volume IV) and
the use of Information Technology to deliver many of the FSU outputs to users via web-
based applications (the FSU Web Portal).
Components of the FSU programme were prescribed in Work Packages. This volume
describes research carried out at NUI Galway and NUI Maynooth between 2006 and 2009.
Volume II Flood Frequency Estimation
2
1.2 Flood data
1.2.1 Introduction
The flood data analysed here are river flows measured at formal gauging stations. The
standard unit of measurement is a cubic metre per second (i.e. m3s
-1). The unit is often
spoken as a “cumec”.
River flow measurement is a central element of hydrometry: the science, technology and
practice of water measurement. The measurement of flood flows is relatively specialised and
typically revolves around establishing a long-term relationship between river flow and water
level known as the flood rating curve. The rating curve is also known as the stage-discharge
relationship. Their review is discussed in Appendix A.
1.2.2 Annual maximum series
The flood frequency analyses reported here are chiefly based on annual maximum data. The
annual maximum flow is the largest instantaneous flow recorded in the hydrological year
beginning 1 October. This is also known as the water-year.
The annual maximum flow is sometimes referred to as the annual maximum flood. This
alternate name is a little loose. In some years, the annual maximum flow is too small to be
considered a flood event. However, the emphasis in analysing annual maxima is on flood
estimation, so the term annual maximum flood is often used.
The time series of annual maximum flows for a particular station is referred to as the annual
maximum (AM) series.
1.2.3 FSU flood datasets
Annual maximum series were provided for some 206 river-flow gauging stations. The
datasets are briefly introduced and some features noted.
It is to be expected that individual researchers favour slightly different interpretations and
groupings. The datasets introduced here are those underpinning the general research on flood
frequency and flood growth rates undertaken at NUI Galway. The variants used in the NUI
Maynooth research of Chapter 2 are discussed in Section 2.1.1.
Attention focused on stations whose quality gradings (see Section A3 of Appendix A) fall
into Categories A1, A2 or B. There are 199 gauged catchments made up of:
45 Grade A1 stations;
70 Grade A2 stations;
67 Grade B stations;
A further 17 stations with both pre and post-drainage records (see Section 2.1.1).
The FSU employs a standard flood known as the index flood. The definition adopted here
(see Section 1.2.5) is the median, QMED, of the annual maximum flows. Estimation of
QMED is relatively tolerant of imperfection in flood rating curves (see Appendix A). Hence,
Volume II Flood Frequency Estimation
3
the Grade B stations are included in the research on QMED estimation from physical
catchment descriptors reported in Chapter 2. However, the analysis of flood growth rates and
other aspects warrants use of only the best datasets. Most recommendations are therefore
based on results obtained from A1 and A2 graded stations only.
For consistency and completeness, results are sometimes presented or referred to for the
Grade B stations. However, other than in the Chapter 2 work, no weight is given to them
when making general deductions about Irish flood behaviour.
For various reasons, some of the studies described are based on subsets of the overall dataset
of A1 and A2 graded stations. For example, some analyses needed:
Annual maximum flow series that were strictly complete, i.e. without any gaps;
Multiple gauges on a river (i.e. sets of gauges up and down the river system);
Stations with gauged estimates of the baseflow index (BFI).
In addition, some studies were completed before data for all stations became available. Such
features or exceptions are generally noted in the relevant part of the report.
1.2.4 Expressions of flood rarity
Strictly, any measure of event rarity should refer to a precise feature such as the peak flood
level or the peak flow in m3s
-1. Except where expressly stated otherwise, the feature dealt
with in this report is the peak flow.
The recommended measure of flood rarity for communication to the public is the annual
exceedance probability (AEP) of the flood rather than the return period in years. For
example, a large and damaging flood may have an AEP of 0.01, meaning that there is a one
in a hundred chance of its magnitude being exceeded in any one year.
Despite this clear statement of preferred terminology, the principal measure of rarity adopted
here is the return period (T) in years. The return period is simply more convenient in
technical reports. The two measures of rarity are fully interchangeable, with:
AEP = 1/T 1.1
The return period is defined as the average number of years elapsing between successive
exceedances of that flood magnitude. The word average needs to be stressed as a reminder
that any flood magnitude can be exceeded at any time, regardless of the recent flood history.
Some further remarks on return period are made in Section 1.5.
1.2.5 Index flood
An index flood is a reference flood that can be relatively reliably estimated from gauged data.
The index flood adopted in the FSU is the median annual flood, QMED. This is the median
of the annual maximum (AM) flow series.
Half of AM floods are larger than QMED and half are smaller. Thus, the annual exceedance
probability associated with QMED is precisely 0.5. QMED is said to have a return period of
two years on the AM scale of frequency.
Volume II Flood Frequency Estimation
4
1.3 Historical floods
Remarkable floods have occurred in Ireland that do not feature in the systematically
measured data series. Most often this is because gauging stations did not exist on the relevant
river. One example is the notable flood on the River Dargle at Bray on 25/26 August 1986
occasioned by Hurricane Charley (Charlie in some sources). Even on this one river, there are
other known historical floods, notably in August 1905, September 1931 and November 1965.
More widely across Ireland, there are many late 19th
century and early 20th
century floods for
which there is photographic or documentary evidence but no formal gauged flow data. A
notable example is for the town of Mallow, for which documentary evidence suggests that
flooding from the Munster Blackwater in November 1853, November 1916 and December
1948 exceeded the largest flows in the formal gauged record.
A more recent example of a notable event that passed largely unmeasured by the river
gauging network is the Newcastle West flood of 1 August 2008. The town is on the River
Arra, an ungauged tributary of the River Deel.
There are gauging stations on some catchments known to have experienced significant floods
– e.g. the Deel at Crossmolina – but data have not been supplied. This is generally because
the station does not meet the required standard for inclusion in Grade A1, A2 or B.
Further examination of notable floods is warranted, especially where they may be
symptomatic of damaging events on small steep catchments of a kind not well represented in
the main database of Grade A1 and A2 stations.
The National Flood Hazard Mapping website (http://www.floodmaps.ie/) provides
information on these and other historically important floods. It is essential viewing when
estimating flood frequency on a specific river. Some notable events are summarised by Met
Éireann (e.g. http://www.met.ie/climate-ireland/weather-events/Aug1986_HurCharlie.pdf).
British Rainfall yearbooks – now available online – sometimes provide useful information in
earlier eras (e.g. http://www.metoffice.gov.uk/media/pdf/7/a/British_Rainfall_1905.pdf).
1.4 Structure of volume
Chapter 2 reports the NUI Maynooth study of the relationship between the index flood
(QMED) and Physical Catchment Descriptors (PCDs). Section 2.6 discusses the important
topic of data transfer methods, for which a detailed worked example follows in Section 2.7.
Later chapters report research undertaken at NUI Galway, with Chapters 3 to 5 describing
relatively general statistical analyses of the AM flow series:
Trend and randomness (Chapter 3);
Descriptive statistics (Chapter 4);
Probability plots (Chapter 5).
Volume II Flood Frequency Estimation
5
Later chapters review the general problem of determining the T-year return period flood
magnitude, QT. Topics include:
Introduction to QT estimation by the index flood method (Chapter 6);
Determining the flood growth factor, xT, by single-site and pooled analysis of the AM
flow data (Chapter 7), including the construction of pooling groups (Section 7.3);
Studying the effect of catchment type and period of record on pooled estimates of QT
(Chapter 8);
Assessment of uncertainty in estimates of QMED, xT and QT (Chapter 9).
Guidelines for determining the design flood QT are presented in Chapter 10, together with
some challenging examples where application of a single prescribed strategy may generate
QT estimates that are counter-intuitive. This leads on to a discussion in Section 10.5 of
additional topics such as:
Problems in the use of 3-parameter distributions in single-site analysis;
The choice between single-site and pooled estimates of QT;
Flood growth curves with an upper bound.
Data transfers and the selection of the pivotal catchment are further discussed in Chapter 11.
One of many conclusions drawn in Chapter 12 is that blind application of a rule-based
procedure for flood frequency estimation is both inappropriate and impractical. Factors
beyond codification in any determination of QT include:
The history of flooding at the site;
The trade-off between single-site and pooled estimates of flood growth;
The plausibility (or not) of a derived flood frequency curve that has an upper bound.
Supporting material is presented in appendices.
1.5 Material in common with rainfall frequency studies
Some topics encountered in flood frequency estimation overlap with the rainfall frequency
research presented in Volume I. They are therefore only briefly mentioned here.
1.5.1 Annual maximum and peaks-over-threshold series
By definition, the annual maximum (AM) series comprises the highest values in each year.
The second highest value in a year is ignored, whether or not it exceeds the highest values in
other years. In contrast, the peaks-over-threshold (POT) series consists of all extreme values
exceeding a certain threshold. Although not considered in the FSU flood frequency research,
POT methods can sometimes assist in the estimation of QMED when the length of record is
shorter than ten years (but see also Section 2.1.2).
Volume II Flood Frequency Estimation
6
The probability of the T-year extreme event being exceeded at least once in L years is:
L)
T
1(11r 1.2
Hydrologists typically refer to this as the risk equation. It can be reasoned by rewriting:
L)T
1(1r1
and noting that the probability of the T-year extreme not being exceeded in L years (i.e.
the LH side of the equation) is simply the probability of the T-year extreme not being
exceeded in any of the L individual years (i.e. the RH side of the equation).
Risks are not always as one imagines. For example, there is an even chance (r = 0.5) that
the 100-year rainfall is exceeded in any 69-year period:
0.500)100
1(11 69
The risk equation can be applied to flood peaks as well as to extreme rainfall depths of a
given duration. It should be noted that such applications assume that the system
producing the extremes is stationary, i.e. that annul maxima are statistically independent
and drawn from the same underlying distribution. This assumption is compromised by
climate change. In the case of flood risk, it may also be compromised by catchment
change.
1.5.2 Return period
The return period T is best thought of as the inverse of the annual exceedance probability.
For example, the peak flow corresponding to T = 50 has a probability of 0.02 of being
exceeded in any year. It can be helpful to refer to this return period as the return period on
the annual maximum scale, to avoid possible confusion with other measures of frequency
(see Sections 1.5.3 and 1.5.4).
Though it is often misunderstood by the public, return period is a useful concept to the
professional. Its importance is perhaps best conveyed in understanding the often appreciable
risk of an extreme event being experienced in the long run. The risk equation (see
Equation 1.2 in Box 1.1) expresses the chance of the T-year event being exceeded in a long
run or lifetime of L years.
Box 1.1: The risk equation
1.5.3 Average recurrence interval
The analysis of peaks-over-threshold series gives the average interval between events that
exceed a particular value. This is often termed the average recurrence interval (ARI). For
high values of T, values of ARI and T are nearly equal. But for T less than 20 years the
difference is large enough to be important. Some analysts refer to ARI as the return period
on the peaks-over-threshold scale, and denote it by TPOT.
Volume II Flood Frequency Estimation
7
1.5.4 Langbein’s formula
Langbein (1949) provides a formula relating the two measures of frequency, i.e. T and ARI:
ARI
1exp1
T
1 1.3
This yields pairs of values such as:
T = 1.16 years when ARI = 0.5 years (twice per year frequency);
T = 1.58 years when ARI = 1 year (once per year frequency);
T = 2.54 years when ARI = 2 years (one in two years frequency).
For return periods of five years or longer, the approximation T = ARI + 0.5 suffices.
Langbein’s formula converts any ARI of interest into an equivalent return period or vice
versa.
For those favouring the notation TAM and TPOT, Langbein’s formula can be written:
POTAM T
1exp1
T
1 1.4
Volume II Flood Frequency Estimation
8
2 Estimation of the index flood, QMED
This chapter reports research undertaken by NUI Maynooth. The research is targeted at
relating the index flood to Physical Catchment Descriptors (PCDs). This is necessary to
allow QMED to be estimated at ungauged sites.
Even where flood data are available elsewhere on the same river, the model relating QMED
to PCDs is an intrinsic part of the procedure for transferring information from the gauged site
to the subject site. This topic is discussed in Sections 2.6 and 2.7, and revisited in
Chapter 11.
2.1 Exploratory data analysis
2.1.1 Introduction to the datasets
Estimation of the index flood for ungauged catchments is based on the construction of an
empirical model. This links the index flood (known only at gauged sites) to PCDs (known at
all sites). The PCDs are described in Volume IV.
The annual maximum flow series and values for QMED were provided for a total of 206
gauging stations. Not all stations provided were used. Some were discarded following an
exploratory data analysis. Other questionable stations came to light during model-building.
Appendix B identifies 16 stations omitted from QMED modelling, and justifies the decisions
taken. This left 190 stations for model-building.
Arterial drainage
Where arterial drainage has taken place within the record, the series in question was treated in
two parts and separate values of QMED extracted for the pre and post-drainage periods. In
total, 15 stations were divided into pre and post-drainage records.
The set of 190 catchments was thereby increased to 205 for model-building. These 205
comprise 190 catchments nominally in their current condition plus 15 in their pre-drainage
condition. The make-up of the 205 catchments can be further summarised as: 131 (i.e. 116 +
15) stations without arterial drainage, and 74 stations with arterial drainage (i.e. 59 + 15).
Flood data quality
Figure 2.1 shows the number of stations in each category of flood data quality. There are 58
stations graded A1, 78 stations graded A2 and 69 stations graded B.
Length of record
In total, the full dataset of 205 stations comprises 6350 annual maximum events. The stations
have a mean length of 31 years, with a maximum of 65 and a minimum of seven years. The
distribution of record-lengths is summarised in Histogram 2.1.
Volume II Flood Frequency Estimation
9
Grade A1
Grade A2
Grade B
69
78
58
Figure 2.1: Number of stations in each quality category (full dataset of 205 stations)
706050403020100
15
10
5
0
Number of annual maxima from which QMED estimated
Fre
quency
EPA stations
OPW stations
Histogram 2.1: Length of available record for QMED estimation (205 stations)
Most of the longer records are from OPW stations. The bimodality evident in the distribution
of record lengths likely reflects changing priorities and budgets. It may also reflect that
arterial drainage has divided some otherwise long flood series into shorter records. The FSU
treats pre-drainage and post-drainage flood series as separate stations.
Histogram 2.2 shows the spread of QMED values, which are seen to be distributed
approximately lognormally. The 205 stations are marked in Map 2.1. Because of the
collocation of pre-drainage and post-drainage stations, and the limited resolution of plotting,
fewer than 205 symbols are discernible on the map.
Volume II Flood Frequency Estimation
10
6.21
46
5.29
83
4.60
52
3.91
20
2.99
57
2.30
26
1.60
94
0.69
31
0.00
00
20
10
0
QMED m s
Fre
quency
1 2 5 10 20 50 100 200 5003 -1
[The 205 QMED values are seen to be approximately lognormally distributed]
[Logarithmic scale]
Histogram 2.2: QMED values across full dataset of 205 stations
2.1.2 Adjustments for period-of-record effects
Natural variability in climate can lead to flood series exhibiting so-called flood-rich and/or
flood-poor periods. Several major floods may occur in a flood-rich period only to be
succeeded by a flood-poor period in which no major flood occurs for many years. In
consequence, QMED estimates from short records can be unrepresentative of the longer term.
Estimates of QMED were therefore adjusted for period-of-record effects using a method
broadly based on Robson (1999c).
Stations having a flood series shorter than 20 years were taken to be short records in need of
adjustment. The adjustment is accomplished by transferring information from long-record
sites. To be considered for this role, the donor station had to have a Grade A1 or A2 flood
rating, have a record length of over 30 years and be free from arterial drainage.
Long-record catchments within a 50 km radius of the short-record (subject) site were initially
selected. Only donors that overlapped with at least three-quarters of the subject site’s annual
maximum data were considered.
The correlation between subject and donor sites was assessed using Spearman’s rank
correlation, ρ (e.g. Kendall, 1975). An observation has rank r if it is the rth
largest in a
dataset. Donors that fully overlap the period of record at the subject site were prioritised but
those revealing a weak or negative rank correlation were excluded from the adjustment
process. The number of donors used ranged from one (where the donor fully overlapped the
record at the subject site and ρ > 0.8) to five (where no outstanding donor was identified).
Volume II Flood Frequency Estimation
11
Map 2.1: Spatial distribution of 205 catchments
In order to transfer information from the donor site to the subject site, QMED at the donor
site was estimated using all available data and then recalculated using only the period of
record available at the subject site. The ratio of these two values provides the adjustment
factor at the subject site. The QMED estimate adjusted to the donor period was taken as:
ρM
od,
dsadjs,
QMED
QMEDQMEDQMED
2.1
Here, QMEDs,adj is the adjusted QMED at the subject site, QMEDs and QMEDd are estimates
at the subject site and the donor site based on their respective periods of record, and QMEDd,o
is QMED at the donor site evaluated for the period of overlap.
[Symbols placed at
catchment centroids]
Volume II Flood Frequency Estimation
12
The exponent M(ρ) is a moderating influence based on the strength of correlation between
donor and subject site. [Editorial note: The exponent is later termed the “hardness” of the
data transfer, and denoted by h. See Step 5 in Section 2.7.1.] M(ρ) is defined by:
1ρ4n
ρ3nρM
2
o
3
o
2.2
where no is the length of overlap between subject and donor sites and ρ is Spearman’s rank
correlation between annual maxima at subject and donor sites.
Where only one donor with a very strong correlation (ρ > 0.8) was found, the adjustment
process was deemed complete. However, where several donor sites were identified, a
combined adjustment was made by weighting each donor based on distance from subject site,
additional years of data provided by the donor, and the strength of correlation with the subject
site. The weighting factor is formulated as:
ρnnn100
d1w odo
2.3
where nd is the length of the donor series and d is distance in km.
Table 2.1 details the 28 stations at which the QMED value was adjusted for a period-of-
record effect. As confirmed in Figure 2.2, the procedure led to relatively minor adjustments
in QMED. The most notable exception was the Cammock at Killeen Road (Station 09035),
for which the adjustment raised QMED by 31% (from 11.70 to 15.28 m3s
-1). The Cammock
is a short-record station known to be affected by urbanisation. Five of the nine annual
maxima shown in Figure 2.3 were recorded in summer (Apr-Sep) rather than winter
(Oct-Mar), which is consistent with an urban flood regime.
2.1.3 Physical catchment descriptors
Some 20 physical catchment descriptors (PCDs) were considered. Their derivations are fully
documented in Volume IV. [Editorial note: The flood attenuation index FAI was
unavailable at the time of study. Because QMED represents a relatively frequent (2-year)
flood, the omission of a PCD reflecting floodplain attenuation is not thought to be too
important.] In initial studies, the FLATWET descriptor was found to enter the QMED model
in a counter-intuitive manner. Given the reservations about this PCD expressed by the FSU
Steering Group, FLATWET was omitted from later modelling of QMED.
The remaining 19 PCDs were logarithmically transformed prior to the main model-building.
Where the lower range of a particular descriptor can take a value of zero, 1.0 is added to the
value prior to the log transformation. This applied to the fractions URBEXT, FOREST,
PEAT, PASTURE, ALLUV, ARTDRAIN and ARTDRAIN2. In some tables and figures that
follow, ARTDRAIN and ARTDRAIN2 are abbreviated to AD and AD2 respectively.
PCDs were screened both by the non-parametric approach of calculating rank correlations
(see Section 2.1.4) and by plotting the ℓn-transformed PCDs against ℓnQMED (see
Section 2.1.6). Scatter-plots help to check for possible outliers and non-linear relationships,
as well as for cross-correlations between the PCDs.
Volume II Flood Frequency Estimation
13
Table 2.1: QMED adjustments for period-of-record effects
Station
number
Rating
grade River Station name
# annual
maxima QMED QMEDadj
Adjustt
factor
01055 B Mourne Beg M. Beg Weir 9 2.70 2.80 1.036
07006 A2 Moynalty Fyanstown 19 27.93 25.20 0.902
07041 A2 Boyne Ballinter Br. 7 165.00 165.28 1.002
08007 B Broadmeadow Ashbourne 17 8.24 8.12 0.986
08009 A1 Ward Balheary 14 5.00 5.09 1.018
08012 B Stream Ballyboghill 19 4.35 4.35 1.000
09010 A1 Dodder Waldron’s Br. 18 47.05 46.64 0.991
09035 B Cammock Killeen Road 9 11.70 15.28 1.306
10028 B Aughrim Knocknamohill 16 46.95 46.29 0.986
13002 B Corock Foulk’s Mills 19 7.01 6.98 0.996
14034 A2 Barrow Bestfield 17 117.00 117.07 1.001
15007 A2 Nore Kilbricken 13 53.45 53.58 1.002
15012 B Nore Ballyragget 16 77.11 76.18 0.988
16051 B Suir Clobanna 13 2.85 2.82 0.989
19046 B Martin Station Road 9 29.95 28.33 0.946
22003 B Maine Riverville 8 98.01 98.03 1.000
22035 B Laune Laune Bridge 14 116.40 110.42 0.949
23012 A2 Lee (Kerry) Ballymullen 18 15.66 15.83 1.011
25038 B Tyone Nenagh 17 39.30 37.68 0.959
25124 A2 Brosna Ballynagore 18 13.65 13.36 0.979
25158 A1 Bilboa Cappamore 18 43.88 37.06 0.845
26014 B Lung Banada Br. 16 42.82 42.18 0.985
26108 A2 Owenure Bellavahan Br. 15 57.32 55.92 0.976
30012 B Clare Claregalway 9 126.00 116.97 0.928
34010 B Moy Cloonacannana 12 95.42 99.21 1.040
34029 B Deel Knockadangan 9 110.00 110.00 1.000
36016 B Annalee Rathkenny 14 50.70 50.70 1.000
39001 B Swilly New Mills 17 47.80 47.05 0.984
Volume II Flood Frequency Estimation
14
2.3012.0001.6991.3011.0000.6990.301
2.301
2.000
1.699
1.301
1.000
0.699
0.301
QMED ( m s )
QM
ED
adj
2 5 10 20 50 100 200
1:1 line
2
20
200
50
100
5
10
3 -1
Station 09035
Figure 2.2: QMED adjusted for period-of-record effects
20042002200019981996
30
20
10
0
Water-year
Flo
od p
eak
Winter
Summer
Figure 2.3: Annual maximum flow series for Station 09035 Cammock at Killeen Road
2.1.4 Rank correlations
Table 2.2 shows rank correlations amongst QMED and the PCDs for the 205 catchments.
[Editorial note: The rank correlation is the correlation between the ranks of the variables
within the dataset. It is unaffected by whether or not the ℓn-transformation has been applied,
making it a convenient and relatively robust measure of the degree of association between
values of the PCDs.] The stronger correlations (and anti-correlations) are highlighted, with
values greater than 0.9 (or less than -0.9) shown in red. Values shown in orange and green
mark pairs of PCDs that are progressively less strongly correlated (or anti-correlated).
Flo
od p
eak (
m3 s
-1)
QM
ED
adj (
m3 s
-1)
Volume II Flood Frequency Estimation
15
Table 2.2: Rank correlations amongst QMED and 19 PCDs (205 gauged sites)
Competition is evident amongst the PCDs that reflect catchment size, namely: AREA, MSL,
NETLEN and STMFRQ. [Editorial note: The FSU definition of STMFRQ (see Volume
IV) is the number of streams in the catchment. This is one greater than the number of stream
junctions. STMFRQ is therefore heavily correlated with size variables. In contrast, the FSR
(NERC, 1975) defined STMFRQ as a standardised stream frequency i.e. the number of
junctions per unit area.]
Unsurprisingly, there is strong competition between the PCDs representing arterial drainage
(ARTDRAIN and ARTDRAIN2) and between the slope descriptors (S1085 and TAYLSO).
The marked associations between the land-use fractions FOREST, PEAT and PASTURE
reflect that afforestation tends to be more prevalent on catchments with peat-based soils,
whereas pasture is more typical of catchments with non-peat soils. These patterns are
consistent with the positive association of SAAR with FOREST and PEAT, and its negative
association with PASTURE. SAAR is the long-term average annual rainfall.
2.1.5 Principal component analysis
The transformed PCDs were subject to a principal component analysis to explore the
dominant factors in explaining the variation of ℓnQMED. Table 2.3 displays the results
following Varimax rotation. This special type of orthogonal rotation maximises the sum of
the variance of the loading vectors. Its goal is to simplify the structure of the components by
making the large loadings larger and the small loadings smaller.
[Editorial note: Because of the artificial and important influence of urbanisation on flood
response times and flood magnitudes, the exploratory work studied QMED on the
190 catchments that are largely rural. This explains why URBEXT does not appear in the
principal component analysis (PCA). The mean altitude (ALTBAR) was considered in the
PCA but did not prove influential.]
Volume II Flood Frequency Estimation
16
Table 2.3: Principal component loadings of PCDs on ℓnQMED for 190 rural stations
Component 1 2 3 4 5 6 7 8 9
ℓnAREA 0.961
ℓnMSL 0.965
ℓnNETLEN 0.986
ℓnSTMFRQ 0.957
ℓnDRAIND 0.937
ℓnS1085 -0.620 0.584
ℓnSAAR 0.700 0.307
ℓn(1+FOREST) 0.819
ℓn(1+PEAT) 0.841
ℓn(1+PASTURE) -0.877
ℓn(1+ALLUV) 0.851
ℓnSAAPE 0.902
ℓnFARL 0.870
ℓnBFIsoil 0.836
ℓnTAYSLO -0.603 0.629
ℓn(1+ARTDRAIN) 0.968
ℓn(1+ARTDRAIN2) 0.961
Unsurprisingly, the first component is dominated by terms (directly or indirectly) reflecting
catchment size. Large catchments tend to have small S1085 and TAYSLO slopes. The
second component reflects catchment wetness and land-use, with subsequent components
reflecting arterial drainage, storage attenuation (indexed by ℓnFARL) and drainage density.
2.1.6 Correlations and competing variables
Given the strong correlations between ℓnQMED and the PCDs reflecting catchment size –
ℓnAREA, ℓnMSL, ℓnNETLEN and ℓnSTMFRQ – only ℓnAREA was retained in the model-
building. AREA is the natural descriptor of catchment size.
Strong correlations were also noted between ℓnSAAR, ℓn(1+FOREST) and altitude (indexed
by ALTBAR). These reflect the influence of topography on rainfall and land-use. Forested
areas are chiefly located in upland areas of high rainfall. The traditional descriptor
(ℓnSAAR) was selected.
The slope descriptor ℓnS1085 showed a marginally stronger relationship with ℓnQMED than
did the alternate ℓnTAYLSO. More exhaustive studies in later modelling of ℓnQMED led to
ℓnS1085 being consistently selected in preference to ℓnTAYLSO. The latter was eventually
dropped from further analysis.
Volume II Flood Frequency Estimation
17
The decision to retain only one of the competing variables at an early stage of the analysis
avoids problems of collinearity (see Box 2.1). Although the descriptors of arterial drainage
are highly correlated (see Figure 2.4), it was unclear which variable to prefer. Both were
retained but ARTDRAIN2 ultimately proved the more useful.
Box 2.1: Collinearity
1.00.90.80.70.60.50.40.30.20.10.0
0.5
0.4
0.3
0.2
0.1
0.0
ARTDRAIN2
AR
TD
RA
IN
[Regression: ARTDRAIN = 0.276 ARTDRAIN2 (intercept not significant)]
Figure 2.4: Relationship between PCDs summarising extent of arterial drainage
2.1.7 Scatter-plots and summary information for selected PCDs
Figure 2.5 shows a matrix of scatter-plots and correlations for ℓnQMED and a selected subset
of the ℓn-transformed PCDs. Table 2.4 provides some summary information.
Collinearity refers in a strict sense to the presence of exact linear relationships within a set
of variables. Typically, these are a set of candidate explanatory (i.e. predictor) variables
in a regression-type model. In statistical usage, collinearity also refers to near-
collinearity, i.e. when variables are close to being linearly related.
In a multiple regression with collinearity, least-squares regression coefficients are highly
sensitive to very minor changes in the input data. The least-squares problem or the dataset
is said to be ill-conditioned. Some or all of the regression coefficients are likely to be
meaningless.
A typical approach to overcoming collinearity is to simplify the problem, e.g. by retaining
only one of the subset of variables that are highly correlated. Relatively arbitrary
decisions – as to which variables to retain and which to remove – are sometimes
unavoidable and inevitably influence the final model achieved.
Volume II Flood Frequency Estimation
18
Figure 2.5: Association of ℓnQMED and selected PCDs for 205 catchments
2.2 Rural-catchment model for estimating QMED from PCDs
2.2.1 Regression methods
Regression has long been used to relate a desired flood quantile to catchment physiographic,
geomorphologic and climatic characteristics (e.g. Nash and Shaw, 1965; NERC, 1975). The
outcome is an estimation equation. This is typically referred to as the catchment-
characteristic or catchment-descriptor equation.
The model typically takes the form of a power-law equation:
p321 β
p
β
3
β
2
β
1T .....xxxxaQ 2.4
where:
QT is the flood quantile of interest;
xi is the ith
physical catchment descriptor;
βi is the related model parameter;
a is a multiplier (which is a further parameter of the model);
p is the number of catchment descriptors.
Lower triangle shows scatter-plot. Upper triangle shows (Pearsonian) correlation coefficient. ℓnQMED 0.83 0.09 -0.39 0.25 -0.03 -0.10 -0.01 ℓnAREA -0.27 -0.71 -0.06 -0.13 0.29 0.05 ℓnDRAIND 0.45 0.49 -0.16 -0.49 -0.13 ℓnS1085 0.31 0.14 -0.44 -0.21 ℓnSAAR -0.40 -0.37 -0.24 ℓnFARL -0.24 0.03 ℓnBFI
soil 0.16
ℓn(1+AD2)
Volume II Flood Frequency Estimation
19
Table 2.4: Summary information for PCDs at 190 rural stations used in model-building
PCD
notation Unit Meaning Min
Geom
mean Max
Main form in which
variable is used
AREA km2
Catchment area from
DTM 5.46 218 7980 ℓnAREA
DRAIND km/km2 Drainage density 0.27 1.03 2.64 ℓnDRAIND
SAAR mm
Standard-period
average annual rainfall
(1961-90)
711 1101 2465 ℓnSAAR
URBEXT
fractions
Urban extent 0.00 0.019 0.683 ℓn(1+URBEXT)
PEAT Peat extent 0.00 0.126 0.802 ℓn(1+PEAT)
ALLUV Alluvium extent 0.00 0.034 0.108 ℓn(1+ALLUV)
PASTURE Pasture extent 0.00 0.697 1.000 ℓn(1+PASTURE)
SAAPE mm
Standard-period
average annual potential
evaporation (1961-90)
448 501 563 ℓnSAAPE
FARL –
Index of flood
attenuation from
reservoirs and lakes
0.632 0.948 1.000 ℓnFARL
BFIsoil –
Soil baseflow index
(estimate of BFI
derived from soils,
geology and climate
data)
0.294 0.574 0.814 ℓnBFIsoil
ARTDRAIN
(AD) –
Proportion of catchment
area mapped as
benefitting from arterial
drainage
0.00 0.047 0.367 ℓn(1+ARTDRAIN)
ARTDRAIN2
(AD2) –
Proportion of river
network length mapped
as included in Arterial
Drainage Schemes
0.00 0.150 0.846 ℓn(1+ARTDRAIN2)
S1085 m/km
Mainstream slope
(excluding top 10% and
bottom 15%)
0.24 2.61 30.8 ℓnS1085
The index flood adopted in the Flood Studies Update is the median annual flood, QMED.
This is the median of the annual maximum flood values. Equation 2.4 holds that changes in
physical catchment descriptors (PCDs) have a scaling effect on the index flood, with the
degree of scaling indicated by the exponent terms.
Several techniques are available for estimating the p+1 model parameters. The most common
is multiple linear regression. This is invoked by linearising Equation 2.4 through a
logarithmic transformation, leading to the form:
pp332211 xnβ...xnβxnβxnβanQMEDn 2.5
Volume II Flood Frequency Estimation
20
The basic principle of regression is to achieve a best fit to the data by minimising a measure
of the deviation of the estimated values (of ℓnQMED) from the observed values (of
ℓnQMED). The criterion typically applied is to minimise the sum of squares of the
deviations. This gives rise to the term least-squares regression.
Models of the form of Equation 2.5 can be fitted using regression techniques supported by
standard statistical packages. Some methods begin by selecting the best one-variable model
(i.e. p = 1) and then adding further variables one by one. This is stepwise regression.
However, for many datasets, computer power is typically sufficient to allow exhaustive
searching (across all possible linear models) for the best 1-variable model, the best 2-variable
model, etc. This is the technique generally employed here.
There are several versions of multiple linear regression. The formulation chiefly used is
ordinary least-squares (OLS). However, consideration is also given to weighted least-squares
(WLS) and generalised least-squares (GLS). Some details of linear least-squares regression
are given in Section C1 of Appendix C.
2.2.2 Alternative methods
McCuen et al. (1990) highlight that techniques such as OLS, while leading to an unbiased
estimate of ℓnQMED, can lead to a biased estimate of QMED. In addressing this problem, a
number of authors have applied more complicated procedures such as non-linear and non-
parametric regression (e.g. Pandey and Nguyen, 1999). The use of such techniques was not
attempted here. [Editorial note: In most flood-risk applications, proportional changes
and/or factorial errors in peak flow are of greater relevance than absolute changes/errors.
This makes estimates of ℓnQMED typically more relevant than estimates of QMED.]
As a precaution, and given the impressive results achieved by Dawson et al. (2006), the use
of Artificial Neural Networks (ANNs) was also examined. However, the selected approach
was found to outperform ANNs for the FSU dataset. The alternative method is not reported.
2.2.3 Selection of catchments for calibration and validation
The approach taken began by deriving a method of estimating QMED for use on essentially
rural catchments. Urban adjustments are considered in Section 2.5.
Thirty-five of the 205 catchments have an urban fraction (i.e. URBEXT value) greater than
0.015. Excluding these catchments left 170 stations for modelling QMED on essentially rural
catchments. These 170 catchments include 15 pairs of catchments for which pre-drainage
and post-drainage behaviour were both considered (see also Section 2.4).
Split-sampling allowed some testing of the derived models. The 170-catchment rural dataset
was split 85%/15% to provide 145 stations for calibration and 25 stations for validation. The
allocation of stations to the validation group was made at random. As seen in Histograms
2.3, this resulted in a fair distribution of catchment sizes across the two datasets.
Volume II Flood Frequency Estimation
21
15
10
5
0
5000200010005002001005020105
10
5
0
Num
ber
of
cat
chm
ents
in A
RE
A b
and
Calibration dataset (145 rural catchments)
Validation dataset (25 rural catchments)
Histograms 2.3: Catchment sizes for calibration and validation datasets
2.2.4 Selecting PCDs
Selecting the combination of PCDs to be included in the final QMED model was a lengthy
and iterative process. Not every stage is reported here.
An exhaustive search found the best five sets of variables by fitting every combination of
descriptors (up to a maximum of nine PCDs) to the 145-catchment calibration dataset.
The fitted models were assessed on size (i.e. number of PCDs included), the coefficient of
determination (r2), the root mean square error (RMSE) of prediction, their hydrological
realism and the behaviour of the model residuals.
The RMSE is also known as the standard error of estimate. Here, the RMSE applies to
ℓnQMED. The factorial standard error (FSE) of the index flood itself, i.e. of QMED, is
given by:
RMSEeFSE 2.6
Table 2.5 indicates the PCDs that feature in the three best-fitting OLS models of each model
size, i.e. using from one to nine PCDs as regressor variables. For example, the best-fitting
2-variable model uses ℓnAREA and ℓnBFIsoil and has a factorial standard error of
e0.465
= 1.59. The names shown in the table denote the relevant transformed variable, e.g.
AREA denotes ℓnAREA and ALLUV denotes ℓn(1+ALLUV). The descriptor PEAT was
included in the analysis but did not feature in any of the best models reported in Table 2.5.
AREA (km2)
Volume II Flood Frequency Estimation
22
Table 2.5: PCDs in best three 1 to 9-variable models for ℓnQMED (with ℓnAREA forced)
2.2.5 Choosing a model
Table 2.6 reports performance statistics for the best 1-variable, 2-variable, 3-variable, …,
9-variable models for estimating ℓnQMED on rural catchments. This was based on an
exhaustive search.
The r2 value increases inexorably as the model increases in size. But the r
2 value levels off at
the 7-variable model (see Figure 2.6). The adjusted r2 allows for the degrees of freedom
consumed by estimating model parameters. An F test (based on the ratio of variances)
confirms that the 8-variable model does not provide a significant improvement on the
7-variable model.
Table 2.6: Performance diagnostics for the addition of each independent variable
Number of PCDs r r2 r
2 change Adjusted r
2 RMSE F ratio
Significance
of F ratio
1 0.786 0.618 0.618 0.615 0.629 229.57 0.000
2 0.890 0.793 0.175 0.790 0.465 119.09 0.000
3 0.918 0.843 0.050 0.839 0.407 44.25 0.000
4 0.936 0.876 0.033 0.872 0.362 37.54 0.000
5 0.945 0.893 0.017 0.889 0.337 22.12 0.000
6 0.951 0.903 0.010 0.899 0.322 14.60 0.000
7 0.954 0.909 0.006 0.905 0.313 8.72 0.004
8 0.954 0.911 0.001 0.905 0.312 2.05 0.154
9 0.955 0.911 0.001 0.905 0.312 1.19 0.278
Volume II Flood Frequency Estimation
23
987654321
1.0
0.8
0.6
0.4
0.2
0.0
Number of PCDs in model
0.9
Figure 2.6: Improvement of r
2 for a model size of one to nine variables
Based on the 145 catchments in the calibration set, the following 7-variable model was
selected for use on essentially rural catchments:
ℓnQMED = –11.300 + 0.937ℓnAREA – 0.922ℓnBFIsoil + 1.306ℓnSAAR +
2.217ℓnFARL + 0.341ℓnDRAIND + 0.185ℓnS1085 + 0.408ℓn(1+ARTDRAIN2) 2.7
The model has an r2 of 0.909 and a root mean square error (RMSE) of 0.313. Exponentiat-
ing, the model can be written:
0.4080.1850.341
2.2171.3060.922
soil
0.9375
ARTDRAIN21S1085DRAIND
FARLSAARBFIAREA101.237QMED
2.8
The factorial standard error (FSE) in estimating QMED is therefore e0.313
= 1.37. This tells us
that, under standard assumptions, 68% of QMED estimates can be expected to lie within the
range 1/1.37 to 1.37 (i.e. 73% to 137%) of the true value. [Editorial note: This is a
strikingly good performance. While there is no directly comparable equation in the FSR
(NERC, 1975), the FEH QMEDrural model (for UK data) has a considerably wider error-band,
with FSE = 1.55 (Robson, 1999b). A revised FEH QMED model (Kjeldsen et al., 2008)
improves this to FSE = 1.43.]
2.2.6 Model performance
Figure 2.7 plots the modelled and observed QMED values and shows a good fit with little
evidence of heteroscedacity (uneven error variance). The labelled stations are discussed later.
Table 2.7 confirms that all coefficients are significant at the 0.05 level (│t statistic│> 1.96).
The standardised coefficients (β in the table) highlight the relative contribution of each
descriptor to explaining the variation in ℓnQMED. As expected, ℓnAREA is by far the most
important predictor.
Co
effi
cien
t o
f d
eter
min
atio
n,
r2
Volume II Flood Frequency Estimation
24
50020010050201052
500
200
100
50
20
10
5
2
Observed QMED (m s )
Pre
dic
ted Q
ME
D1:1 line
3 -1
Station 06030
Station 10004
Station 19046
Figure 2.7: Observed and modelled QMED for the 145 calibration stations
Table 2.7: Coefficient and collinearity statistics for selected rural model for ℓnQMED
Term/regressor Coefficient Standard
error
β
value
t
statistic
95% confidence
interval Variance
inflation
factor (VIF) Lower Upper
Constant -11.300 1.15 -9.82 -13.58 -9.02
ℓnAREA 0.937 0.03 1.02 29.46 0.87 1.00 1.80
ℓnSAAR 1.306 0.17 0.29 7.54 0.96 1.65 2.22
ℓnFARL 2.217 0.33 0.22 6.67 1.56 2.87 1.62
ℓnBFIsoil -0.922 0.17 -0.18 -5.46 -1.26 -0.59 1.70
ℓnS1085 0.185 0.04 0.18 4.41 0.10 0.27 2.45
ℓnDRAIND 0.341 0.07 0.14 4.85 0.20 0.48 1.34
ℓn(1+AD2) 0.408 0.14 0.08 2.95 0.13 0.68 1.11
In order to assess the possible impact of collinearity (see Box 2.1), an additional statistic was
studied: the variance inflation factor (VIF). This is the reciprocal of the tolerance, which in
turn denotes the proportion of the variance in a given catchment descriptor that cannot be
explained by the other regressors.
High VIF values (i.e. small tolerances) indicate that a large amount of the variance in one
regressor can be explained by the other regressors. VIF thus indexes the impact of
collinearity (amongst the regressors) on the stability of the multiple regression model. VIF
values are (by definition) greater than or equal to 1. Whilst only a guide, VIF values greater
than 10 are often regarded as indicating serious problems of collinearity. In weaker models,
values above 2.5 may sometimes be a cause for concern.
It is seen from the final column of Table 2.7 that VIF is less than two for all regressors
excepting ℓnSAAR and ℓnS1085. The modest correlation (rank correlation ρ = 0.30)
Volume II Flood Frequency Estimation
25
between these two PCDs leads them to compete in explaining the variation in ℓnQMED.
This is because higher values of SAAR and S1085 both tend to be associated with upland
areas, and lower values with lowland areas. That their VIF values are reasonably healthy (i.e.
less than 2.5) supports the retention of both descriptors in the model, and suggests that
ℓnS1085 is adding something useful.
2.2.7 Checking for logical consistency
Logical consistency – in the sense of being in accord with common reasoning – is often the
overriding factor in the final choice of a regression model. The 7-variable model for
ℓnQMED on rural catchments performs well in this respect:
QMED increases with AREA;
QMED increases with greater wetness (indexed by SAAR);
QMED increases with FARL, meaning that it decreases for increased attenuation;
QMED decreases with greater permeability (indexed by BFIsoil);
QMED increases with catchment steepness (indexed by S1085);
QMED increases with drainage density (DRAIND);
QMED increases with the extent of arterial drainage (indexed by ARTDRAIN2).
All these features are logically consistent with what is known about flood behaviour.
2.2.8 Checking and investigating the model residuals
The ordinary least-squares (OLS) approach to judging and testing the model requires that the
residuals are Normally-distributed with constant variance. The plots of Figure 2.8 suggest
that the residuals to the ℓnQMED model are well behaved. Visually, the fit to the assumed
Normal distribution is good, even in the tails of the distribution. There is little evidence of
changes in variance with increasing ℓnQMED.
Figure 2.8: Diagnostic plots of ℓnQMED model performance
0.0 0.2 0.4 0.6 0.8 1.0
Observed cumulative probability
Ex
pec
ted c
um
ula
tive
pro
bab
ilit
y
0.0
0.2
0.4
0
.6
0.8
1
.0
0 1 2 3 4 5 6
Fitted ℓnQMED
Res
idual
(on ℓ
n s
cale
) -0
.5 0
.0
0.5
1
.0
(a) Normal quantile-quantile (b) Residual versus fitted
1:1 line
Volume II Flood Frequency Estimation
26
Map 2.2 indicates the residuals from the 7-variable rural QMED model. The small grey
symbols mark the catchments on which the model performs particularly well. It is seen from
the graduated symbols that there is some semblance of a tendency for the model to over-
estimate ℓnQMED in the Midlands and West, and to underestimate it in the East and South.
Map 2.2: Factorial error in QMED estimated by selected rural model
In analysing the residuals further, Figure 2.9 shows scatter-plots of the selected rural model
against each of the seven selected catchment descriptors. Interest centres on:
Examining the relationship between residuals and PCDs to assess the success of the
model in capturing the range of catchment types represented;
Identifying the possible presence of a curved pattern in the residuals when plotted
against any descriptor (curvature would indicate non-linear relationships and suggest
the need to include additional transformations of the PCDs in the model-building).
● Model underestimates QMED ● Model overestimates QMED
Factorial error
in QMED
Volume II Flood Frequency Estimation
27
6.90784.60522.3026
1.0
0.5
0.0
-0.5
-1.0
-0.1054-0.5108-1.2040 7.60097.31326.9078
0.0000-0.2231-0.5108
1.0
0.5
0.0
-0.5
-1.0
0.69310.0000-0.6931-1.3863 3.21891.60940.0000-1.6094 0.69310.40550.0000
Figure 2.9: Residuals versus individual PCDs for rural ℓnQMED model
In terms of curvature, the residual plots are found to be well behaved, with little evidence of
non-linear relationships between the residuals and the PCDs. In examining model
performance for the range of catchment types, particular interest was directed at how well the
model performs for permeable catchments (with high BFIsoil) and how well the full range of
catchment sizes is represented.
Res
idual
(on ℓ
n s
cale
)
AREA (km2) BFIsoil SAAR (mm)
10 100 1000 0.3 0.6 0.9 1000 1500 2000
FARL DRAIND (km/km2) S1085 (m/km) ARTDRAIN2
0.6 0.8 1 .25 .5 1 2 0.2 1 5 25 0 0.5 1
Res
idual
(on ℓ
n s
cale
)
Volume II Flood Frequency Estimation
28
Catchments smaller than about 50 km2 are not well represented in either the calibration or
validation dataset (see Histograms 2.3 in Section 2.2.3). From the spread of residuals in the
AREA plot in Figure 2.9, the selected rural model performs somewhat better for larger
catchments. However, there is a hint (from the positive residuals) of a tendency towards
underestimation on the very largest catchments.
The weaker performance on smaller catchments is a typical finding of such studies e.g.
Kjeldsen et al. (2008). Inevitably this prompts concern, given that many applications of
PCD-based models are to smaller catchments.
2.2.9 Validation of model performance
The performance of the rural ℓnQMED model was validated by application to records held
back for the purpose (see Section 2.2.3). The 25 randomly selected stations provide a
relatively robust method of assessment.
The outcome of the test is fully satisfactory. Figure 2.10 shows observed and predicted
QMED values for the 25 validation stations. The coefficient of determination obtained is
r2 = 0.906. The absence of heteroscedacity (uneven error variance) in Figure 2.10 is
reassuring, and vindicates the approach taken to fitting the QMED model: namely, OLS
regression on ℓn-transformed variables.
The validation eases concern over possible weak performance on small catchments. There is
no evidence from Figure 2.10 that the QMED model performs worse on catchments smaller
than 100 km2 than on catchments larger than 300 km
2.
50020010050201052
500
200
100
50
20
10
5
2
Observed QMED (m s )
Pre
dic
ted Q
ME
D
1:1 line
3 -1
Station 25017
Figure 2.10: Observed and modelled QMED for the 25 validation stations
[Editorial note: An incorrect value of AREA appears to have been used for Station 25017
and accounts for the disappointing performance on this catchment. Overall, the model
performance on the validation set is excellent.]
AREA < 100 km2
100 ≤ AREA < 300 km2
AREA ≥ 300 km2
Volume II Flood Frequency Estimation
29
2.3 Assessing model robustness
2.3.1 Model robustness
Model robustness is an inevitable concern when a 7-variable model has been fitted to a
dataset of 145 observations. Stringent tests were therefore performed to make sure that the
inclusion/omission of certain catchments did not impact significantly on model coefficients.
In general terms, the robustness of a prediction model can be thought of as its ability to
remain stable against external disturbances. One concern is that the presence of an unusual
catchment in the calibration dataset might exert undue influence on the model derived.
2.3.2 Bootstrap resampling
As a prelude to testing the robustness of the model to the inclusion/omission of particular
catchments, the coefficients of the rural ℓnQMED model were “bootstrapped”. In conducting
the resampling, 1000 new samples – each of the same size as the observed data – were drawn
with replacement from replications of the observed data.
The model coefficients were first calculated for the observed data and then recalculated for
each of the 1000 resamples. This yielded the confidence intervals (for the model
coefficients) shown in Table 2.8. The BCa method constructs confidence intervals using the
“Bias-Corrected and accelerated” bootstrap introduced by Efron (1987). This adjusts for bias
and for skewness in the bootstrap distribution.
Table 2.8: Bootstrapped confidence intervals for model coefficients (BCa method)
Boot-strapped
confidence intervals 2.5% 5%
OLS
result 95% 97.5%
Constant -13.70 -13.26 -11.300 -9.73 -9.41
ℓnAREA 0.86 0.87 0.937 0.98 0.99
ℓnBFIsoil -1.21 -1.17 -0.922 -0.45 -0.36
ℓnSAAR 1.05 1.09 1.306 1.63 1.70
ℓnFARL 1.65 1.77 2.217 2.76 2.86
ℓnDRAIND 0.26 0.29 0.341 0.65 0.70
ℓnS1085 0.08 0.09 0.185 0.23 0.24
ℓn(1+AD2) 0.12 0.18 0.408 0.63 0.68
Figure 2.11 shows the Normal quantile-quantile plots of the bootstrapped model coefficients.
Volume II Flood Frequency Estimation
30
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
-14
-12
-10
-8(Intercept)
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
0.8
50
.90
0.9
51
.00
1.0
5
LnArea
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
-1.4
-1.2
-1.0
-0.8
-0.6
-0.4
-0.2
LnBFIsoils
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
0.8
1.0
1.2
1.4
1.6
1.8
LnSAAR
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
1.5
2.0
2.5
3.0
3.5
LnFARL
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
0.2
0.4
0.6
0.8
LnDRAIND
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
0.0
0.0
50
.10
0.1
50
.20
0.2
5
LnS1085
Qu a n ti l e s o f Sta n d a rd No rm a l
Qu
an
tile
s o
f R
ep
lica
tes
-2 0 2
-0.2
0.0
0.2
0.4
0.6
0.8
Ln1ARTDRAIN2
Figure 2.11: Normal quantile-quantile plots of the bootstrapped model coefficients
2.3.3 Model robustness to influential stations
The influence of individual catchments on the final model coefficients was assessed using a
technique referred to as “Jackknife after bootstrap”. Jack-knifing recalculates model
coefficients exhaustively: omitting each data point (i.e. each catchment) in turn.
In testing the sensitivity of model coefficients to the data they were trained on, the model was
held to be overly sensitive if the removal of any individual catchment from the calibration
dataset resulted in coefficients becoming insignificant or falling outside the 95% confidence
intervals of the BCa percentiles. Groups of influential catchments were also considered.
Figure 2.12 depicts the absolute relative influence of individual observations on the model
formulation. Observations with absolute relative influence values greater than 2.0 were
selected for further testing. In testing the sensitivity of each parameter, the model was rerun
with influential observations omitted sequentially without replacement. The significance of
the resulting change in the model parameter was noted.
Intercept ℓnAREA ℓnBFIsoil ℓnSAAR V
erti
cal
scal
es m
ark
qu
anti
les
of
mo
del
coef
fici
ent
ℓnFARL ℓnDRAIND ℓnS1085 ℓn(1+AD2)
Horizontal scales mark quantiles of standard Normal distribution
Volume II Flood Frequency Estimation
31
(Intercept)
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
5
2 4
9 19 5
1 0 1
1 2 8
LnArea
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
5
2 4
5 5 7 5
8 5
9 1
1 0 7
1 0 91 3 9
LnBFIsoils
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
5
2 44 6
4 9
5 4
5 5
9 2
LnSAAR
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
2 4
9 1
9 5
1 0 1
1 2 8
1 3 8
LnFARL
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
5
2 44 9
8 81 0 7
1 0 9
1 2 8
LnDRAIND
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
5
2 4 7 5 9 09 39 5
1 0 7
1 2 21 3 7
LnS1085
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
24 2 43 0 4 96 6
8 5
8 79 11 3 1
Ln1ARTDRAIN2
Ob s e rv a ti o n
Ab
solu
te R
ela
tive
In
flu
en
ce
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0
02
46
2
4
1 76 26 66 7
9 01 0 71 0 9 1 4 4
Figure 2.12: Influence of individual stations in determining model coefficients
[Editorial note: Particular catchments (e.g. Sequential Numbers 5, 24, 49 and 109) are seen
to be rather influential. Sequential Numbers 5, 24 and 49 correspond respectively to
Stations 06030 Big at Ballygoly, 10004 Glenmacnass at Laragh and 19046 Martin at Station
Road. These are all stations for which the 7-variable model underestimates QMED
appreciably (see Figure 2.7). The AM flood data for Station 06030 Big at Ballygoly appear
suspect, although this may not compromise the gauged value of QMED. See Box 4.3.
Sequential Number 109 corresponds to Station 31002 Cashla at Cashla, for which QMED is
well-predicted by the model. This station is influential on the exponent of FARL because the
catchment has the greatest attenuation (FARL=0.632) of any within the calibration dataset.]
It was found that, even when all influential points were removed, the eight model coefficients
(i.e. the intercept term and the seven coefficients in Equation 2.7) remained significant and
within the BCa percentiles given in Table 2.8. It was therefore concluded that the QMED
model is not unduly sensitive to individual observations.
2.3.4 Robustness to spatial drift in model coefficients
An underlying assumption of the regression approach – undertaken here by ordinary least-
squares – is that the derived relationship is spatially constant, i.e. that the estimated
parameters remain constant over space. In hydrological modelling applications, this
assumption of spatial invariance can sometimes be unconvincing.
Geographically weighted regression (GWR) is a technique that expands standard regression
for use with spatial data (Fotheringham et al., 2002). GWR assesses local influences and
Labels (and horizontal scales) mark sequential number in calibration dataset of 145 catchments
Ver
tica
l sc
ales
mar
k “
Act
ual
rel
ativ
e in
flu
ence
”
Intercept ℓnAREA ℓnBFIsoil ℓnSAAR
ℓnFARL ℓnDRAIND ℓnS1085 ℓn(1+AD2)
Volume II Flood Frequency Estimation
32
allows a spatial shift in parameters. This can provide a more appropriate and accurate basis
for descriptive and predictive purposes. Some details are given in Section C4.
The GWR technique was employed to test if the coefficients of the ℓnQMED model are
spatially constant. The assumption of spatial invariance was found reasonable except in
regard of the FARL term. There is some limited evidence (see Map C.1 in Section C4) that a
higher coefficient of ℓnFARL may be appropriate in the East and North West, while a lower
coefficient may be apt around the Upper Shannon basin and in the West.
This test result was of borderline significance, and use of a spatially-varying model for
ℓnQMED is not recommended. However, the finding endorses the recommendation to
favour geographical closeness – as well as similarity in key PCDs such as FARL – when
selecting the pivotal catchment in a particular flood estimation problem (see Section 11.2).
2.4 Investigating the effect of arterial drainage
2.4.1 Approach
The effect of arterial drainage on QMED is represented by the appearance of ARTDRAIN2
in the 7-variable model of Equation 2.8 in Section 2.2.5. However, arterial drainage is a
sufficiently important feature of Irish flood hydrology to warrant more explicit investigation.
Its impact on the incidence of flooding downstream has long been a source of controversy,
with the opposing points of view well highlighted by Robinson (1990).
The further work took two main forms. Section 2.4.2 looks at stations for which both pre and
post-drainage records are held, while Section 2.4.3 considers the separate modelling of
QMED on drained and undrained catchments.
2.4.2 Analysis of stations for which both pre and post-drainage records held
Amongst the flood peak data studied in the FSU, 15 stations have both pre and post-drainage
records. Gauged values of QMED are shown in Table 2.9 and Figure 2.13.
Table 2.9: Factorial change in QMED following drainage
Station
#
QMED (m3s
-1)
Factorial
change
Station
#
QMED (m3s
-1)
Factorial
change Pre-
drainage
Post-
drainage
Pre-
drainage
Post-
drainage
03051 21.50 40.10 1.865 24004 39.27 62.41 1.589
07002 17.91 19.22 1.073 25017 391.74 414.17 1.057
07003 12.71 21.87 1.721 26012 29.61 47.68 1.610
07005 86.10 104.98 1.219 30004 42.30 90.34 2.136
07007 37.15 35.70 0.961 30005 22.88 36.78 1.608
07010 32.87 70.72 2.152 30061 247.97 250.07 1.008
07012 149.61 265.86 1.777 35011 86.71 132.23 1.525
24001 80.84 114.59 1.417 15 FSU catchments, pre and post-drainage
Volume II Flood Frequency Estimation
33
It is evident that, following arterial drainage, there are substantial increases in QMED at 11 of
the 15 stations. QMED is found to more than double at Stations 07010 and 30004. In
contrast, Station 07007 shows a slight reduction, while Stations 25017 and 30061 show little
or no change in QMED following drainage.
Figure 2.13: QMED for pre and post-drainage records at 15 stations
Disappointingly, there is no discernible association between the factorial change and either of
the PCDs summarising the extent of arterial drainage (see Figures 2.14).
0.50.40.30.20.10.0
2.5
2
1.5
1
0.5
ARTDRAIN
Facto
rial change
1.00.80.60.40.20.0
2.5
2
1.5
1
0.5
ARTDRAIN2
Facto
rial change
Figures 2.14: QMED change following drainage (against ARTDRAIN and ARTDRAIN2)
The one significant association found – the trend is just significant at the 95% level – was
with BFIsoil (see Figure 2.15). It is logically consistent that drainage is somewhat more
effective on naturally less permeable (i.e. low BFIsoil) catchments.
Pre-drainage QMED (m3s
-1)
Post
-dra
inag
e Q
ME
D (
m3s-1
)
- - - - 100% increase – – – 50% increase ––––– no change
Volume II Flood Frequency Estimation
34
1.00.80.60.40.20.0
2.5
2
1.5
1
0.5
BFI
Fac
tori
al c
han
ge
35011
30061
30005
30004
26012
25017
24004
24001
07012
07010
07007
07005
07003
07002
03051
soil Figure 2.15: QMED change following arterial drainage (against BFIsoil)
2.4.3 Modelling QMED on drained and undrained catchments separately
Although represented in the QMED model of Equation 2.8, the effect of arterial drainage was
explored further by partitioning the 145-catchment rural dataset into 50 drained catchments
and 95 undrained catchments. The aim was to assess whether explicit discrimination of
drained and undrained catchments could:
Enhance the overall modelling of QMED from PCDs;
More fully reveal the impact that arterial drainage has on QMED;
Identify the PCDs that best capture the drained and undrained responses.
The further analysis is reported in Appendix D, where it is concluded that there is little gain
in modelling QMED separately on drained and undrained catchments.
However, the analysis suggests that BFIsoil and ARTDRAIN2 are valued descriptors in
predicting QMED in drained catchments, whilst DRAIND and S1085 are more important in
predicting QMED in undrained catchments. Awareness of these sensitivities may help the
experienced user to judge the most appropriate choice of pivotal catchment in a given flood
estimation problem (see Section 11.2).
2.5 Adjusting the selected model for urbanisation
2.5.1 Introduction
The rainfall-runoff response of a catchment can be radically altered by urbanisation.
Impervious surfaces inhibit infiltration and reduce surface retention. Increases in surface
runoff are typically accompanied by a more intense and accelerated response.
In assessing the effect of urbanisation at the river-catchment scale, the approach taken was to
study the change from the natural (rural) condition. The outcome is an adjustment factor that
Volume II Flood Frequency Estimation
35
can be used to augment the performance of the rural model for use on catchments that have
undergone urbanisation.
The adjustment derived is strictly for catchments that have undergone urbanisation. For
reasons discussed by Reed and Robson (1999), such models are not suitable for anticipating
the effect of planned urban development.
The modelling of QMED in Section 2.2 treated catchments with a fractional urban extent
(URBEXT) less than 0.015 as essentially rural. The urban dataset thus comprises the 35
catchments in the FSU dataset having URBEXT≥0.015.
2.5.2 Deviation of QMED on urbanised catchments from rural model
Figure 2.16 summarises the outcome of applying the rural QMED model to the 35 urbanised
catchments. Estimation of QMED from PCDs is inevitably approximate, so it is difficult to
infer a definitive influence of urbanisation. As anticipated, the rural model underestimates
QMED on more catchments than it overestimates (20 as against 15). Regression confirms the
general tendency for the rural model to underestimate QMED when applied to urbanised
catchments.
200100502010521
200
100
50
20
10
5
2
1
Observed QMED (m s )
QM
ED
est
imat
ed b
y r
ura
l m
odel
08005
09010
08009
3 -1
1100108008
Figure 2.16: Performance of rural QMED model on urbanised catchments
Five catchments (labelled in Figure 2.16 and listed in Table 2.10) have a factorial error
greater than 2, although only marginally. QMED at Stations 08005, 08008, 09010 and 11001
is underestimated by a factor of just over 2, while QMED at Station 08009 is overestimated
by a factor of just over 2 (see central column of Table 2.10). [Editorial note: Station 08009
Ward at Balheary attracts comment elsewhere, notably in Box 4.4 and Section 10.4.5.
Bhattarai and Baigent (2009) suggest that the surprisingly low gauged QMED may reflect
incomplete recording. While urbanisation may account for the underestimation at Stations
08005 and 09010, it is likely that some other factor (e.g. extensive arterial drainage) accounts
for the underestimation of QMED at Stations 08008 and 11001.]
Regression
line
Volume II Flood Frequency Estimation
36
Table 2.10: Five catchments poorly predicted by the QMEDrural model
Station
# QMEDobs QMEDrural
QMEDrural ∕ QMEDobs
URBEXT AREA ARTDRAIN2
08005 2.32 1.09 0.470 0.25 9.17 0.00
08008 40.90 19.97 0.488 0.04 107.92 0.44
08009 5.00 10.14 2.028 0.08 61.64 0.55
09010 47.05 22.33 0.475 0.24 94.26 0.00
11001 47.17 21.84 0.463 0.02 155.11 0.36
2.5.3 A special check
Catchments liable to urbanisation are not necessarily typical of the overall set of Irish gauged
catchments. A special check was made by seeking a set of 35 rural catchments that match
the 35 urbanised catchments in respect of catchment features other than urbanisation.
The experiment defined catchment similarity as a standardised distance in the 4-dimensional
space representing differences in values of ℓnAREA, ℓnSAAR, ℓnBFIsoil and ℓnFARL.
Because of the limited number of catchments available, some rural catchments were allowed
to appear more than once in the set of 35 rural catchments formed to match the 35 urbanised
catchments. The most problematic feature to match was catchment wetness (SAAR).
[Editorial note: On average, the matched catchments tend to be somewhat wetter than their
urbanised counterparts. This reflects that settlement patterns are biased towards drier parts of
Ireland.]
The special analysis (not reported) found that the rural QMED model of Section 2.2
represents QMED on the matched set of catchments relatively well, without notable bias. It
was therefore concluded that it was reasonable to use the 7-variable rural model as a vehicle
for estimating the rural element of QMED on urbanised catchments.
2.5.4 Approach to devising an urban adjustment
Despite the difficulty in detecting a clear fingerprint of urbanisation, effort was made to
derive an adjustment factor for urbanisation that could be used to scale up the rural model
results using the form:
ruralQMEDUAFQMED 2.9
UAF is an urban adjustment factor that describes the proportional increase in QMED induced
by urbanisation. QMEDrural denotes the rural estimate for QMED by the Section 2.2 model.
Values of UAF at the gauged sites are inferred by inverting Equation 2.9:
ruralQMED
QMEDUAF 2.10
Volume II Flood Frequency Estimation
37
2.5.5 Exploratory data analysis
UAF was ℓn-transformed and its relationship with catchment descriptors examined. Figure
2.17 shows scatter-plots and correlation coefficients across ℓnUAF and a selection of
catchment descriptors.
Figure 2.17: Correlations between ℓnUAF and selected PCDs
It is seen that ℓnUAF is not strongly correlated with the PCDs. The strongest relationships
found are with ℓn(1+URBEXT), ℓnAREA and ℓnSAAR. The marked negative correlations
between ℓnUAF and ℓnAREA, and between ℓn(1+URBEXT) and ℓnAREA, reflect that the
most heavily urbanised catchments tend to be small catchments.
2.5.6 An urban adjustment model
Bearing in mind the above relationships, several approaches to modelling UAF were
examined beginning with the basic model form of:
URBEXT)n(1gnUAF 2.11
The model was fitted by weighted least-squares (WLS) regression with the weights
proportional to the urban extent. Thus, greater weight is given to data from more heavily
urbanised catchments. This led to the model:
ℓnUAF = 1.482 ℓn(1+URBEXT) 2.12
and its multiplicative form:
1.482URBEXT)(1UAF 2.13
Lower triangle shows scatter-plot. Upper triangle shows correlation coefficient.
ℓnUAF -0.48 0.54 0.13 -0.10 -0.39
ℓnAREA -0.68 -0.03 0.21 0.16
ℓn(1+URBEXT) 0.09 0.07 -0.43
ℓnFARL -0.42 -0.44
ℓnBFIsoil -0.09
ℓnSAAR
Volume II Flood Frequency Estimation
38
The model returned an r2 (in ℓn space) of 0.300, a standard error of 0.735 and a factorial
standard error of 2.09. The coefficient 1.482 has a standard error of 0.139. The UAF model
has the merit of decreasing to 1.0 when URBEXT decreases to zero. The model returns a
value of 2.793 when URBEXT reaches a maximum of 1 (for a fully urbanised catchment).
The value of g in this model is very similar to the coefficient derived for the corresponding
model in the FEH, which gave a value of 1.49 (Table 18.1 of Reed and Robson, 1999). The
large factorial standard error highlights the large uncertainties involved in modelling UAF.
Nonetheless the model represents a theoretically plausible description of the impact of
urbanisation on the index flood and can be interpreted as:
Urban adjustment factor increases with urban extent;
Urban adjustment factor increases to a maximum of 2.793 when a catchment is fully
urbanised (URBEXT=1);
Urban adjustment factor decreases to 1 as URBEXT tends towards zero.
2.5.7 Performance of urban adjustment model
Figure 2.18 shows the performance of the urban-adjusted model on the 35 urbanised
catchments. In comparison with Figure 2.16 – which shows the performance of the
unadjusted model on these catchments – a small but just perceptible improvement is evident,
not least for Stations 08005 and 09010.
200100502010521
200
100
50
20
10
5
2
1
Observed QMED (m s )
QM
ED
model
adju
sted
by U
AF
08005
3 -1
1100108008
08009
09010
Figure 2.18: Performance of UAF-adjustment to QMED
Urbanisation tends to have a greater effect on naturally permeable catchments. However, it
was not practical to consider a permeable-catchment adjustment given the very small number
of such stations in the available dataset. It is recommended that additional gauges be
established in such catchments.
Volume II Flood Frequency Estimation
39
2.5.8 Comparisons and contrasts with the FEH
The approach to modelling the urban adjustment mirrors some aspects of the methods of
Reed and Robson (1999). Important differences in the Irish study are that:
The smaller pool of gauged catchments and the narrower variation of catchment
permeability led to adoption of a simpler model structure in which UAF is determined
solely by the urban fraction URBEXT;
The basis of URBEXT values differs between Ireland and the UK, impeding
comparisons;
Values of URBEXT have not been backdated to represent the state of catchment
development at the mid-point of the period of record used to calculate QMED at the
gauged site;
Values of UAF inferred from Equation 2.10 were specifically constrained to be no
less than 1.0.
[Editorial note: Associating the long-term value of QMED with the current (higher) value
of urban extent will tend to underplay the effect of urbanisation on flood magnitudes.
However, the device of constraining “observed” UAF values to be no less than 1.0 may do
much to compensate.]
Box 2.2: Interpretation of urban adjustment models for QMED
2.6 Improving model performance by data transfer
2.6.1 Introduction
Even though the selected rural model marks a major improvement on previous approaches for
Ireland, the uncertainty ranges are still large. Every effort should be made to adjust the
QMED estimates made from PCDs by reference to flood data available from nearby or
similar catchments.
2.6.2 Review of techniques
In terms of adjusting model predictions based on the transfer of information from gauged
sites, there has been debate within the literature as to whether adjustments should be made
Editorial note: As with the UK work of Reed and Robson (1999), the worry lingers that
the urban adjustment to QMED might be mistaken for representing the total effect of
urbanisation. It should be recalled that the urban adjustment factor represents only the
residual effect of urbanisation on flood magnitudes. It does not represent the effect that
past development-control has successfully ameliorated.
The urban adjustment model has been fitted to the gauged flood data just as they are.
Thus, the model implies that typically this amount of urbanisation produces this amount of
an increase in flooding if development control continues to be as effective as it has been in
the past. It would be negligent to interpret the urban adjustment model as representing the
total effect of urbanisation on what would have hitherto been a greenfield catchment.
Volume II Flood Frequency Estimation
40
using information from catchments that are geographically close to the site of interest
(sometimes called donor catchments) or from catchments that are hydrologically similar in
terms of key descriptors but located anywhere within the study domain (sometimes called
analogue catchments). Previous work has tended to highlight the strong clustering of
residuals in regression models and used this to underscore the recommendation to favour
local catchments in QMED adjustments.
In a comprehensive assessment of the FEH statistical method for adjusting QMED values,
Morris (2003) found inappropriate adjustment of QMED to be a potentially important source
of error and suggested that the selection of gauges for the transfer of information be based on
catchment similarity judged from key PCDs. However, the selection and use of similar (as
opposed to local) catchments is subjective and can greatly affect the QMED estimate. Morris
also concluded that favouring donor sites that are located directly upstream/downstream of
the subject site could potentially help to reduce prediction errors further. In contrast Kjeldsen
et al. (2008) favour a method where the weight given to a particular data transfer is based on
geographical distance, rather than expressly in terms of catchment similarity.
This section summarises some work done on an automated method of data transfer before
explaining why more traditional methods were chosen for implementation. Section 2.7
provides a worked example of QMED adjustment by data transfer. The topic is taken up
further in Chapter 11.
2.6.3 Geostatistical mapping of residuals
Grover et al. (2002) indicate that the performance of global regression models can be
improved by mapping regression residuals using geostatistical methods and using these
mapped residuals to adjust QMED estimates at point locations. To complement the
traditional approaches reviewed above, geostatistical mapping of residuals was considered:
To explore the spatial pattern of errors in the estimation of ℓnQMED;
As a basis for adjusting QMED.
Geostatistical methods were used to interpolate and map the residuals of the rural 7-variable
model. The interpolation methods considered included Kriging, splines and inverse distance
weighting. Based on the assessment of a small validation set, the inverse distance weighting
(IDW) technique was found to be the most appropriate.
The IDW function determines interpolated values using a linear weighted combination of a
set of sample points. The weight assigned to each is a function of the distance of an input
point from an output cell location. The greater the distance, the smaller the influence the
point has on the output value.
A fixed radius of 55 km was adopted to select input stations for modelling ℓnResidual.
Interpolated residuals are shown in Map 2.3. Areas of underestimation of ℓnQMED are
indicated in red, while the blue-grey areas mark overestimation of ℓnQMED. Some
clustering of model errors is evident, with areas of overestimation of ℓnQMED in the South
East, much of the North West, the Mid-West and the South West, and areas of
underestimation in much of the East and South of the country.
Volume II Flood Frequency Estimation
41
Map 2.3: IDW-interpolated residuals from rural ℓnQMED model
[Editorial note: The extent to which the geostatistical mapping approach is driven by the
residual error for particular gauged catchments can be judged by comparing Map 2.3 with
Map 2.2 presented in Section 2.2.8. Regrettably, the colour schemes are transposed. The
overestimation of QMED around Wexford is largely based on the residual for Station 13002
Corock at Foulk’s Mill (QMEDobs = 7.01 and QMEDest = 10.74 m3s
-1). Here, QMEDest
denotes the estimate of QMED from the rural catchment model of Equation 2.8.]
ℓn(Residual)
0.9 – 1.1
0.7 – 0.9
0.5 – 0.7
0.3 – 0.5
0.1 – 0.3
0.0 – 0.1
-0.1 – 0.0
-0.3 – -0.1
-0.5 – -0.3
-0.7 – -0.5
Wexford
Volume II Flood Frequency Estimation
42
In order to extract a correction value from the interpolated map, the centroid of the subject
catchment is overlaid and the interpolated ℓnResidual extracted. The value is then added to
the prediction from the rural ℓnQMED model.
The method was assessed across the 25 validation stations. Whereas the basic model
achieved a coefficient of determination of 0.906 for ℓnQMED (see Section 2.2.9), applying
the IDW-adjustment increased r2 to 0.912. This suggests that the adjustment technique is
moderately effective.
It is recommended that additional monitoring stations are sought in districts where residuals
are uncomfortably large (i.e. in darker-shaded areas of Map 2.3), especially where Map 2.2
indicates the underlying data to be sparse.
2.6.4 Possible disadvantages of automated methods
Automated methods of adjusting estimates are becoming more commonplace in flood
hydrology. Here, the IDW method has been shown (by split-sample testing) to be moderately
effective in adjusting the PCD-based estimates of QMED.
However, the routine adoption of such automated methods comes at a price. It rules out the
use of experience and reasoning to make subjective but informed decisions about appropriate
adjustments. Factors that the human adjuster might consider include:
The perceived degree of similarity between donor and subject catchments;
The likely quality of the gauged estimate of QMED at the donor site;
Whether to involve more than one donor in the data transfer.
There is some support from Morris (2003) for the traditional view that data transfer from an
upstream or downstream site is often of greatest value. This is further supported by an
interim assessment reported in Section 11.1.
It should be noted that the automated method for adjusting QMED by geostatistical mapping
is executed directly in space by reference to distances between catchments. The method does
not map the residual error (in ℓnQMED) up and down a particular river system.
The distances used in the geostatistical mapping method are measured from catchment
centroid to catchment centroid rather than from gauged site to subject site. In cases where the
nearby gauges all lie in the same overall river basin, these inter-centroid distances reflect
something of the river structure, albeit imperfectly.
A particular weakness of an automated method can be a failure to consider “scale” effects.
Should data from a small tributary be used to adjust QMED values in the much larger river to
which it drains? And is a gauging station on the main river suitable for adjusting QMED on a
minor tributary? These are difficult questions to answer algorithmically, rather than by
individual judgement of the particular circumstances.
The FSU concludes that the geostatistical mapping approach embodied in Map 2.3 is useful
for investigating the spatial characteristics of model residuals. However, it does not
recommend its adoption. Instead, the user is exhorted to choose (and make) a data transfer
Volume II Flood Frequency Estimation
43
based on reasoned judgement of the situational context of their flood estimation problem in
relation to the available gauged data. The judgement should take account both of the
characteristics of the particular subject site and of the options available for data transfer.
2.6.5 Recommended procedure for data transfer
It is advocated that the user assesses the most appropriate data transfer by making a reasoned
selection of a pivotal catchment. This is the gauged catchment judged to be most relevant to
the specific flood estimation problem.
It is also a matter for the user to judge whether part or all of the observed ℓnResidual at the
pivotal site is to be added to the catchment-descriptor estimate of ℓnQMED when making the
adjustment at the subject site.
Data transfer methods are discussed further in Chapter 11. But many of the principal
elements are introduced in the worked example that now follows.
2.7 Worked example of QMED estimation at an ungauged site
The catchments in the worked example are all essentially rural; their URBEXT values range
from 0.0021 to 0.0055. The recommended procedure for data transfer is more complicated
when the donor catchment is appreciably urbanised (see Box 2.3).
Box 2.3: Data transfer procedure when one of the catchments is urbanised
While it is recommmended to incorporate urban adjustments in every case, the goal of the
worked example here is to focus on the fundamentals of data transfer. The occasional
references to QMEDrural remind the user that the data transfer is to be applied to the inferred
rural element of QMED only, as indicated in Box 2.3.
2.7.1 Illustrative example
The methodology for estimating the index flood at an ungauged site is illustrated for the Suck
at Rookwood. This corresponds to Station 26002 but is treated here as an ungauged site. The
Editorial note: The FSU recommendation is that adjustment factors are derived from,
and applied to, the inferred rural element of QMED, i.e. to QMEDrural. The full procedure
is therefore to:
1. Infer values of QMEDrural at gauged sites by inverse application of the urban
adjustment model (i.e. Equation 2.10);
2. Assess adjustment factors as the ratio of gauged to modelled values of QMEDrural;
3. Make the data transfer from gauged to subject site, thereby adjusting the estimate
of QMEDrural at the subject site;
4. Apply the urban adjustment factor in the normal manner (i.e. Equation 2.9) to
obtain the adjusted estimate of QMED at the subject site.
In all cases, the urban adjustment factor UAF is given by Equation 2.13:
1.482URBEXT)(1UAF
Volume II Flood Frequency Estimation
44
following provides a step-by-step guide to deriving an estimate of the index flood at this
location.
Step 1 Derive coordinates for ungauged location: In this case the catchment centroid has an
Easting of 172050 and a Northing of 270500.
Step 2 Derive catchment descriptor information: The relevant PCDs are given in Table 2.11.
They derive from the digital datasets made available with the FSU.
Table 2.11: PCDs for Suck at Rookwood worked example
PCD value unit PCD value unit
AREA 641.45 km2 URBEXT 0.0026 –
DRAIND 0.799 km/km2 FARL 0.979 –
S1085 0.500 m/km BFIsoil 0.6036 –
SAAR 1067.03 mm ARTDRAIN2 (AD2) 0.000 –
Step 3 Apply rural QMED model
From Equation 2.8:
0.4080.1850.341
2.2171.3060.922
soil
0.9375
ARTDRAIN21S1085DRAIND
FARLSAARBFIAREA101.237QMED
408.0185.0341.0217.2306.1922.0937.05 000.01500.0799.0979.003.10676036.045.641102373.1
18796.09263.09540.090125928.19.426102373.1 5
= 58.93 m3s
-1
Step 4 Apply urban adjustment factor
From Equation 2.13:
1.482URBEXT)(1UAF
1.482(1.0026)UAF
=1.0039
Thus, the PCD-based estimate of QMED is reached by Equation 2.9:
ruralQMEDUAFQMED
QMED = 1.0039 × 58.93
= 59.16 m3s
-1
The gauged QMED for the Suck at Rookwood (Station 26002) is 56.56 m3s
-1. So
QMED at this site is overestimated by a factorial error of 59.16/56.56 = 1.046 or
4.6%. This represents a very good performance. However, the user will not know
this for an ungauged site!
Volume II Flood Frequency Estimation
45
Step 5 Transfer data from gauged locations to improve model prediction at subject site:
The general procedure is to infer an adjustment factor, AdjFac, by reference to the
performance of the PCD-based model of QMEDrural at a nearby gauged site:
PCDrural,
gaugedrural,
QMED
QMEDAdjFac 2.14
The adjustment is then partially or fully transferred to the subject site:
PCDrural,
h
adjustedrural, QMEDAdjFacQMED 2.15
The typical procedure is to apply a full transfer by setting the exponent h to 1.0. If
QMEDrural is found to be 20% greater than the PCD-based estimate, it is assumed that
the model will be similarly in error at the subject site. Thus, the estimate of
QMEDrural at the subject site is adjusted by multiplying by 1.20.
The exponent h can be thought of as the hardness of the data transfer. h=1 denotes a
full (or “hard”) transfer. A partial (or “softer”) transfer might set h=0.5. In this case,
if QMEDrural is found (by examining QMEDrural,gauged) at the donor site to be 20%
greater than given by the PCD model (i.e. QMEDrural,PCD) the estimate of QMEDrural at
the subject site is adjusted by multiplying by a factor of 1.200.5
or 1.095. [Editorial
note: The relatively complex notation here reflects that data transfers are made to the
rural element of QMED only. See Box 2.3 above.]
Much skill attaches to deciding which of several possible donor catchments is pivotal
to improving QMED estimation at the subject site. With gauged sites both upstream
and downstream of the subject site, the choice is not clear-cut for the Suck at
Rookwood. Table 2.12 summarises data transfers from the gauged sites upstream and
downstream of the subject site. A clear head is required to track the adjustments.
Table 2.12: Data transfers for Suck at Rookwood worked example
Method
Data
transfer
from
AREA ℓnAREA Gauged
QMED
Modelled
QMED Implied
factorial
adjustment
QMED
at subject site
Gauged Adjusted
km2 – m
3s
-1 m
3s
-1
No data
transfer 641.5 6.46 59.16
Upstream
donor
Station
26006 184.8 5.22 24.23 28.05 0.864 51.1
Downstream
donor
Station
26005 1085.4 6.99 93.21 87.39 1.067 63.1
Compromise 0.960 56.8
IDW
interpolated several 59.8
Analysis of
gauged data
Station
26002 56.56
Volume II Flood Frequency Estimation
46
2.7.2 Merging data transfers from two sites
The adjustment factors implied by the model performance upstream and downstream of the
subject site are seen to be contradictory: the PCD model is overestimating QMED at Station
26006 Suck at Willsbrook but underestimating it at Station 26005 Suck at Derrycahill. A
reasonable approach is to seek a compromise by taking a weighted geometric mean of the
adjustment factors:
w1
Donor2
w
Donor1Compromise AdjFacAdjFacAdjFac
2.16
The weight, w, is chosen to reflect the relative quality or relevance of the donors to QMED
estimation at the subject site. The value chosen for w might reflect the quality or length of
record at the gauged sites, or their inter-centroid distances from the subject catchment.
It transpires that the QMED estimates at 26005 and 26006 both derive from very long
records. The flood rating at Station 26006 is rated the more highly (Grade A1 as against
Grade A2 for Station 26005). However, the inter-centroid distance to the subject catchment
is shorter for Station 26005 (73 km rather than 129 km for Station 26006). The pragmatist
might decide to give the donor sites equal weight, setting w = 0.5. Thus:
0.50.50.5
26006
0.5
26005Compromise 1.0670.864AdjFacAdjFacAdjFac 0.960
The final estimate of QMED at the subject site is therefore 0.96059.16 = 56.8 m3s
-1.
It transpires that this would be a very good call for the Suck at Rookwood, where QMED is
gauged to be 56.6 m3s
-1. The excellent outcome in this example is, however, of little comfort
to the user faced with making adjustments at genuinely ungauged sites.
2.7.3 Geostatistical mapping method
It transpires that the geostatistical mapping method of Section 2.6.3 also performs well on
this example. The IDW-technique of Map 2.3 yields an adjusted QMED estimate for the
Suck at Rookwood of 56.9 m3s
-1.
Were an automated method of QMED adjustment implemented in Ireland, it would be
important to determine an appropriate method of incorporating the urban adjustment. Most
likely, the geostatistical mapping would be applied only to the inferred rural element of
QMED, in keeping with the FSU recommendation of Box 2.3.
Volume II Flood Frequency Estimation
47
3 Trend and randomness
It is assumed in flood frequency analysis that successive values in an annual maximum (AM)
series have emerged randomly and independently from a common population of flood values.
In particular, this means that there should be no upward or downward trend in the series of
data nor should there be any strong clustering of groups of larger-than-average or smaller-
than-average values during the passage of time. Recognised tests of trend and randomness
are conducted to determine how well these assumptions are met.
The tests were applied to the AM flow series that lie at the heart of the FSU research. When
seeking to detect trend or non-randomness in hydrological variables, there can be merit in
also studying extremes abstracted in peaks-over-threshold (POT) format (e.g. Robson,
1999a). If such series are available, methods put forward by Buishand (1982) may be used.
The testing was limited to Grade A1 and A2 gauging stations. These include some post-
drainage records on catchments which had earlier experienced arterial drainage works.
3.1 Tests
3.1.1 Methods
Building on experience gained and reported in the UK Flood Studies Report (NERC, 1975)
and the WMO Report Detecting trend and other changes in hydrological data (Kundzewicz
and Robson, 2000), six tests were used to check for trends, shift and serial dependency in the
AM flood series:
Tests for trend
Mann-Kendall (non-parametric test for trend);
Spearman’s ρ (non-parametric test for trend);
Mean-weighted linear regression test (parametric test for trend);
Test for step-change
Mann-Whitney U test (non-parametric test for step-change);
Tests for serial dependency in time series
Turning points (non-parametric test for randomness);
Rank difference (non-parametric test for randomness).
Spearman’s ρ is the rank correlation. An observation has rank r if it is the rth
largest in a
dataset. Rank-based tests use the ranks of the data values rather than the values themselves.
Most rank-based tests assume that data are independent and identically distributed but
typically have the advantage of being robust and simple to apply.
Volume II Flood Frequency Estimation
48
3.1.2 Formats
Statistical trend analyses were carried out on the annual maxima in two formats. The main
analyses were undertaken for 94 AM flow series for stations graded A1 or A2. Their
selection is discussed in Sections E5.1 and E6.1 of Appendix E.
Additionally, a special analysis of the median of annual maxima taken in five-year blocks
was applied to 117 series. The technique of splitting the annual maximum series into (non-
overlapping) five-year segments – and retaining only the median (i.e. middle-ranking) values
– provides a macro view of the data that is unaffected by occasional unusually large or small
annual maxima. This format is reasonably tolerant of minor gaps in the AM series.
Results obtained for the two formats were generally in good agreement. Those for the
standard AM series are now summarised. Supporting material is presented in Appendix E.
3.2 Findings
The exploratory data analysis found 11 AM flow series that showed a significant trend with
time. Eight of the significant trends were upward and three downward. [Editorial note: Of
the three sites showing downward trend, Station 25002 Newport at Barrington’s Bridge and
Station 25003 Mulkear at Abington were later withdrawn from study for reasons given in
Appendix B. The downward trend for Station 25014 Silver at Millbrook might reflect rating
changes in 1971.] Testing for trend and step-change are a valuable element of quality
inspection. Close inspection is warranted before judging whether a statistically significant
change likely reflects a physical effect.
Overall, the results show an unexpected degree of non-randomness in the flood series.
Across all tests on the AM series (Table 3.1), approximately 10% of stations reject the null
hypothesis at the 1% level, whereas in a truly random situation only one station (from about
100 stations) might be expected to reject the null hypothesis. Further details of the testing are
reported in Appendix E.
Table 3.1: Number of cases (out of 94) in which the null hypothesis is rejected
Significance
level
Trend Step-
change Randomness Number
expected
(from 94
cases) by
chance
alone
Mann-
Kendall
test
Spearman’s
ρ
Mean-
weighted
linear
regression
Mann-
Whitney
U
test
Turning
point
test
Rank
difference
test
1% 12 13 8 24 5 9 ≈1
5%
(but not 1%) 14 13 8 15 4 6 ≈4
5%
(all cases) 26 26 16 39 9 15 ≈5
Volume II Flood Frequency Estimation
49
It is concluded that the Irish flood series exhibit more non-stationarity (i.e. trends and step-
changes) than is to be expected by chance. They also exhibit some tendency to non-
randomness. This casts doubt on the validity of the independent and identically distributed
(iid) assumption which underpins all standard methods of flood frequency analysis.
3.3 Pragmatism
International experience with flood frequency analysis nevertheless suggests that making the
iid assumption is the most suitable means of estimating probabilities within a national flood
frequency methodology. It is especially helpful that the iid assumption allows AM series
with gaps – e.g. due to recorder or processing malfunction – to be treated in the same way as
continuous records. For instance, if five annual maxima are missing from 40 years of data,
the series is taken to be equivalent to a record of 35 years, regardless of whether the five
unavailable years are contiguous or not.
While it is pragmatic to make the classical assumption that AM flows are independent and
identically distributed, it is important to recognise that the assumption is not strictly valid.
This highlights the importance of examining AM series closely in all cases: both in terms of
their summary statistics (Chapter 4) and by visual inspection of probability plots (Chapter 5).
Volume II Flood Frequency Estimation
50
4 Descriptive statistics – and inferences therefrom
4.1 Descriptive statistics
4.1.1 Introduction
The research reported in Chapters 4 and 5 represents a detailed exploratory data analysis.
Much of the thinking behind such analyses is to allow the data to “speak” without
prejudgement. Although complex to the non-statistician, the summary statistics employed in
Chapter 4 – and the conventions followed in Chapter 5 – are all relatively standard in the
analysis of extreme values such as annual maximum flows.
4.1.2 Summary statistics – the idea
The AM) flow values provide the basic information on which the entire study rests. The
distribution of values can be viewed as a histogram or in a probability plot (see Chapter 5).
However, for the purposes of applying statistical models – such as probability distributions
based on the iid assumption – it is helpful to summarise the data by a small number of
statistical measures. These are referred to as summary statistics.
Summary statistics expressed in dimensionless form are especially useful, not least for:
Comparing flood frequency behaviour on different catchments;
Drawing inferences about the suitability of probability distributions for describing
flood data.
4.1.3 Summary statistics based on moments and L-moments
In flood hydrology, the most useful statistics relate to:
The typical magnitude of flood flows, e.g. their mean or median value
The “scale” or spread of the data, e.g. indexed by their standard deviation
The skewness (i.e. asymmetry) of the data
The kurtosis or “peakedness” of the data.
Box 4.1 defines dimensionless forms of the scale and skewness statistics based on ordinary
moments. These are the coefficient of variation, CV, and the coefficient of skewness, g. As
well as defining the basic population quantities, Box 4.1 provides formulae for calculating the
statistics from sample data. It also introduces the basic notation {Q1, Q2, Q3, …, QN} to
denote the AM flow series for N years of data.
Skewness values tend to increase slightly with sample size N. Hazen skewness is ordinary
skewness g multiplied by the factor (1 + 8.5/N). This correction factor – introduced by
Hazen (1930) – was shown by Wallis et al. (1974) to provide an almost unbiased estimator of
skewness for lognormal and EV1 distributions and for (mildly) skewed distributions in
general. Hazen skewness is listed here as H-skew.
Volume II Flood Frequency Estimation
51
Notation
The AM flow series is denoted by: {Q1, Q2, Q3, …, QN}.
Population quantities Sample estimates
= mean
N
1i
iQN
1Q
= standard deviation
N
1i
2
i
1N
)Q(Qσ
3 = 3rd
central moment
N
1i
3
i3 )Q(Q2)1)(N(N
Nμ
μ
σCV = coefficient of variation
Q
σCV
3
3
σ
μg = coefficient of skewness
3
3
σ
μg
Hazen’s unbiased skewness H-skew
N
8.51g
Box 4.1: Calculation of moments and their dimensionless ratios
Box 4.2 provides corresponding information for summary statistics based on probability-
weighted moments (PWMs) and L-moments. Prominent amongst these are the L-moment
ratios: the L-CV, L-skewness and L-kurtosis (Hosking, 1990). Relatively comprehensive
presentations of these measures are given in texts such as Hosking and Wallis (1997) and
Robson and Jakob (1999).
4.1.4 Additional summary statistics
Dimensionless measures of scale and skewness are useful for detecting common behaviour
among several series in a dataset. Dimensionless measures of flood magnitude are not
feasible although the specific flood (i.e. the peak flow per unit area) can sometimes be useful
for comparison purposes.
Other dimensionless quantities such as Qmax/Qmed, Qmax/Qmean or Qmean/Qmed can also be
compared between catchments although it has to be realised that Qmax/Qmean and Qmax/Qmed
can be sensitive to the period of record available and have high sampling variability.
Fourth-order moment ratios tend to have very high sampling variability. Nevertheless, values
of L-kurtosis can sometimes be useful when judging the relative merit of particular
3-parameter distributions.
Volume II Flood Frequency Estimation
52
Box 4.2: Calculation of L-moments and their dimensionless ratios
4.2 Summary statistics for Irish flood data
4.2.1 Summary statistics for 181 FSU stations
Principal summary statistics are presented in Table 4.1 for 45 stations graded A1, 70 stations
graded A2, and 66 stations graded B. Columns headed H-skew denote values of Hazen
skewness. Columns headed L-kurt show L-kurtosis values. The notations Qmed and QMED
are interchangeable. Appendix F reports additional summary statistics for these 181 stations.
Notation
In some analyses and plots, data are ranked in ascending order. The ordered AM
series is denoted by: {Q(1), Q(2), Q(3), …, Q(N)} where Q(1) denotes the smallest AM
flow in the series. Thus, the largest flood in the N years of record is Q(N).
Population quantities Sample estimates
M100 = 1st PWM QM100 = Qmean = sample mean
M110 = 2nd
PWM
N
1i
(i)110 Q1)(N
1)(i
N
1M
M120 = 3rd
PWM
N
1i
(i)120 Q2)1)(N(N
2)1)(i(i
N
1M
M130 = 4th
PWM
(i)
N
1i
130 Q3N
3i
2N
2i
1N
1i
N
1M
1 = M100 = 1st L-moment
2 = 2 M110 – M100 = 2nd
L-moment
3 = 6 M120 – 6 M110 + M100 = 3rd
L-moment
4 = 20 M130 – 30 M120 + 12 M110 – M100 = 4th
L-moment
1
2
λ
λτ = L-CV
2
33
λ
λτ = L-skewness
2
44
λ
λτ = L-kurtosis
Above-listed sample values of
M100, M110, M120 and M130 are
inserted into these expressions to
yield the sample estimates of
L-CV, L-skewness and L-kurtosis.
L-moment
ratios
L-moments
Probability-
weighted moments
(PWMs)
Volume II Flood Frequency Estimation
53
Statistics for six stations asterisked in the tables are provided for information only. Stations
25001, 25002, 25003, 25004, 25005 and 25158 play no role in inferences later about the
suitability of particular distributions to describe annual maximum flows. This is because
embankments in the Mulkear catchment contain about four out of five AM flows but are
overtopped in larger floods. This leads to damping of the larger flood magnitudes and
extremely concave downwards probability plots (see Section 5.3). Further details of these
exclusions are given in Appendix B. Box 4.3 refers to a further exceptional station.
Box 4.3: Station 06030 Big at Ballygoly
Table 4.1: Principal summary statistics
Station number and name Grade N Qmean Qmed H-skew CV L-CV L-skew L-kurt
01041 Deele at Sandy Mills B 32 85.08 82.61 0.22 0.29 0.17 0.02 0.13
01055 Mourne Beg at Mourne Beg Weir B 9 2.92 2.70 1.03 0.38 0.23 0.16 0.02
06011 Fane at Moyles Mill A1 48 15.86 15.39 0.81 0.20 0.11 0.09 0.07
06013 Dee at Charleville Weir A1 30 27.81 27.37 0.10 0.27 0.16 0.02 0.01
06014 Glyde at Tallanstown A1 30 22.56 21.46 1.31 0.27 0.15 0.22 0.12
06021 Glyde at Mansfieldstown B 50 21.54 21.50 0.45 0.24 0.13 0.09 0.09
06025 Dee at Burley Bridge A1 30 18.32 18.69 -0.33 0.16 0.09 -0.07 0.19
06026 Glyde at Aclint Bridge A1 46 13.87 12.30 1.05 0.32 0.18 0.24 0.09
06030 Big at Ballygoly [see Box 4.3] B 30 20.58 10.03 3.42 1.61 0.61 0.68 0.48
06031 Flurry at Curralhir A2 18 13.58 11.70 3.12 0.52 0.26 0.39 0.32
06033 White Dee at Coneyburrow Bridge B 25 27.88 18.60 2.49 0.84 0.41 0.46 0.27
06070 Muckno L. at Muckno A1 24 13.32 13.19 0.78 0.25 0.14 0.14 0.14
07006 Moynalty at Fyanstown A2 19 26.73 27.93 -1.09 0.21 0.12 -0.20 0.07
07009 Boyne at Navan Weir A1 29 162.64 134.80 0.93 0.38 0.21 0.21 0.11
07033 Blackwater at Virginia Hatchery A2 25 14.93 14.62 1.74 0.24 0.13 0.16 0.27
08002 Delvin at Naul A1 20 5.62 5.32 1.98 0.21 0.11 0.27 0.17
08003 Broadmeadow at Fieldstown B 18 26.88 22.55 4.21 0.87 0.39 0.36 0.34
08005 Sluice at Kinsaley Hall A2 18 3.04 2.50 1.54 0.68 0.38 0.23 0.19
08007 Broadmeadow at Ashbourne B 15 9.88 8.24 0.89 0.49 0.29 0.18 -0.01
08008 Broadmeadow at Broadmeadow A2 25 44.55 40.90 1.74 0.63 0.34 0.28 0.16
08009 Ward at Balheary A1 11 10.38 6.59 5.43 1.41 0.56 0.68 0.65
08011 Nanny at Duleek Road Bridge B 23 31.00 32.22 -0.93 0.25 0.14 -0.15 0.20
08012 Stream at Ballyboghill B 19 4.21 4.35 -0.54 0.58 0.33 -0.09 0.11
09001 Ryewater at Leixlip A1 48 38.71 35.46 1.17 0.44 0.24 0.19 0.15
09002 Griffeen at Lucan A1 24 7.24 5.40 2.40 0.83 0.42 0.39 0.25
Editorial note: Station 06030 Big at Ballygoly is also marked in the tables because the
AM flood series analysed for this Grade B station is highly questionable. A radically
different rating curve has been applied to the first five years of data, leading to peak flows
that appear unrealistically large. The effect is not enough to perturb the estimate of
QMED greatly. However, the values of variability and skewness quoted in Table 4.1 (and
the maximum recorded flood quoted in Appendix F) are not to be relied on.
Volume II Flood Frequency Estimation
54
Station number and name Grade N Qmean Qmed H-skew CV L-CV L-skew L-kurt
09010 Dodder at Waldron’s Bridge A1 19 70.15 48.00 3.24 0.86 0.42 0.42 0.30
09035 Cammock at Killeen Road B 9 12.04 11.70 2.75 0.60 0.33 0.33 0.23
10002 Avonmore at Rathdrum B 47 88.19 83.49 2.85 0.43 0.21 0.27 0.31
10021 Shanganagh at Common’s Road A1 24 7.87 7.36 0.95 0.37 0.21 0.19 0.07
10022 Cabinteely at Carrickmines A1 18 3.84 3.85 0.36 0.40 0.23 0.06 0.06
10028 Aughrim at Knocknamohill B 16 56.69 46.95 1.57 0.38 0.21 0.30 0.08
11001 Owenavorragh at Boleany B 33 49.85 47.17 3.03 0.33 0.16 0.24 0.30
12001 Slaney at Scarawalsh A2 50 169.50 157.00 1.59 0.36 0.19 0.18 0.17
12013 Slaney at Rathvilly B 30 45.16 43.55 -0.01 0.27 0.15 0.03 0.12
14005 Barrow at Portarlington A2 48 40.81 38.27 2.01 0.29 0.15 0.29 0.22
14006 Barrow at Pass Bridge A1 51 83.76 80.52 1.46 0.20 0.11 0.24 0.21
14007 Stradbally at Derrybrock A1 25 16.94 16.20 1.12 0.30 0.17 0.22 0.07
14009 Cushina at Cushina A2 25 6.69 6.79 1.42 0.23 0.13 0.14 0.24
14011 Slate at Rathangan A1 26 12.07 12.30 0.19 0.25 0.14 0.02 0.16
14013 Burrin at Ballinacarrig A2 50 16.54 16.05 0.37 0.26 0.15 0.07 0.06
14018 Barrow at Royal Oak A1 51 141.83 147.98 0.22 0.24 0.14 0.04 0.06
14019 Barrow at Levitstown A1 51 103.46 102.41 0.61 0.24 0.14 0.09 0.13
14029 Barrow at Graiguenamanagh A2 47 162.54 160.74 0.18 0.14 0.08 0.06 0.07
14033 Owenass at Mountmellick B 22 22.59 19.50 0.53 0.28 0.16 0.14 -0.07
14034 Barrow at Bestfield A2 14 137.30 125.00 2.23 0.32 0.17 0.33 0.17
15001 Kings at Annamult A2 42 89.39 88.75 0.14 0.28 0.16 0.00 0.08
15002 Nore at John’s Bridge A2 35 211.98 197.00 0.59 0.31 0.18 0.08 0.09
15003 Dinan at Dinan Bridge A2 50 143.58 150.76 -0.78 0.20 0.11 -0.15 0.09
15004 Nore at McMahons Bridge A2 51 38.96 37.28 0.96 0.31 0.17 0.13 0.15
15005 Erkina at Durrow Foot Bridge B 50 28.47 27.44 1.78 0.34 0.18 0.19 0.21
15012 Nore at Ballyragget B 16 77.16 77.11 0.90 0.30 0.17 0.08 0.21
16001 Drish at Athlummon A2 33 15.65 15.66 0.30 0.22 0.12 0.01 0.12
16002 Suir at Beakstown A2 51 55.40 52.66 1.79 0.30 0.16 0.17 0.19
16003 Clodiagh at Rathkennan A2 51 31.17 29.98 0.85 0.18 0.10 0.21 0.05
16004 Suir at Thurles A2 48 22.17 21.37 0.53 0.20 0.11 0.07 0.11
16005 Multeen at Aughnagross A2 30 23.11 21.79 1.33 0.18 0.10 0.20 0.13
16006 Multeen at Ballinclogh Bridge B 33 30.37 27.87 0.34 0.39 0.23 0.06 0.03
16007 Aherlow at Killardry B 51 79.18 75.84 0.46 0.32 0.18 0.09 0.06
16008 Suir at New Bridge A2 51 90.66 92.32 -0.32 0.13 0.07 -0.06 0.05
16009 Suir at Cahir Park A2 52 159.29 162.21 -0.41 0.17 0.10 -0.10 0.05
16011 Suir at Clonmel A1 52 234.52 223.00 0.42 0.30 0.17 0.09 0.05
16012 Tar at Tar Bridge B 36 55.20 54.57 0.39 0.28 0.16 0.06 0.08
16013 Nire at Fourmilewater B 33 101.69 93.21 0.86 0.42 0.24 0.16 0.08
Volume II Flood Frequency Estimation
55
Station number and name Grade N Qmean Qmed H-skew CV L-CV L-skew L-kurt
16051 Rossestown at Clobanna B 13 2.95 2.85 2.53 0.36 0.19 0.33 0.26
18001 Bride at Mogeely Bridge B 48 71.07 71.49 -0.16 0.19 0.11 -0.03 0.10
18002 Ballyduff at Muns Blackwater B 49 353.65 344.00 0.24 0.16 0.09 0.06 0.10
18003 Blackwater at Killavullen B 49 282.76 266.15 1.10 0.23 0.13 0.16 0.10
18004 Ballynamona at Awbeg A2 46 30.96 31.20 1.25 0.17 0.09 0.04 0.33
18005 Funshion at Downing Bridge A2 50 56.69 53.05 1.63 0.27 0.14 0.22 0.18
18006 Blackwater at CSET Mallow B 27 291.30 286.00 1.10 0.14 0.08 0.19 0.12
18016 Blackwater at Duncannon B 24 80.99 79.65 0.49 0.23 0.14 0.12 0.00
18048 Blackwater at Dromcummer B 23 222.77 220.00 1.11 0.08 0.05 0.15 0.22
18050 Blackwater at Duarrigle B 24 121.96 124.50 0.12 0.20 0.11 -0.01 0.09
19001 Owenboy at Ballea Upper A2 48 15.87 15.42 0.48 0.17 0.09 0.09 0.12
19014 Lee at Dromcarra B 20 79.69 71.89 1.67 0.38 0.21 0.30 0.12
19016 Bride at Ovens Bridge B 8 28.74 29.58 -1.35 0.16 0.09 -0.14 0.22
19020 Owennacurra at Ballyedmond A2 28 24.63 22.40 0.12 0.34 0.20 0.03 0.04
19031 Sullane at Macroom B 9 131.09 135.90 1.46 0.27 0.16 0.16 0.17
19046 Martin at Station Road B 9 31.09 29.95 -0.34 0.26 0.16 -0.04 0.06
20002 Bandon at Curranure B 31 140.60 126.28 2.13 0.37 0.19 0.37 0.28
20006 Argideen at Clonakilty WW B 25 30.25 27.70 1.67 0.30 0.16 0.27 0.18
22006 Flesk at Flesk Bridge B 51 165.89 169.09 0.66 0.25 0.14 0.07 0.12
22009 Dreenagh at White Bridge B 24 11.91 11.47 1.17 0.16 0.09 0.18 0.19
22035 Laune at Laune Bridge B 14 112.81 116.40 -0.86 0.20 0.12 -0.15 0.03
23001 Galey at Inch Bridge A2 45 97.39 99.05 1.22 0.33 0.18 0.13 0.18
23012 Lee at Ballymullen A2 18 16.87 15.66 2.98 0.29 0.15 0.39 0.33
24002 Camogue at Gray’s Bridge A2 27 24.06 23.49 0.32 0.19 0.11 0.06 0.17
24004 Maigue at Bruree B 52 54.86 50.63 0.91 0.39 0.22 0.18 0.10
24008 Maigue at Castleroberts A2 30 120.96 119.13 0.28 0.26 0.15 0.05 0.09
24011 Deel at Deel Bridge B 33 103.01 104.55 0.42 0.22 0.12 -0.02 0.24
24012 Deel at Grange Bridge B 41 110.45 109.99 -0.02 0.16 0.09 0.01 0.13
24022 Mahore at Hospital A2 20 9.83 9.80 1.12 0.41 0.23 0.12 0.16
24030 Deel at Danganbeg B 25 52.89 52.00 0.88 0.15 0.08 0.11 0.10
24082 Maigue at Islandmore A2 28 135.47 140.01 -0.22 0.26 0.15 -0.04 0.13
25001 Mulkear at Annacotty* A2 49 133.95 132.88 -0.19 0.16 0.09 -0.02 0.16
25002 Newport at Barringtons Bridge* A2 51 61.15 62.64 -0.65 0.16 0.09 -0.14 0.06
25003 Mulkear at Abington* A1 51 69.45 68.98 0.05 0.15 0.08 0.00 0.11
25004 Bilboa at Newbridge* B 30 41.70 42.30 -0.16 0.23 0.13 -0.04 0.12
25005 Dead at Sunville* A2 46 28.73 29.63 -0.98 0.11 0.06 -0.19 0.10
25006 Brosna at Ferbane A1 52 86.77 81.91 0.71 0.25 0.14 0.14 0.18
25011 Brosna at Moystown B 51 85.64 82.02 1.31 0.34 0.18 0.14 0.23
Volume II Flood Frequency Estimation
56
Station number and name Grade N Qmean Qmed H-skew CV L-CV L-skew L-kurt
25014 Silver at Millbrook Bridge A1 54 17.67 17.25 0.57 0.23 0.13 0.10 0.13
25016 Clodiagh at Rahan A2 42 23.04 22.57 0.62 0.22 0.13 0.08 0.16
25017 Shannon at Banagher A1 55 413.25 407.68 0.18 0.20 0.12 0.04 0.08
25020 Killimor at Killeen B 35 46.60 43.65 0.93 0.33 0.19 0.16 0.11
25021 Little Brosna at Croghan A2 44 28.03 28.58 -0.13 0.14 0.08 -0.03 0.07
25023 Little Brosna at Milltown A1 52 12.14 11.22 0.57 0.29 0.16 0.14 0.06
25025 Ballyfinboy at Ballyhooney A1 31 10.15 10.18 0.57 0.29 0.17 0.08 0.15
25027 Ollatrim at Gourdeen Bridge A1 43 23.32 22.10 0.27 0.28 0.16 0.05 0.13
25029 Nenagh at Clarianna A2 33 54.12 56.48 -0.09 0.24 0.14 -0.02 -0.02
25030 Graney at Scarriff Bridge A1 48 43.80 40.64 1.01 0.32 0.18 0.18 0.13
25034 L. Ennell Trib at Rochfort A2 24 1.50 1.48 -0.48 0.29 0.17 -0.08 0.12
25038 Nenagh at Tyone B 17 42.08 39.30 1.30 0.27 0.15 0.17 0.22
25040 Bunow at Roscrea A2 20 3.78 3.59 1.32 0.27 0.15 0.20 0.18
25044 Kilmastulla at Coole A2 33 25.38 22.70 1.23 0.34 0.19 0.25 0.15
25124 Brosna at Ballynagore A2 18 12.79 13.65 -0.31 0.36 0.20 -0.07 0.23
25158 Bilboa at Cappamore* A1 18 47.66 43.88 0.17 0.29 0.17 0.05 0.13
26002 Suck at Rookwood A2 53 56.98 56.56 2.29 0.22 0.11 0.22 0.29
26005 Suck at Derrycahill A2 51 92.80 93.21 0.27 0.18 0.10 0.03 0.14
26006 Suck at Willsbrook A1 53 26.57 24.23 3.87 0.37 0.15 0.40 0.41
26007 Suck at Bellagill Bridge A1 53 91.75 88.15 1.05 0.19 0.11 0.16 0.16
26008 Rinn at Johnston’s Bridge A1 49 23.68 22.94 1.49 0.19 0.10 0.19 0.19
26009 Black at Bellantra Bridge A2 35 13.66 13.22 0.93 0.16 0.09 0.20 0.11
26010 Cloone at Riverstown B 35 20.03 17.17 1.64 0.41 0.22 0.35 0.18
26014 Lung at Banada Bridge B 16 44.10 42.82 1.80 0.23 0.12 0.19 0.26
26018 Owenure at Bellavahan A2 49 9.19 8.95 0.77 0.20 0.11 0.13 0.11
26019 Camlin at Mullagh A1 51 22.34 21.18 1.17 0.25 0.14 0.23 0.12
26020 Camlin at Argar Bridge A1 32 11.21 11.27 0.16 0.19 0.11 0.03 0.07
26021 Inny at Ballymahon A2 30 65.88 66.34 -0.86 0.25 0.14 -0.11 0.19
26022 Fallan at Kilmore A2 33 6.64 6.49 0.47 0.30 0.17 0.09 0.05
26058 Inny Upper at Ballinrink Bridge B 24 5.98 5.35 1.60 0.38 0.21 0.28 0.18
26059 Inny at Finnea Bridge A1 17 12.98 12.20 0.31 0.18 0.10 0.11 0.17
26108 Owenure at Boyle Abbey Bridge B 15 56.29 57.32 0.28 0.18 0.11 0.06 -0.06
27001 Claureen at Inch Bridge A2 30 20.65 20.10 1.55 0.20 0.11 0.19 0.25
27002 Fergus at Ballycorey A1 51 34.22 32.60 1.37 0.23 0.12 0.18 0.22
27003 Fergus at Corofin A2 48 24.01 22.92 0.71 0.24 0.13 0.09 0.22
28001 Inagh at Ennistimon B 17 52.69 47.58 4.89 0.40 0.16 0.47 0.61
29001 Raford at Rathgorgin A1 40 14.17 13.46 0.34 0.19 0.11 0.09 0.10
29004 Clarinbridge at Clarinbridge A2 32 11.39 11.30 0.77 0.15 0.08 0.14 0.08
Volume II Flood Frequency Estimation
57
Station number and name Grade N Qmean Qmed H-skew CV L-CV L-skew L-kurt
29007 L. Cullaun at Craughwell B 22 27.83 26.49 1.03 0.22 0.12 0.15 0.20
29011 Dunkellin at Kilcolgan Bridge A1 22 31.94 28.89 3.37 0.30 0.14 0.41 0.32
29071 L. Cutra at Cutra A2 26 16.00 15.70 0.74 0.23 0.13 0.11 0.19
30007 Clare at Ballygaddy A2 31 61.93 62.98 1.30 0.20 0.11 0.11 0.20
30012 Clare at Claregalway B 9 126.89 126.00 1.00 0.12 0.07 0.13 0.15
30021 Robe at Christina’s Bridge B 26 28.17 27.20 2.17 0.34 0.18 0.21 0.29
30031 Cong at Cong Weir B 24 94.35 93.88 -0.62 0.18 0.10 -0.02 0.13
30037 Robe at Clooncormick B 21 1.80 1.79 -0.42 0.37 0.21 -0.09 0.18
30061 Corrib Estuary at Wolfe Tone Bridge A2 33 274.97 247.97 3.04 0.32 0.15 0.41 0.38
31002 Cashla at Cashla A1 26 12.89 12.16 1.92 0.24 0.13 0.31 0.18
31072 Cong at Cong Weir B 26 49.08 43.20 2.04 0.38 0.20 0.31 0.20
32011 Bunowen at Louisberg Weir B 26 74.88 64.87 0.75 0.30 0.17 0.17 0.06
32012 Newport at Newport Weir A2 24 30.06 29.85 -0.12 0.12 0.07 0.00 0.18
33001 Glenamoy at Glenamoy B 25 62.11 59.30 1.65 0.28 0.15 0.20 0.19
33070 Carrowmore L. at Carrowmore A1 28 7.90 7.67 1.19 0.16 0.09 0.10 0.18
34001 Moy at Rahans A2 36 174.76 174.61 1.08 0.19 0.10 0.08 0.21
34003 Moy at Foxford A2 29 180.42 178.00 1.26 0.17 0.09 0.11 0.25
34007 Deel at Ballycarroon B 53 90.37 84.48 0.96 0.36 0.20 0.14 0.10
34009 Owengarve at Curraghbonaun A2 33 28.37 27.48 0.43 0.17 0.10 0.08 0.16
34010 Moy at Cloonacannana B 12 123.29 113.72 1.31 0.30 0.17 0.21 0.11
34011 Manulla at Gneeve Bridge A2 30 18.80 18.73 0.78 0.16 0.09 0.10 0.23
34018 Castlebar at Turlough A1 27 11.50 11.28 0.93 0.20 0.11 0.18 0.03
34024 Pollagh at Kiltimagh A2 28 20.70 20.80 -0.39 0.12 0.07 -0.05 0.14
35001 Owenmore at Ballynacarrow A2 29 30.52 31.16 0.10 0.21 0.12 -0.01 0.19
35002 Owenbeg at Billa Bridge A2 34 51.78 50.48 0.07 0.17 0.10 0.03 0.08
35005 Ballysadare at Ballysadare A2 55 77.78 75.42 1.04 0.26 0.14 0.20 0.14
35011 Bonet at Dromahair B 36 116.02 115.36 0.04 0.30 0.17 0.01 0.08
35071 L. Melvin at Lareen A2 30 26.95 26.29 0.76 0.18 0.10 0.12 0.20
35073 L. Gill at Lough Gill A2 30 54.81 54.05 0.34 0.22 0.12 0.08 0.14
36010 Annalee at Butlers Bridge A1 50 66.56 66.80 1.05 0.22 0.12 0.15 0.22
36011 Erne at Bellahillan B 49 17.91 18.23 -0.51 0.18 0.10 -0.10 0.11
36012 Erne at Sallaghan A1 47 14.22 14.12 0.21 0.22 0.13 0.03 0.10
36015 Finn at Anlore A1 33 23.14 22.08 2.62 0.32 0.16 0.32 0.30
36018 Dronmore at Ashfield Bridge A1 50 15.84 16.25 0.43 0.18 0.10 0.04 0.10
36019 Erne at Belturbet A2 47 89.60 89.95 -0.16 0.18 0.10 -0.03 0.06
36021 Yellow at Kiltybarden A2 27 24.96 23.37 1.85 0.22 0.12 0.20 0.22
36031 Cavan at Lisdarn A2 30 6.85 6.45 4.19 0.22 0.10 0.37 0.39
36071 L. Scur at Gowly B 20 6.36 6.49 -0.06 0.15 0.09 -0.04 0.01
Volume II Flood Frequency Estimation
58
Station number and name Grade N Qmean Qmed H-skew CV L-CV L-skew L-kurt
38001 Owenea at Clonconwal B 33 70.02 70.63 1.78 0.16 0.09 0.08 0.26
39001 New Mills at Swilly B 30 44.88 44.25 0.06 0.20 0.12 0.01 0.08
39008 Leannan at Gartan Bridge A2 33 28.34 28.18 0.73 0.26 0.15 0.13 0.11
39009 Fern O/L at Aghawoney A2 33 45.91 45.72 1.03 0.26 0.14 0.16 0.15
4.2.2 Variability and skewness of Irish flood data
Table 4.2 reports average values of the summary statistics describing variability and
skewness. Inclusion of both measures based on moments (i.e. CV and Hazen skewness) and
measures based on L-moments (i.e. L-CV and L-skewness) facilitates comparisons with other
studies. Values of Qmax ∕ Qmean are of interest because they convey general information about
the flood regime of a region that is additional to the measures of variability and skewness.
[Editorial note: The averages presented in Table 4.2 have been calculated as arithmetic
means. Given that each of the summary statistics is defined as a ratio, use of a geometric
mean would have been more appropriate. The numbers of stations shown in each grade
reflect the six excluded stations in the Mulkear basin (see Section 4.2.1 and Appendix B).]
Table 4.2: Average values of some statistics for gauging stations (by station grade)
Station
grade
#
stations
Qmed ∕
Qmean CV
L-
CV
H-
skew
L-
skew
L-
kurt
Qmax ∕
Qmean
Qmax ∕
AREA
A1 43 0.943 0.311 0.165 1.153 0.172 0.158 1.852 0.278
A2 67 0.975 0.249 0.137 0.838 0.106 0.160 1.651 0.307
A1+A2 110 0.963 0.273 0.148 0.961 0.132 0.159 1.730 0.296
B 65 0.947 0.325 0.174 1.024 0.140 0.155 1.839 0.787
A1+A2+B 175 0.957 0.293 0.157 0.984 0.135 0.158 1.770 0.473
Here, attention is chiefly paid to the results for the Grade A1 and A2 stations. The average
CV of 0.273 is identical to that reported for Ireland in Table I.2.3 of the FSR (NERC, 1975)
based on 63 stations with records of 15 years or more. The average Hazen skewness of 0.963
is larger than the average weighted skewness of 0.662 reported in the FSR table, though the
FSR value is without Hazen’s correction. [Editorial note: For reasons discussed in Section
I.2.3.3 of the FSR, derivation of a pooled value of skewness warrants greater sophistication
than an arithmetic mean of individual values of skewness: regardless of whether the Hazen
correction is applied.]
The average value of Qmed/Qmean for the A1 + A2 stations is 0.963. The expected ratio for an
EV1 distribution with CV = 0.273 is 0.955.
In comparison to many countries, Ireland has a flood hydrology regime that can be
characterised as “low CV and low skewness”. This is typical of very humid conditions where
between-year variation in flood magnitudes is relatively small.
Volume II Flood Frequency Estimation
59
4.2.3 Comparisons of CV with L-CV and of H-skew with L-skew
The distributions of CV and L-CV are summarised in Histograms 4.1, and those of Hazen
skewness and L-skewness in Histograms 4.2. The distribution of AM values is positively
skewed at most sites (i.e. H-skew > 0 and L-skew > 0). This is particularly the case for the
Grade A1 stations, which are considered the most reliable at measuring across the full range
of flows.
30
20
10
0
30
20
10
0
1.61.41.21.00.80.60.40.20.0
30
20
10
0
CV of 43 A1 stations
Fre
quency CV of 67 A2 stations
CV of 110 A1 + A2 stations
30
20
10
0
30
20
10
0
0.60.50.40.30.20.10.0
30
20
10
0
L-CV of 43 A1 stations
Fre
quency
L-CV of 67 A2 stations
L-CV of 110 A1 + A2 stations
Histograms 4.1: CV and L-CV at Grade A1 and A2 stations
20
10
0
20
10
0
6420-2
20
10
0
H-skew of 43 A1 stations
Fre
quency
H-skew of 67 A2 stations
H-skew of 110 A1 + A2 stations
20
10
0
20
10
0
0.80.60.40.20.0-0.2-0.4
20
10
0
L-skew of 43 A1 stations
Fre
quen
cy
L-skew of 67 A2 stations
L-skew of 110 A1 + A2 stations
Histograms 4.2: Hazen skewness and L-skewness at Grade A1 and A2 stations
Volume II Flood Frequency Estimation
60
Figure 4.1 shows the relationship between L-CV and CV, and Figure 4.2 shows that between
L-skewness and Hazen skewness.
1.61.41.21.00.80.60.40.20.0
0.8
0.6
0.4
0.2
0.0
CV
L-C
V
Forcing intercept to zero yields: L-CV = 0.515 CV
Regression of y on x yields: L-CV = 0.441 CV + 0.027
Figure 4.1: Relationship between L-CV and CV for 110 A1 + A2 stations
6420-2
0.8
0.6
0.4
0.2
0.0
-0.2
Hazen skewness
L-s
kew
ness
Regression of y on x yields: L-skew = 0.122 H-skew + 0.014
Forcing intercept to zero yields: L-skew = 0.129 H-skew
Figure 4.2: Relationship between L-skew and H-skew for 110 A1 + A2 stations
While CV and skewness vary between gauging stations some of this variation is due to
sampling variability rather than true differences. [Editorial note: Das and Cunnane (2012)
explore the sensitivity of L-CV and L-skewness to record length. Exploring the same group
of 110 stations, they note that the largest values are typically associated with the shortest
records. Further work might explore whether the higher values are artefacts of the limited
period of record or a product of the nature of those catchments with shorter records.]
Volume II Flood Frequency Estimation
61
From Table 4.2 it can be seen that – for the 110 A1 and A2 stations – the ratio L-CV / CV =
0.148/0.273 = 0.542 and the ratio L-skew / H-skew = 0.132/0.961 = 0.137. These compare
reasonably well with the corresponding theoretical values of 0.540 and 0.149 for an EV1
distribution. [Editorial note: The first ratio is √6 ℓn2 / π. The second ratio is evaluated as
L-skewness/skewness = 0.1699/1.139 = 0.149.]
4.3 Geographical traits
Map 4.1 shows the pattern of specific QMED values, i.e. QMED/AREA. Unsurprisingly,
many of the higher values lie in the wetter regions (e.g. the North West and the South West).
Map 4.1: Specific QMED (m3s
-1/km
2) for 176 A1 + A2 + B stations
Legend
Sp.Qmed
!( 0.006829 - 0.151544
!( 0.151545 - 0.270321
!( 0.270322 - 0.418750
!( 0.418751 - 0.812329
!( 0.812330 - 1.823686
´
!(
!(
!( !(
!(
!(
!(
!(
!(
!(!(
!( !(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
[Symbols placed at catchment outlets]
Specific
QMED
0.007 to 0.152
0.152 to 0.270
0.270 to 0.419
0.419 to 0.812
0.812 to 1.824
Volume II Flood Frequency Estimation
62
The patterns of CV and Hazen skewness are illustrated in Map 4.2 and Map 4.3. It is
noticeable that many of the stations around Dublin have high variability in annual maximum
floods. The geographical pattern of Hazen skewness (see Map 4.3) is less clear. There may
be some weak tendency for higher values in the North Midlands and at stations nearer the
coast.
[Editorial note: Geographical traits are also discussed by Ahilan et al., 2012.]
Map 4.2: CV at 110 A1 + A2 stations
´Legend
CV
!( 0.110000 - 0.190000
!( 0.190001 - 0.270000
!( 0.270001 - 0.440000
!( 0.440001 - 0.860000
!( 0.860001 - 1.410000
!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
901090029001
800980088005
8002
7033
7009
7006
6070 6031
6026
6025 6013
6011
3900939008
3603136021
3601936018
36015
36012
36011
36010
35073
35071
3500535002
35001
3402434018
34011
3401034003
34001
33070
32012
31002 30061
30007
29071
2901129004
27003
2700227001
26059
26022
26021
26020
26019
2601826017
2600926008
26007
26006
26005
26002
25158
25124
2504425040
25034
2503025029
25027
25025
25023
25021
25017
25016
2501425006
250052500325002
2408224022
24008
24002
23012
23001
19020
19001
1800518004
16009
16008
16005
16004
1600316002
16001
15004
15003
15001
14034
14029
14019
14018
14013
1401114009
14007
1400614005
12001
11001
1002210021
CV
0.11 to 0.19
0.19 to 0.27
0.27 to 0.44
0.44 to 0.86
0.86 to 1.41
[Symbols placed at catchment outlets]
Volume II Flood Frequency Estimation
63
Map 4.3: Hazen skewness at 110 A1 + A2 stations
4.4 Preliminary distribution choice from skewness v. record length plot
Another way of viewing skewness values is explored in Figure 4.3 and Figure 4.4, which
respectively show skewness and L-skewness plotted against length of record. In the interest
of model parsimony, it is desirable to consider the adequacy of 2-parameter models such as
the Normal and EV1 distributions. The confidence intervals marked here are based on the
assumption of (a) the Normal distribution and (b) the EV1 distribution.
´Legend
H_SKEW
!( -1.090000 - -0.090000
!( -0.089999 - 0.620000
!( 0.620001 - 1.460000
!( 1.460001 - 2.620000
!( 2.620001 - 5.430000
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
901090029001
800980088005
8002
7033
7009
7006
6070 6031
6026
60256014 6013
6011
15002
16011
3900939008
3603136021 36019 36018
36015
36012
36010
35073
35071
3500535002
35001
3402434018
34011
34009
34001
33070
32012
31002 30061
30007
29071
290112900429001
27003
2700227001
26059
26022
26021
26020
26019
2601826009
26008
26007
26006
26005
26002
25124
2504425040
25034
250302502925027
25025
25023
25021
25017
25016
25014
25006
24082 24022
24008
24002
23012
23001
19020
19001
1800518004
16009
16008
16005
1600416003
16002
15004
15003
15001
14034
14029
14019
14018
14013
1401114009
14007
1400614005
12001
1002210021
- 1.09 to - 0.09
- 0.09 to 0.62
0.62 to 1.46
1.46 to 2.62
2.62 to 5.43
Hazen
skewness
[Symbols placed at catchment outlets]
Volume II Flood Frequency Estimation
64
The Normal distribution confidence intervals for skewness are calculated using the standard
expression for samples of size N:
3)2)(N1)(N(N
1)6N(NskewnessSE
4.1
Other confidence intervals were obtained by simulation.
Figure 4.3: Skewness at 110 A1 + A2 stations versus record length
with confidence intervals for (a) Normal samples and (b) EV1 samples
Skewness vs RL (Normal Dist.)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60
Record Length
Skew
ness
Observed Theoretical Skew. Value for N.D. 95% C.I. 67% C.I.
Skewness vs RL (EV1)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60Record_Length
Skew
ness
Observed Theoretical Skew. value for EV1 67% C.I. 95% C.I.
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60L-S
kew
ness
Skewness vs RL (Normal Dist.)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60
Record Length
Skew
ness
Observed Theoretical Skew. Value for N.D. 95% C.I. 67% C.I.
Skewness vs RL (EV1)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60Record_Length
Skew
ness
Observed Theoretical Skew. value for EV1 67% C.I. 95% C.I.
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
Skewness vs RL (Normal Dist.)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60
Record Length
Skew
ness
Observed Theoretical Skew. Value for N.D. 95% C.I. 67% C.I.
Skewness vs RL (EV1)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60Record_Length
Skew
ness
Observed Theoretical Skew. value for EV1 67% C.I. 95% C.I.
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
Skewness vs RL (Normal Dist.)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60
Record Length
Skew
ness
Observed Theoretical Skew. Value for N.D. 95% C.I. 67% C.I.
Skewness vs RL (EV1)
-2
-1
0
1
2
3
4
0 10 20 30 40 50 60Record_Length
Skew
ness
Observed Theoretical Skew. value for EV1 67% C.I. 95% C.I.
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
L-S
kew
ness
Sk
ewnes
s
Record length (years)
(a) Samples from Normal distribution
Skew
nes
s
Record length (years)
(b) Samples from EV1 distribution
Volume II Flood Frequency Estimation
65
Figure 4.4: L-skewness against record length for 110 A1 + A2 gauging station
with confidence intervals for (a) Normal samples and (b) EV1 samples
Because of the relatively low values of observed skewness at many gauging stations, it is
appropriate to ask whether the data as a whole could be considered to have come from a
Normal distribution. Looking at the upper end of the skewness range in Figure 4.3a, it is seen
that 46 stations fall above the upper confidence interval for the Normal distribution. If the
data as a whole were Normal then only 2.5% of the 110 values (say three values) would be
expected to lie above the upper 95% confidence interval. Thus the Normal hypothesis cannot
be accepted.
L-Skewness vs RL (Normal Dist.)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60Record_Length
L-S
kew
ness
Observed Theoretical L-Skew.value for N.D. 95% C.I. 67% C.I.
L-Skewness vs RL (EV1)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
Record_Length
L-S
kew
ness
Observed Theoretical L-Skew value for EV1 67% C.I. 95% C.I.
L-Skewness vs RL (Normal Dist.)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60Record_Length
L-S
kew
ness
Observed Theoretical L-Skew.value for N.D. 95% C.I. 67% C.I.
L-Skewness vs RL (EV1)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
Record_Length
L-S
kew
ness
Observed Theoretical L-Skew value for EV1 67% C.I. 95% C.I.
L-Skewness vs RL (Normal Dist.)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60Record_Length
L-S
kew
ness
Observed Theoretical L-Skew.value for N.D. 95% C.I. 67% C.I.
L-Skewness vs RL (EV1)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
Record_Length
L-S
kew
ness
Observed Theoretical L-Skew value for EV1 67% C.I. 95% C.I.
L-Skewness vs RL (Normal Dist.)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60Record_Length
L-S
kew
ness
Observed Theoretical L-Skew.value for N.D. 95% C.I. 67% C.I.
L-Skewness vs RL (EV1)
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0 10 20 30 40 50 60
Record_Length
L-S
kew
ness
Observed Theoretical L-Skew value for EV1 67% C.I. 95% C.I.
L-s
kew
nes
s
Record length (years)
(a) Samples from Normal distribution
L-s
kew
nes
s
Record length (years)
(b) Samples from EV1 distribution
Volume II Flood Frequency Estimation
66
Looking at the lower end of the skewness range in Figure 4.3b, it is seen that 14 stations fall
below the lower confidence interval for the EV1 distribution. Although less divergent (than
for the Normal hypothesis in Figure 4.3a), the number is still many times too large (i.e.
14 >>3) for the hypothesis that all the samples come from the EV1 family to be accepted as a
whole.
The preliminary conclusion on the basis of skewness values is that the stations with the larger
values of skewness might be modelled by an EV1 distribution but with the reservation that
there is a serious number (19 out of 110 stations, or 17%) which are not in keeping with this
choice. Thus the choice of the EV1 distribution to represent all stations in Ireland is of
doubtful validity.
Skewness in the sample sizes typically found in hydrology is not renowned for its precision.
L-skewness offers some advantages. However, very similar findings are obtained from the
plots of L-skewness against record length shown above in Figure 4.4.
An alternative assessment approach is to use L-moment ratio diagrams.
4.5 Preliminary distribution choice aided by L-moment ratio diagrams
4.5.1 L-moment ratio diagrams
Another way to explore the suitability of different probability distributions is to use
L-moment ratio diagrams. These provide a useful diagnostic tool (Hosking and Wallis, 1997)
though cannot be regarded as entirely reliable in the context of the sample sizes typically
available. Diagrams based on conventional moment ratios are also possible but are regarded
as having weaker discriminating capability (see Appendix 3 of Cunnane, 1989).
The most useful L-moment ratio diagram is a graph of L-kurtosis against L-skewness.
Usually a 2-parameter distribution with a location and a scale parameter plots as a single
point on the diagram. A 3-parameter distribution with location, scale and shape parameters
typically appears as a line or curve. The distribution selection process begins with a scatter-
plot of the sample L-moment ratios. These are compared with the points/curves that
represent the theoretical L-moment ratios for the candidate distributions. Figure 4.5 shows
this for the AM flows at the 110 A1 + A2 sites.
4.5.2 Implication for choice of 2-parameter distribution
In the context of the 2-parameter distributions alone, the sample data for the 110 stations are
on average appreciably closer to the population L-moment ratios of an EV1 distribution (●)
than to those of the LO (♦) or LN (■) distribution. Here, LO denotes the Logistic distribution
and LN denotes the lognormal distribution.
Figure 4.6 shows sample L-moments of data simulated from EV1, LO and LN distributions:
each with the same record lengths and parameters as the actual records. In the case of the
EV1 simulated data, the scatter of points is narrower than in the observed data, especially on
the LH side. Many more of the observed values of L-skewness are negative than is the case
for the simulated data. The LO and LN simulated L-moments cover the range of observed
values in the negative L-skewness domain more adequately than do the EV1 simulated data.
Volume II Flood Frequency Estimation
67
However, the LO and LN simulated L-moments cover the range of observed values less well
in the positive L-skewness domain, where the EV1 simulated data perform better. Further
realisations of the experiment were made, for which the same general patterns were noted.
Figure 4.5: L-moment ratio diagram for annual maximum floods
Figure 4.6: Comparison with equivalent samples drawn from particular distributions
L-Moment Ratio( Observed data)
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (EV1 simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( LN simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (LO simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( Observed data)
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (EV1 simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( LN simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (LO simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( Observed data)
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (EV1 simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( LN simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (LO simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( Observed data)
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (EV1 simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio( LN simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio (LO simulated data)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7L-Skewness
L-K
urt
osis
L-Moment Ratio Diagram for 110 A1 & A2 Stations
-0.20
-0.10
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
-0.30 -0.20 -0.10 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70L-Skewness
L-K
urt
osis
sample_data
sample_avg
EV1
LO
LN
GEV
LN3
GLO
Normal
110 A1 + A2 stations
L
-kurt
osi
s
L-skewness
L-k
urt
osi
s
L-skewness
(a) Observed data (b) EV1 simulated data
(c) LN simulated data (d) LO simulated data
Station 08009
Volume II Flood Frequency Estimation
68
On the basis of average values, EV1 looks a more suitable candidate than LO or LN.
However, the occurrence of so many negative values of L-skewness among the observed data
throws doubt on the suitability of EV1 as a universal 2-parameter model for Ireland.
[Editorial note: It is understood that Station 08009 Ward at Balheary was omitted from
these simulations (see Box 4.4 below).]
4.5.3 Implication for choice of 3-parameter distribution
From Figure 4.5 it is seen that the average of the 110 data points falls roughly half way
between the GEV and GLO curves. GLO denotes the Generalised Logistic distribution. The
preliminary conclusion from the L-moment ratio diagrams is that, in addition to the
2-parameter EV1 distribution, the 3-parameter GEV and GLO distributions are possible
candidates to describe Irish flood data.
4.6 Exceptional recorded values and outliers
The largest flow values, Qmax, in each record and the dimensionless value Qmax/Qmean are
amongst the additional summary statistics presented in Appendix F.
4.6.1 Qmax/Qmean ratios
The largest Qmax/Qmean among the 45 Grade A1 stations is 5.17 at Station 08009 Ward at
Balheary (see Box 4.4). There are only two other stations where this statistic exceeds 3.0,
namely: Station 09002 Griffeen at Lucan and Station 09010 Dodder at Waldron’s Bridge.
Box 4.4: Station 08009 Ward at Balheary
The large AM value at Waldron’s Bridge occurred in August 1986 as a result of Hurricane
Charlie. Of all the exceptional flows which occurred in South Dublin and Wicklow at that
time, it is the only flow to have been measured. Even allowing for the difficulties of
measurement, and extrapolation of the rating curve, it is truly a very large flood in the Irish
context.
At the 42 remaining Grade A1 stations, and at all 70 Grade A2 stations, the Qmax/Qmean ratio
is less than 3.0. These values are extremely low by European standards. It should be noted
that these quantities depend on record length and not all records in the dataset are of equal
length, varying from 15 to 55 years.
The Station 08009 Ward at Balheary record is relatively short and there are several years
with incomplete data. Regardless of whether the record is short or long, the 1992/93
annual maximum of 53.6 m3s
-1 on 12 June 1993 – even allowing for difficulties of
measurement and extrapolation of the rating curve – is a large outlier. As a precaution,
the station’s data were not included in the inferences and summaries of preceding sections.
[Editorial notes: Section 10.4.5 uses the AM flow series at this station to illustrate the
difficulties that a high outlier presents to the analyst. The series is also discussed by
Bhattarai and Baigent (2009) who assign a much lower peak flow to the 12 June 1993
event and suggest that the unusually small gauged QMED at this station may be because
other large floods are missing from the record. Classification of a station as Grade A1 is
not always a guarantee that flow records are acceptably complete.]
Volume II Flood Frequency Estimation
69
4.6.2 High outliers
An outlier is an observation that lies an abnormal distance from other values in a random
sample from a population. In other words, an outlier lies significantly away from the general
range or pattern of the remaining data. The presence of outliers can create problems: both for
the summary statistics and in the interpretation of probability plots (see Chapter 5).
Most interest typically centres on high outliers, i.e. unusually large AM values. The 17
stations with the most obvious high outliers have been selected from the probability plots and
are listed in Table 4.3, where the large values are compared with Qmed rather than with Qmean.
Qmed is the index flood adopted in the FSU. [Editorial note: It is written QMED elsewhere
in the volume so that subscripts can be attached to it with greater clarity.]
It is seen from Table 4.3 that only three stations have a Qmax/Qmed ratio in excess of 3.0 and
only 12 stations have values in excess of 2.0. These numbers are low by comparison with
UK data, for instance.
Table 4.3: High outliers amongst 110 stations graded A1 + A2
Station
number
Station
grade
#
years
Largest
AM flow
[Qmax]
2nd
largest
[Qmax2]
3rd
largest
[Qmax3]
Qmed Qmax/
Qmed
Qmax/
Qmax2
Qmax2/
Qmax3
08009 A1 11 53.6 11.8 10.3 6.59 8.13 4.54 1.15
09010 A1 19 269.__ 156.__ 112.__ 48. 5.60 1.72 1.39
06031 A2 18 35.8 24.4 18.1 11.7 3.06 1.47 1.35
26006 A1 53 70.06 68.66 41.81 29.23 2.40 1.02 1.64
16002 A2 51 123.88 84.42 81.86 52.66 2.35 1.47 1.03
29011 A1 22 66.52 44.34 40.61 28.89 2.30 1.50 1.09
36015 A1 33 49.99 43.30 31.77 22.08 2.26 1.15 1.36
36031 A2 30 13.70 8.87 8.32 6.45 2.12 1.54 1.07
14005 A2 48 80.42 80.42 59.99 38.27 2.10 1.00 1.34
24022 A2 20 20.5 15.0 14.2 9.79 2.09 1.37 1.06
23012 A2 18 31.74 26.27 19.12 15.66 2.03 1.21 1.37
15004 A2 51 74.96 73.61 56.88 37.28 2.01 1.02 1.29
36021 A2 27 43.57 32.16 31.79 23.37 1.86 1.35 1.01
07033 A2 25 26.58 20.56 18.71 14.62 1.82 1.29 1.10
06011 A1 48 26.36 21.07 19.99 15.39 1.71 1.25 1.05
34001 A2 36 286.56 224.44 219.12 174.61 1.64 1.28 1.02
33070 A1 28 11.97 9.44 9.35 7.67 1.56 1.27 1.01
Volume II Flood Frequency Estimation
70
4.6.3 Low outliers
A low outlier is an AM value that is unusually small. This may arise from data error (e.g.
where the annual maximum derives from only a few months of gauged record) or a chance
absence of any high flow in that year (e.g. in 1976 in some permeable UK catchments). Low
outliers can influence skewness values unduly, leading to unsuitable model fits. The remedy
is to pay close attention to the probability plot (see Chapter 5).
Table 4.4 lists two stations with low outliers. Although not remarkably low, such outliers
may influence distributions fitted to the AM series. This highlights the importance of always
inspecting probability plots.
[Editorial note: Another series with a low outlier appears in EV1 Probability Plot 10.7 for
Station 26008 Rinn at Johnston's Bridge.]
Table 4.4: Low outliers amongst 110 stations graded A1 + A2
Station
number
Station
grade
#
years
Smallest
AM flow
[Qmin]
2nd
smallest
[Qmin2]
3rd
smallest
[Qmin3]
Qmed Qmed/
Qmin
Qmin2/
Qmin
Qmin3/
Qmin2
15003 A2 50 61.13 96.72 98.02 150.76 2.47 1.58 1.01
26059 A1 17 8.23 10.7_ 11.4_ 12.2 1.48 1.30 1.07
Volume II Flood Frequency Estimation
71
5 Probability plots … and inferences from them
5.1 Introduction
5.1.1 The idea of a probability plot
Probability plots are useful for the display and analysis of flood data, particularly to
determine whether or not a given sample is consistent with a particular population
distribution. The plots employ an inverse distribution scale so that a simple (usually
2-parameter) cumulative distribution function (CDF) plots as a straight line. On this scale,
points (xi, yi) for i = 1, 2, ..., n are expected to lie close to the line y = a + b x, where a and b
are respectively the location and scale parameters of the distribution. Conversely, strong
deviation from the line is evidence that the distribution did not produce the data.
What constitutes “strong deviation” from such a line often has to be judged subjectively.
Nevertheless, probability plots are widely used in flood hydrology for data display and
exploratory analysis. Experienced analysts use them to inform – rather than determine – the
final choice of distribution.
5.1.2 Synthetic examples
EV1 Probability Plot 5.1 provides a small selection of probability plots of samples of size 25
drawn randomly from an EV1 population. EV1 Probability Plot 5.2 does the same for
samples of size 50. In a proportion of cases, the plots reveal departures from a linear pattern
that might – on the basis of visual inspection – lead the analyst to reject the EV1 hypothesis.
This illustrates that probability plots are no guarantee of taking a good decision.
For real samples of annual maximum flood data of size 25 or 50, we do not know the
population from which they are drawn. Hence, inferences from probability plots have to be
considered carefully.
5.1.3 Standard plots for the FSU flood peak data
Three versions of probability plot were considered based on the Gumbel (EV1), the
2-parameter Logistic (LO) and the 2-parameter lognormal (LN) distributions. Plots were
drawn and studied for the AM flows of 186 Irish stations. The flood series range in length
from eight to 55 years and the associated catchment areas range from 10.1 to 2780 km2.
Appendix G shows a sample collection of EV1, LO and LN probability plots and summary
information for Station 14018 Barrow at Royal Oak. Further examples of EV1 probability
plots appear throughout the volume, most notably in Section 10.4 and Appendix J.
Volume II Flood Frequency Estimation
72
EV1 Probability Plot 5.1: Nine random samples of size 25 drawn from EV1
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
-2 -1 0 1 2 3 4 5
EV1 y
Q
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
EV1 reduced variate, y
EV1 reduced variate, y
EV1 reduced variate, y
EV1 reduced variate, y EV1 reduced variate, y
EV1 reduced variate, y EV1 reduced variate, y
EV1 reduced variate, y EV1 reduced variate, y
Random samples of size 25 drawn from
an EV1 population with
mean = 100 and SD = 30
(i.e. CV = 0.333)
Volume II Flood Frequency Estimation
73
EV1 Probability Plot 5.2: Nine random samples of size 50 drawn from EV1
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
400
450
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5
EV1 y
Q
0
50
100
150
200
250
300
350
-2 -1 0 1 2 3 4 5
EV1 y
Q
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
Val
ue
EV1 reduced variate, y
EV1 reduced variate, y
EV1 reduced variate, y
EV1 reduced variate, y EV1 reduced variate, y
EV1 reduced variate, y EV1 reduced variate, y
EV1 reduced variate, y EV1 reduced variate, y
Random samples of size 50 drawn from
an EV1 population with
mean = 100 and SD = 30
(i.e. CV = 0.333)
Volume II Flood Frequency Estimation
74
5.1.4 Approach taken to assessing plots
It is acknowledged that inferences drawn from probability plots have low statistical power,
i.e. there is a high probability that a conclusion could be adopted even though it is not true.
Nevertheless probability plots have played a prominent role in hydrological frequency
analysis in the past and it is considered worthwhile to use them in this study as a way of
viewing the available data. Plotted data for the 110 A1 + A2 stations were therefore
investigated from the standpoint of both linearity (Section 5.2) and curvature (Section 5.3).In
the linearity investigation, a straight line is fitted to the plotted data by least-squares
regression. The perceived quality/adequacy of fit is scored by personal (visual) judgement,
with 1 denoting a very poor fit and 5 denoting a very good fit.
In the curvature investigation, an assessment of the curve pattern of the plotted data is made
by personal (visual) judgement and the pattern classified.
5.2 Linear patterns
For each (A1 and A2) station in turn, and for each type of plot (EV1, LO and LN), the
probability plot was examined from the point of view of linearity and subjectively scored
from 1 to 5 by visual judgement, where:
1 = very poor (fit)
2 = poor
3 = moderate
4 = good
5 = very good (fit)
The linear pattern scores for the 110 A1 + A2 stations are given in Appendix H and
summarised in Table 5.1.
Table 5.1: Linear pattern statistics for 110 A1 + A2 stations
A1 stations A2 stations A1 + A2 stations
Type of plot EV1 LO LN EV1 LO LN EV1 LO LN
Modal Score 4 2 3 4 3 4 4 3 4
Mean Score 3.3 2.6 3.3 3.4 3.1 3.4 3.3 2.9 3.3
Table 5.1 reveals that EV1 and LN give almost identical results and that either provides a
higher score (i.e. better fit) than by LO. It can be deduced from the linear patterns that –
amongst 2-parameter distributions – Irish flood data are more likely to be distributed as EV1
or LN than as LO.
5.3 Curve patterns
5.3.1 Assignment of patterns
For each (A1 and A2) station in turn, and for each type of plot (EV1, LO and LN), the
probability plot was judged from the point of view of curvature and a curve pattern assigned.
Volume II Flood Frequency Estimation
75
The assessment scheme was based on four curve patterns: linear, concave, convex and
S-curve (i.e. an elongated S shape) coded as:
L straight line
U concave upwards
D concave downwards (i.e. convex upwards)
S S-curve
To allow greater discrimination, each pattern is subdivided into 1 = mild and 2 = severe
categories. Finally, a further classification X (for extremes) is added if the patterns of the end
points of the plot differ from the pattern of the main body of the data. In summary, the
15 curve patterns allowed are:
L1 perfectly straight line
L2 little deviation from straight line
L2X body pattern is quite straight but with end disturbance
U1 mild concave upwards
U2 severe concave upwards
U1X mild concave but with end disturbance
U2X severe concave but with end disturbance
D1 mild convex upwards
D2 severe convex upwards
D1X mild convex upwards but with end disturbance
D2X severe convex upwards but with end disturbance
S1 mild S-curve
S2 severe S-curve
S1X mild S-curve but with end disturbance
S2X severe S-curve but with end disturbance
The curve patterns for the 110 A1 + A2 stations are presented in Appendix H.
5.3.2 Probability plots for 43 Grade A1 stations
The curve patterns of the probability plots for the 43 Grade A1 stations are summarised in
Figure 5.1. The pie charts indicate the relative frequency of the different curve patterns. A
different pie is shown for each of the three probability plots considered: EV1, LO and LN.
Just over half of the EV1 plots in Figure 5.1a show a linear trend (Prefix L) and eight a
convex pattern (Prefix D). Only six stations exhibit a concave pattern (Prefix U). The
scarcity of concave patterns reflects that in many cases the largest floods on record are close
together in magnitude. In very few cases are there outstandingly large floods in the series.
The influence of the choice of probability plot on visual assessment is seen in the contrasting
judgements for the Logistic (LO) plots in Figure 5.1b. Concave patterns (Prefix U) are the
most prevalent. Many linear and S-curve patterns are also found but hardly any convex
(Prefix D) patterns.
The curve patterns for the lognormal (LN) probability plots are summarised in Figure 5.1c.
These show some similarity to those for the EV1 plots. A linear trend (Prefix L) is again the
Volume II Flood Frequency Estimation
76
dominant pattern. However, considerably more S-curve patterns and fewer convex patterns
(Prefix D) are found than for the EV1 plots.
(a) EV1 plots (b) LO plots (c) LN plots
Figure 5.1: Relative frequency of curve patterns – 43 Grade A1 stations
5.3.3 Probability plots for 110 Grade A1+ A2 stations
When results for the 67 Grade A2 stations are included, half of the EV1 plots (see Figure
5.2a) continue to exhibit a linear trend (Prefix L). However, the proportion exhibiting a
concave pattern (Prefix U) is reduced and is now outstripped by S-curve patterns (Prefix S) as
well as by convex (Prefix D) patterns.
(a) EV1 plots (b) LO plots (c) LN plots
Figure 5.2: Relative frequency of curve patterns – 110 Grade A1 + A2 stations
Concave patterns (Prefix U) remain the most prevalent in the LO plots (Figure 5.2b), with
convex patterns (Prefix D) remaining scarce. Across the 110 A1 + A2 stations, not even one
of the LO probability plots was judged a perfect straight line (Code L1).
The curve patterns in the LN plots (Figure 5.2c) are largely as for the Grade A1 stations
alone. However, convex patterns (Prefix D) are somewhat more prevalent, and concave
patterns (Prefix U) somewhat less prevalent, than in the smaller dataset.
In many cases it was observed that those stations which fit quite straight on EV1 plots show a
concave (Prefix U) fit on Logistic paper. The reason might be due to the range of the
probability axis, because the probability axis of LO is relatively compressed compared to
EV1. That is why it was also noticed that those which fit as convex (Prefix D) on EV1 plots
are found to be broadly linear on LO.
Volume II Flood Frequency Estimation
77
5.3.4 Curve pattern in relation to skewness coefficient
Understandably, there is a link between the skewness of a data sample and the curve pattern
of the resulting probability plot. Stations are ordered in Table 5.2 according to Hazen
skewness. The curve patterns shown are those based on EV1 probability plots.
Fourteen of the 110 A1 + A2 stations (shown in bold) have negative Hazen skewness. It is
confirmed that:
Stations with high outliers tend to have high skewness;
Stations with low outliers tend to have low skewness;
EV1 plots for stations with H-skew < 0.2 predominantly have a convex (Prefix D)
curve pattern;
EV1 plots for stations with H-skew > 2.5 (shown in red) mostly have a concave
(Prefix U) curve pattern.
Table 5.2: EV1 plot curve patterns – 110 A1 + A2 stations ordered by Hazen skewness
H-s
kew
Station number
and river
# y
ears
Cu
rve
patt
ern
(EV
1 p
lot)
Ou
tlie
r
Sta
tion
gra
de
H-s
kew
Station number
and river
# y
ears
Cu
rve
patt
ern
(EV
1 p
lot)
Ou
tlie
r
Sta
tion
gra
de
-1.09 07006 Moynalty 19 D A2 0.78 34011 Manulla 30 S A2
-0.86 26021 Inny 30 D A2 0.81 06011 Fane 48 L High A1
-0.78 15003 Dinan 50 D Low A2 0.85 16003 Clodiagh 51 L A2
-0.48 25034 L Ennell trib 24 D A2 0.93 34018 Castlebar 27 L A1
-0.41 16009 Suir 52 D A2 0.93 07009 Boyne 29 L A1
-0.39 34024 Pollagh 28 D A2 0.93 26009 Black 35 S A2
-0.33 06025 Dee 30 D A1 0.95 10021 Shanganagh 24 L A1
-0.32 16008 Suir 51 D A2 0.96 15004 Nore 51 L High A2
-0.31 25124 Brosna 18 D A2 1.01 25030 Graney 48 S A1
-0.22 24082 Maigue 28 D A2 1.03 39009 Fern O/L 33 L A2
-0.16 36019 Erne 47 D A2 1.04 35005 Ballysadare 55 S A2
-0.13 25021 Brosna 44 D A2 1.05 26007 Suck 53 L A1
-0.12 32012 Newport 24 D A2 1.05 06026 Glyde 46 S A1
-0.09 25029 Nenagh 33 S A2 1.05 36010 Annalee 50 S A1
0.07 35002 Owenbeg 34 D A2 1.08 34001 Moy 36 L High A2
0.10 06013 Dee 30 D A1 1.12 14007 Derrybrock 25 L A1
0.10 35001 Owenmore 29 L A2 1.12 24022 Mahore 20 L High A2
0.12 19020 Owennacurra 28 D A2 1.17 26019 Camlin 51 L A1
0.14 15001 Kings 42 D A2 1.17 09001 Ryewater 48 L A1
0.16 26020 Camlin 32 L A1 1.19 33070 Carrowmore 28 L High A1
0.18 25017 Shannon 55 D A1 1.22 23001 Galey 45 L A2
Volume II Flood Frequency Estimation
78
H-s
kew
Station number
and river
# y
ears
Cu
rve
pa
tter
n
(EV
1 p
lot)
Ou
tlie
r
Sta
tio
n g
rad
e
H-s
kew
Station number
and river
# y
ears
Cu
rve
pa
tter
n
(EV
1 p
lot)
Ou
tlie
r
Sta
tio
n g
rad
e
0.18 14029 Barrow 47 D A2 1.23 25044 Kilmastulla 33 S A2
0.19 14011 Rathangan 26 L A1 1.25 18004 Ballynamona 46 S A2
0.21 36012 Erne 47 D A1 1.26 34003 Moy 29 S A2
0.22 14018 Barrow 51 L A1 1.30 30007 Clare 31 L A2
0.27 25027 Ollatrim 43 D A1 1.31 06014 Glyde 30 L A1
0.27 26005 Suck 51 L A2 1.32 25040 Bunow 20 L A2
0.28 24008 Maigue 30 L A2 1.33 16005 Multeen 30 L A2
0.30 16001 Drish 33 L A2 1.37 27002 Fergus 51 L A1
0.31 26059 Inny 17 D Low A1 1.42 14009 Cushina 25 S A2
0.32 24002 Camogue 27 L A2 1.46 14006 Barrow 51 L A1
0.34 29001 Raford 40 D A1 1.49 26008 Rinn 49 L A1
0.34 35073 L. Gill 30 D A2 1.54 08005 Sluice 18 L A2
0.36 10022 Carrickmines 18 L A1 1.55 27001 Claureen 30 L A2
0.37 14013 Burrin 50 S A2 1.59 12001 Slaney 50 L A2
0.42 16011 Suir 52 D A1 1.63 18005 Funshion 50 L A2
0.43 36018 Dronmore 50 L A1 1.74 08008 Broadmeadow 25 L A2
0.43 34009 Owengarve 33 L A2 1.74 07033 Blackwater 25 S High A2
0.47 26022 Fallan 33 L A2 1.79 16002 Suir 51 L High A2
0.48 19001 Owenboy 48 L A2 1.85 36021 Yellow 27 L High A2
0.53 16004 Suir 48 L A2 1.92 31002 Cashla 26 U A1
0.57 25014 Silver 54 L A1 1.98 08002 Delvin 20 L A1
0.57 25023 Little Brosna 52 L A1 2.01 14005 Barrow 48 L High A2
0.57 25025 Ballyfinboy 31 L A1 2.23 14034 Barrow 14 L A2
0.59 15002 Nore 35 L A2 2.29 26002 Suck 53 S A2
0.61 14019 Barrow 51 D A1 2.40 09002 Griffeen 24 S A1
0.62 25016 Clodiagh 42 L A2 2.62 36015 Finn 33 U High A1
0.71 25006 Brosna 52 L A1 2.98 23012 Lee 18 U High A2
0.71 27003 Fergus 48 L A2 3.04 30061 Corrib Estuary 33 U A2
0.73 39008 Gartan 33 S A2 3.12 06031 Flurry 18 U High A2
0.74 29071 L. Cutra 26 L A2 3.24 09010 Dodder 19 U High A1
0.76 35071 L. Melvin 30 L A2 3.37 29011 Dunkellin 22 U High A1
0.77 26018 Owenure 49 L A2 3.87 26006 Suck 53 S High A1
0.77 29004 Clarinbridge 32 L A2 4.19 36031 Cavan 30 S High A2
0.78 06070 Muckno 24 L A1 5.43 08009 Balheary 11 U High A1
Volume II Flood Frequency Estimation
79
5.4 Flood volumes associated with largest peaks on convex probability plots
5.4.1 Stations with a convex curve pattern
For many of the stations which display a convex appearance on EV1 probability plots, it is
noted that the largest and 2nd
largest floods are little greater than the 3rd
largest. This
contributes to the flattening out of the probability plot and the low skewness of the data
sample. Such stations are identified in Appendix H as having a D curve pattern (D for
concave downwards). There are 26 such cases (i.e. stations exhibiting a convex EV1 plot)
amongst the 110 A1 + A2 stations.
If one were to include a few of the stations for which the pattern is classed as S-curve, the
convex behaviour is characteristic of a quarter of the A1 + A2 gauged catchments. The
behaviour is important in itself but also because a statistical distribution fitted to such data
may imply an absolute upper bound to flood magnitudes that is little greater than the largest
flood already observed. Investigation of such controversial cases is clearly warranted.
5.4.2 Hypothesis
In the mildly graded rivers characteristic of much of Ireland, the tendency for the largest
flood peaks to be of similar magnitude is typically attributed to the influence of floodplain
storage (e.g. Mason 1992; Ahilan et al. 2012). The theory is that the magnitude of large
flood peaks is naturally limited by the wide floodplains that become available to water when
river flow exceeds the bankfull capacity of the river system.
A specific attempt was therefore made to test the hypothesis that large floods of similar
magnitude might differ volumetrically. Could flood volume (amongst the highest floods) be
increasing, even though the peak discharge is not?
5.4.3 Arrangement of study
To investigate this question, hydrograph volumes have been calculated for the three or four
largest AM flood peaks at seven stations. These are drawn from the 26 stations identified as
yielding a convex curve pattern (Prefix D) on EV1 probability paper (see Appendix H).
The availability of hydrograph data was a factor in the selection of the examples listed in
Table 5.3. Comparison of the relevant periods of record in the final columns of the table
indicates that hydrograph data are mainly available for the required periods. However, there
were some exceptions. The hydrographs found are presented in Appendix J.
[Editorial note: An eighth station – the Deel at Rathkeale (24013) – had been identified as
having a strongly convex probability plot. However, it transpired that pre- and post-drainage
records had been inadvertently combined. Ahilan et al. (2012) also fail to distinguish pre and
post-drainage records at this station. The nine lowest AM flows in their frequency analysis of
49 annual maxima for the Deel at Rathkeale derive from the ten years of pre-drainage record.
The post-drainage series is retained in the analysis of hydrograph volumes reported in
Appendix J, although EV1 Probability Plot J.5 is found to be only weakly convex.]
Volume II Flood Frequency Estimation
80
Table 5.3: Stations for which flood volumes were specially investigated
Station
number Station name
Station
grade
Hazen
skewness
Period for which:
Hydrograph
data available
AM series
analysed
07006 Moynalty at Fyanstown A2 -1.09 1956 – 2005 1986 – 2004
15003 Dinan at Dinan Bridge A2 -0.78 1972 – 2005 1954 – 2004*
16008 Suir at New Bridge A2 -0.32 1954 – 2005 1954 – 2004
16009 Suir at Cahir Park A2 -0.41 1940 – 2005 1953 – 2004
24013 Deel at Rathkeale A1 0.05 1972 – 2003 1969 – 2004
24082 Maigue at Islandmore A2 -0.22 1975 – 2001 1977 – 2004
25017 Shannon at Banagher A1 0.18 1989 – 2003 1950 – 2004
25021 Little Brosna at Croghan A2 -0.13 1961 – 2003 1961 – 2004 *AM flood for 2001 water-year missing at this station
5.4.4 Calculation of hydrograph volumes
On examination of hydrographs for the three or four largest AM floods, it is seen that some
exhibit a unimodal (i.e. one-peaked) hydrograph reflecting their likely origin in a single
period of heavy rainfall. However, many of them occur as a result of longer and more
fluctuating periods of rainfall which lead to a multimodal hydrograph. In consequence, it is
not practical to specify a unique time duration over which to evaluate the hydrograph volume.
Hydrograph volumes were therefore calculated across a number of time durations or
windows. It should be noted that these are set symmetrically about the time of the peak. The
windows correspond to ±12 hours, ±24 hours, ±84 hours, ±1 week and ±15 days, leading to
window sizes of 1, 2, 7, 14 and 30 days respectively.
5.4.5 Example
The method is illustrated for Station 24082 Maigue at Islandmore. This 764 km2 catchment
has 28 annual maxima (1977 to 2004); see EV1 Probability Plot 5.3. The hydrographs for the
four largest AM floods are shown in Figure 5.3. The Hazen skewness is -0.22.
EV1 Probability Plot 5.3: Station 24082 Maigue at Islandmore
2 5 10 25 50 100 500
0
50
100
150
200
250
-2 -1 0 1 2 3 4 5 6 7
AM
flo
w (
m3
s-1
)
EV1 reduced variate, y
winter peak
summer peak
Feb
'90 Nov
'00 Oct
'88 Dec
'98
Volume II Flood Frequency Estimation
81
Figure 5.3: Hydrographs of four largest AM events, Station 24082 Maigue at Islandmore
Hydrograph during Maximum Flood in the year 1989
0
50
100
150
200
250
01/02/1990 00:00 03/02/1990 00:00 05/02/1990 00:00 07/02/1990 00:00 09/02/1990 00:00 11/02/1990 00:00
Time
Dis
charg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 2000
0
20
40
60
80
100
120
140
160
180
200
31/10/2000 00:00 02/11/2000 00:00 04/11/2000 00:00 06/11/2000 00:00 08/11/2000 00:00 10/11/2000 00:00 12/11/2000 00:00
Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1988
0
20
40
60
80
100
120
140
160
180
200
14/10/1988
00:00
16/10/1988
00:00
18/10/1988
00:00
20/10/1988
00:00
22/10/1988
00:00
24/10/1988
00:00
26/10/1988
00:00
28/10/1988
00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1998
0
20
40
60
80
100
120
140
160
180
200
23/12/1998 00:00 25/12/1998 00:00 27/12/1998 00:00 29/12/1998 00:00 31/12/1998 00:00 02/01/1999 00:00 04/01/1999 00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Rank 1 AM event,
6 February 1990
Rank 2 AM event,
6 November 2000
Rank 3 AM event,
21 October 1988
Rank 4 AM event,
30 December 1998
Flo
w (
m3 s
-1)
Flo
w (
m3 s
-1)
Flo
w (
m3 s
-1)
Flo
w (
m3 s
-1)
±12 hr, , ±84 hr from peak
Volume II Flood Frequency Estimation
82
The flood volumes across 1, 2, 7, 14 and 30-day windows are illustrated in Figure 5.4. The
label Rank 1 AM flood denotes the event with the greatest peak flow. The 1-day and 2-day
flood volumes are noticeably similar across these four events. The Rank 1 AM flood event
yields the largest volume at all durations, but the difference is marked only at the longer
durations of 7, 14 and 30 days.
Figure 5.4: Hydrograph volumes of four largest AM events at 24082 Maigue at Islandmore
5.4.6 Summary of findings
Appendix J provides results for the eight stations studied. Their behaviours are little different
to the example above. No marked trend is noted in the 1 and 2-day flood volumes of the
hydrographs associated with the largest AM flood peaks.
It is concluded that flood volumes do not reveal any noticeable growth with rank among the
highest-ranking floods in those AM series which display convex patterns on EV1 (and LN)
probability plots. In other words, no explanation for the flattening out (of the upper part) of
the probability plots can be found among the flood volumes. The hypothesis put forward in
Section 5.4.2 is therefore not supported.
Because of the importance of such effects, it was decided to make a further check on the
hypothesis by examining whether the skewness of AM flood series might be related to the
Flood Attenuation Index developed in the FSU (see Chapter 3 of Volume IV).
5.4.7 A further check on the hypothesis
The Flood Attenuation Index, FAI, falls in the range 0 to 1. Values of the index indicate the
fraction of the catchment area potentially inundated in very large floods. Under the
Section 5.4.2 hypothesis that floodplain storage tends to lead to convex EV1 probability
plots, it is expected that catchments with high values of FAI might be associated with convex
curve patterns.
It was seen in Table 5.2 that skewness is a relatively good indicator of the general curve
pattern of the EV1 probability plots. Convex patterns predominate when the Hazen skewness
is less than 0.2, whilst concave patterns predominate when the Hazen skewness is greater
than 2.5.
Volume of hydrographs of different year during max peak
0
20
40
60
80
100
120
140
160
1 2 7 14 30Days
Mill
cu
. met
er
1998
1988
2000
1989
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Feb 1990
Rank 4 AM flood, Dec 1998
Rank 2 AM flood, Nov 2000
Rank 3 AM flood, Oct 1988
Volume II Flood Frequency Estimation
83
Extracting FAI values for the 110 A1 + A2 stations and plotting them against Hazen
skewness, it is demonstrated emphatically in Figure 5.5 that the Flood Attenuation Index
provides no explanation for the convex curve pattern found in 26 of the 110 EV1 probability
plots. The hypothesis put forward in Section 5.4.2 is finally rejected.
6543210-1-2
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Hazen skewness
FA
I
LS S
L
DL
U
DSD L
SDLD
LS
LSLL
D U
U
LL ULD
LL
L
DL
DL
LL SL
L SL SD
SL
D
SS D
LLD
D
LLLD L
LL U
L
DL L
SDDD L
L
L
LL L
D
LD
LDD
L
SL
S
L
L
L
L
L
LU
S
L U
L
LL
S
LD
L
U
SD LD
L
Mainly convex Mainly concave
Mainly linear or S-curve
Figure 5.5: FAI v. Hazen skewness, labelled by EV1 curve pattern (110 A1 + A2 stations)
5.5 Flood seasonality
5.5.1 Seasonality of annual maximum floods
Annual maximum flood data were available for a total of 203 Grade A1, A2 or B stations.
One station did not have dates attached to the annual maximum flood data. Although the
flood peak data for some stations were subsequently rejected from wider analysis (see
Section 4.2.1 and Appendix B), the worries did not in general compromise the date
information used to characterise flood seasonality. Consequently, the seasonality study
reported here is based on 202 stations. Collectively, the series provide a total of 6969 station-
years of data.
Although floods are seen to occur at all times of year, most rivers register their annual
maximum in the winter (October-March) half-year. The percentages of AM floods occurring
in the winter half-year at each station are listed in Table K.1 of Appendix K. The months
December and January are associated with the greatest number of AM flood events followed
by November and February. In all, 6094 of the 6969 annual maxima occurred in the winter
half-year. In the summer half-year, considerable numbers of flood peaks were observed in
August. July has the least number of AM floods. The general pattern of flood seasonality is
summarised in Histogram 5.1.
Histogram 5.1: Seasonal occurrence of annual maximum floods
Seasonal Flood Frequency
Total no. of station
years: 6969
0
200
400
600
800
1000
1200
1400
1600
1800
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun JulMonth
No
. o
f O
cc
ura
nc
es
Num
ber
of
occ
urr
ence
s
Based on 6969 station-years
across 202 A1 + A2 + B stations
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul
Volume II Flood Frequency Estimation
84
5.5.2 Seasonality of largest floods
The months of occurrence of the maximum flow in each flood series are given in Table K.2
and summarised in Histogram 5.2. Most of the series maxima occurred in the winter half-
year, with the single month of December provided the maximum at 91 of the 202 stations. At
only 20 stations was the maximum recorded flood in the summer half-year. However, the
single month of August supplied the series maximum at eight of the 202 stations examined.
Histogram 5.2: Month corresponding to series maximum flow (202 A1 + A2 + B stations)
5.5.3 Circular diagrams
The circular diagram in Figure 5.6 illustrates the seasonal distribution of AM flood peaks
together with their magnitudes in m3s
-1. The angular position indicates the flood month and
the radial distance shows the magnitude of the flood peak.
Figure 5.6: Seasonality and magnitudes of 6969 AM floods (202 A1 + A2 + B stations)
Month corresponding to Max. flow in an AM series
0
10
20
30
40
50
60
70
80
90
100
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov DecMonth
No
. o
f o
cc
ure
nc
es
Num
ber
of
occ
urr
ence
s
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
800
Oct
Apr
JanJul
200
400
600
-800 -340 120 580 January
October
July
April
Radial distance indicates
AM flood peak in m3s
-1
Volume II Flood Frequency Estimation
85
5.6 Flood statistics on some rivers with multiple gauges
As part of exploratory data analysis it is valuable to examine flow data from cases where
several long-term gauging stations exist on the same river. This gives an opportunity to see
how flood magnitude increases in the downstream direction.
5.6.1 Down-river growth in QMED
Flood data for multiple stations on the Barrow, Suir, and Suck rivers are considered here. It
is seen from Figure 5.7 that QMED increases down-river in a generally systematic manner.
The variation with catchment size (Figure 5.7a) is confirmed to be more regular than the
variation with distance along the main channel (Figure 5.7b). MSL denotes the mainstream
length. The down-river growth in QMED is seen to be more pronounced on the Suir than on
the Barrow or Suck.
500020001000500200100
500
200
100
50
20
10
AREA (km )
QM
ED
Barrow
Suir
Suck
2 2001005020
500
200
100
50
20
10
MSL (km)
QM
ED
Figure 5.7: Variation of QMED down the Barrow, Suir and Suck
5.6.2 Down-river variation in probability plots
Down-river variation in probability plots is found to be relatively complex (see EV1
Probability Plot 5.4). This is especially the case for the Barrow, although the more marked
differences in pattern are in the less important LH part of the plot. The arrow indicates the
down-river sequence of stations.
The probability plots for the three most upstream stations on the Suir are similar in shape
through much of the range. [Editorial note: O’Sullivan et al. (2012) study the reach from
New Bridge to Cahir Park.] The plots for the other two stations are notably steeper,
especially for the most downstream station (16011 Suir at Clonmel). This highlights the
important role of tributaries in influencing flood frequency on the Lower Suir.
The probability plots for stations on the Suck are notable for their mild slope over much of
the range but with a kick in the RH part of the plot, reflecting the scope for occasional much
larger flood peaks. The disturbance is more marked at upstream stations than at downstream
stations.
QM
ED
(m
3s-1
)
AREA (km2) MSL (km)
● Barrow
● Suir
● Suck
● Barrow
● Suir
● Suck
(a) Variation with catchment size (b) Variation with distance along main channel
Volume II Flood Frequency Estimation
86
EV1 Probability Plot 5.4: Multiple stations on the Rivers Barrow, Suir and Suck
ST14005
ST14006
ST14019
ST14018
ST14029
Probability Plot Summary for stations on Barrow River
2 5 10 25 50 100 5000
50
100
150
200
250
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
14005
14006
14018
14019
14029
Qmed vs Catchment Area relationship for stations on Barrow River
0
20
40
60
80
100
120
140
160
180
0 500 1000 1500 2000 2500 3000
Catchment Area (km2)
Qm
ed
(m
3/s
)
16002
16008
16009
16011
Probability plot summary for stations on Suir River
50010050251052
0
50
100
150
200
250
300
350
400
450
500
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
16004
16002
16008
16009
16011
Qmed vs Catchment Area Relationship for stations on Suir River
0
50
100
150
200
250
0 500 1000 1500 2000 2500
Catchment Area (km2)
Qm
ed
(m
3/s
)
26002
26005
26007
Probability Plot Summary for stations on Suck River
2 5 10 25 50 100 500
0
20
40
60
80
100
120
140
160
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
) 26006
26002
26005
26007
Qmed vs Catchment Area relationship for stations on Suck River
0
10
20
30
40
50
60
70
80
90
100
0 200 400 600 800 1000 1200 1400
Catchment Area (km2)
Qm
ed
(m
3/s
)
Suir
● 16011 Clonmel
16009 Cahir Park
× 16008 New Bridge
▲ 16002 Beakstown
+ 16004 Thurles
EV1 reduced variate, y
EV1 reduced variate, y
EV1 reduced variate, y
Annual
max
imum
flo
w (
m3s-1
) A
nnual
max
imum
flo
w (
m3s-1
) A
nnual
max
imum
flo
w (
m3s-1
)
Barrow
● 14029 Graiguenamanagh
× 14018 Royal Oak
14019 Levitstown
▲14006 Pass Bridge
+ 14005 Portarlington
Suck
26007 Bellagill Br
× 26005 Derrycahill
▲26002 Rookwood
+ 26006 Willsbrook
Return period in years
Return period in years
Return period in years
Volume II Flood Frequency Estimation
87
6 Determining T-year flood magnitude QT by index flood method
6.1 Introduction
Important developments in flood frequency analysis in the early 20th
Century were made by
Fuller (1914), Foster (1924) and Hazen (1930), amongst others. Gumbel (1941) brought the
basis of analysis to a new level by applying extreme value theory and the findings of Fisher
and Tippet (1928). Gumbel introduced the Extreme Value Type I distribution to flood
frequency analysis. This distribution is often known as the Gumbel distribution but is
denoted EV1 in this report. Powell (1943) introduced extreme value probability paper, which
in one form or another (ruled in terms of probability or reduced variate) is still widely in use
today. Then Langbein (1949) provided the link between the T-year flood estimated by
annual maximum analysis and that derived from partial duration (i.e. peaks-over-threshold)
models.
The concept of standard error and bias of statistical estimates of QT was introduced by
Kimball (1949) for the EV1 distribution and by Kaczmarek (1957) for EV1 and other
distributions although this information filtered only slowly into hydrological practice. The
first simulation-based studies of standard errors were reported by Nash and Amorocho (1966)
and by Lowery and Nash (1970). By the mid-1970s, extensive simulation studies on bias and
standard error were reported by e.g. Matalas et al. (1975). Many analogous studies followed,
up to including that by Lu and Stedinger (1992).
Since Fuller (1914), flood magnitudes have been estimated by fitting distributions, either
graphically or numerically, to series of annual maximum floods treated as if they are random
samples from known distributions. The T-year return period flood QT is the flood magnitude
which is exceeded on average once in every T years and has a probability of being exceeded
in any one year of p = 1/T. p is the so-called annual exceedance probability.
Let F(Q) be the cumulative distribution function of AM flood magnitudes. Then the non-
exceedance probability of magnitude Q in any one year is:
F(Q) = 1 – 1/T 6.1
so that
QT = F-1
(1-1/T) 6.2
The cumulative distribution function (CDF) is sometimes referred to as simply the
distribution function. Its connection with the probability density function is illustrated in
Figure 6.1.
Figure 6.1: Probability density function and cumulative distribution function
Probability Density Function
0
0.1
0.2
0.3
0.4
Pro
babili
ty D
ensity F
unction
area = 1/Tarea =
1 - 1/T
QT
Distribution Function
0
0.2
0.4
0.6
0.8
1
No
n E
xce
ed
an
ce
Pro
ba
bili
ty, F
(q)
F(q) = 1 - 1/T
1 - F(q) = 1/T
QT
Probability Density Function
0
0.1
0.2
0.3
0.4
Pro
babili
ty D
ensity F
unction
area = 1/Tarea =
1 - 1/T
QT
Distribution Function
0
0.2
0.4
0.6
0.8
1
No
n E
xce
ed
an
ce
Pro
ba
bili
ty, F
(q)
F(q) = 1 - 1/T
1 - F(q) = 1/T
QT
(b) Cumulative distribution function (a) Probability density function
Volume II Flood Frequency Estimation
88
6.2 Regional flood frequency analysis and the index flood approach
6.2.1 Index flood approach
Regional or geographical pooling of dimensionless flow values was introduced by Dalrymple
(1960) to provide regional growth factors, xT, for use at ungauged sites. The flood frequency
curve, QT, is estimated in the form:
QT = QI xT 6.3
where QI is the index flood and xT is the flood growth factor. The index flood is usually
taken as the mean or median of the AM flow series. The index flood adopted in the FSU is
the median annual flood, QMED.
The FSR (NERC, 1975) introduced the idea that the pooled approach embodied by
Equation 6.3 should be used for QT estimation at gauged sites if T > 2N, where N is the
number of years of gauged record. More stringently, the FEH recommended that the pooled
approach of Equation 6.3 should prevail if T > N and should play at least some role
(alongside single-site analysis) when N/2 ≤ T ≤ N (Reed, 1999).
In between publication of FSR in 1975 and FEH in 1999, considerable research was carried
out on the efficiency of the index flood approach, including: Greis and Wood (1981),
Hosking et al. (1985a and 1985b), Lettenmaier et al. (1987), Cunnane (1989) and Hosking
and Wallis (1997). The index flood approach is now widely accepted as a valid and well
understood method, both in its implementation and its properties.
6.2.2 Type of region for pooling flood data
Dalrymple (1960) and FSR (NERC, 1975) applied the index flood method in fixed
geographical regions. All sites in the region are deemed to share a common flood growth
curve, xT.
In contrast, the Region Of Influence (ROI) method developed by Burn (1990) defines a region
centred on the specific (subject) catchment. Gauged catchments are selected to join the
region according to their nearness to the subject catchment in geographic or catchment-
similarity space. The ROI method was favoured in the FEH project, where it was found to be
more effective than the use of fixed geographical regions (Jakob et al., 1999).
6.2.3 Choice of index flood
The index flood approach uses a typical flood such as the mean (Qmean) or median (Qmed) of
the annual maximum flows as the index variable. FSR chose Qmean (often written QBAR) for
flood frequency but adopted the 5-year return period rainfall depth, M5, as the index variable
for rainfall frequency. In the present study, Qmed is taken as the index flood and is generally
written QMED.
QMED has a fixed frequency. It is the 2-year return period flood on the annual maximum
scale. One in every two annual maxima exceeds QMED.
Volume II Flood Frequency Estimation
89
6.2.4 A 2-stage approach
Use of the index flood approach to estimate QT by Equation 6.3 involves
Estimation of the index flood for the site in question, i.e. the subject site;
Estimation of the growth factor xT.
The graph of xT against T is referred to as the growth curve.
These two steps are more or less independent and each is dealt with separately. QMED
estimation from physical catchment descriptors (PCDs) was dealt with in Chapter 2. Though
a crucial building block, the relevant model for estimating QMED from PCDs (i.e.
Equation 2.8) is rarely applied in isolation. Practical application to estimate QMED at an
ungauged subject site requires both estimation of QMED from PCDs and a data transfer from
the most relevant gauged site. The latter is known as the pivotal catchment.
Though simple in principle, data transfers from gauged to subject sites tend to be complex in
practice. Section 2.7 provides an example, with further discussion chiefly deferred until
Chapter 11. The second stage of QT estimation by the index flood approach is estimation of
the flood growth curve, xT.
Volume II Flood Frequency Estimation
90
7 Flood growth curve estimation
7.1 Introduction to xT
The growth factor, xT, is the factor which when multiplied by the index flood QI gives QT, the
flood magnitude of return period T. In this study QMED is used as the index flood. Thus:
QT = QMED × xT 7.1
The relationship between xT and T is generally referred to as the flood growth curve.
The chapter begins by outlining two methods of determining the growth factor xT and the
resulting QT. The effects of catchment type, geographical location and period of record on xT
and QT are then examined. The effectiveness of different combinations of PCDs in defining
the distance measure on which ROI pooling groups are formed is examined in Appendix L.
7.2 Single-site and pooled estimates of QT
7.2.1 Advantages and drawbacks of the two approaches
If a subject site coincides with a gauging station – and if a sufficiently long record of AM
flows exists at the station – a suitable distribution can be fitted to the data of that site for
direct estimation of QT. This means that QT can be estimated without going through the
intermediate steps of determining QMED and xT. The FSU refers to this direct estimation of
QT as single-site analysis. Some practitioners call it at-site analysis.
What constitutes a sufficiently long record is arguable. The FSR (NERC, 1975) adopted the
criterion N > 0.5T, where N is the number of years of data available (i.e. the record length)
and T is the required return period. In contrast, the FEH adopted the much more stringent
criterion N > 2T (Reed, 1999). Introductory remarks and background information about the
general topic of QT estimation were presented in Chapter 6.
If the record is insufficiently long, the standard error of estimate of QT i.e. SE(QT) will be
large. Adopting a pooled estimate of xT, and hence of QT, reduces SE(QT).
The choice between a single-site estimate and a pooled estimate also involves a trade-off
between the convenience of single-site estimation and the extra work involved in assembling
and analysing data for a pooled estimate. While a pooled estimate of xT has a smaller
standard error, the necessary assumption of pooling-group homogeneity is uncomfortable.
7.2.2 Choice of distribution
When undertaking single-site analysis, another question arises as to what constitutes a
suitable distribution to describe flood frequency. Generally, it is better to adopt a
2-parameter distribution than a 3-parameter distribution because SE(QT) is much smaller in
the former case even though the bias may be increased.
Some guidance on the choice of 2-parameter distributions is available from Chapter 5, where
EV1 and lognormal (LN) distributions are seen to be typically better descriptors of gauged
AM data in Ireland than the 2-parameter Logistic (LO). However, many samples of data do
not seem amenable to description by a 2-parameter distribution. Such cases include instances
Volume II Flood Frequency Estimation
91
where the probability plot exhibits marked concave, convex or S-shaped curvature. These
cases were denoted in Chapter 5 by the curve-pattern classifications U (for upwards concave),
D (for downwards concave) and S (for S-shaped). While it is possible that this is a period-
of-record effect that will be neutralised when much longer records are available, it seems
more likely that the cases represent a genuine and permanent departure from 2-parameter
behaviour.
One approach might be to avoid using a single-site estimate at a station where the sample
shows a strong departure from 2-parameter behaviour and to adopt a pooled estimate instead,
on the basis that this reflects an average type of Q~T behaviour for the type of catchment in
question. It should be recalled from EV1 Probability Plot 5.1 that random samples drawn
from a 2-parameter distribution can display non-linear behaviour in a small proportion of
cases.
7.3 Pooling groups
7.3.1 The idea of pooling
Although xT can be estimated at a gauged site by conventional flood frequency analysis of the
AM flood series there, it is generally accepted that an advantage can be gained – in terms of
reduction of standard error of estimate of QT – if the assumption is made that xT does not vary
between gauging stations that belong to a homogeneous pooling group. A pooling group of
gauging stations is said to be homogeneous if the same growth curve applies to all stations in
the group (or, even more strictly, if the same value of xT for stated T can be applied to all
stations in the group).
The regional approach of pooling data exploits a trade-off between space and time. If the
group of stations is fully homogeneous, n years of record at each of m stations can be taken to
be equivalent to n × m years of record at a single site. Hence, if xT is estimated from the
n × m AM values – after they have been suitably standardised (e.g. by division by their
respective QMED values) – it will have a smaller standard error of estimate than an xT value
obtained by single-site analysis.
For a perfect case, the reduction in the standard error is by a factor of the order of m0.5
. The
gain is more limited if the pooling group is only weakly homogeneous or if the period of
record available at the stations is not very typical of the longer term.
7.3.2 Construction of pooling groups
Pooling groups can be formed by using geographical regions but Jakob et al. (1999,
specifically Figure 16.5) found such pooling groups to be typically less homogeneous than
those formed by a region-of-influence (ROI) approach of the type proposed by Burn (1990).
The ROI approach forms a pooling group for a particular subject site by selecting the nearest
stations in catchment-descriptor space. The space is designed to reflect the hydrological
similarity of catchments.
In choosing a pooling scheme, a decision has to be made about the PCDs to be included in
the metric (i.e. the distance measure dij) and what weightings and transformations are to be
applied to them. The FEH pooling scheme provides an example (see Box 7.1).
Volume II Flood Frequency Estimation
92
Box 7.1: FEH pooling scheme
7.3.3 The 5T rule
In addition, a decision has to be taken about how many stations to include in the pooling
group. Jakob et al. (1999) investigated pooling groups of various sizes before adopting the
5T rule. This states that at least 5T station-years of data should be used when estimating the
T-year flood.
Adoption of such a rule is a compromise. If too few station-years of data are pooled,
precision of the xT estimate is sacrificed. If too many stations are included, the assumption of
homogeneity may be compromised. Hosking and Wallis (1997) show that a small departure
from homogeneity can be tolerated. Having slightly too many stations included is therefore
deemed preferable to pooling too few station-years of data.
The 5T rule is adopted here as a pragmatic means of determining how many stations to
include in the xT estimation. It is recognised that it may not be practical or appropriate to
apply the 5T rule in every situation.
7.3.4 Recommended metric for pooling-group construction
Research on pooling schemes is reported in Appendix L. The distance measure (or metric)
ultimately recommended is:
2
nBFI
ji
2
nSAAR
ji
2
nAREA
ji
ijσ
nBFInBFI
σ
nSAARnSAAR
σ
nAREAnAREAd
7.2
Here, σℓnAREA, σℓnSAAR and σℓnBFI are the standard deviations of ℓnAREA, ℓnSAAR and ℓnBFI.
Values to be used in implementation are 1.265, 0.173 and 0.219 respectively.
The use of AREA and SAAR mirrors the FEH scheme of Box 7.1, in full knowledge that
both feature strongly in the estimation of QMED from PCDs (see Section 2.2). The third
variable in the distance metric is gauged BFI. This differs in that values of the baseflow
index (BFI) derived from daily mean flow data are used in preference to estimates of BFI
from soil modelling. This is in keeping with the philosophy (where possible) of basing
In the original FEH pooling scheme (Jakob et al., 1999), the distance between
subject site i and gauged site j is taken to be:
2
BFIHOST
ji
2
lnSAAR
ji
2
lnAREA
ji
ijσ
BFIHOSTBFIHOST
σ
lnSAARlnSAAR
σ
lnAREAlnAREA
2
1d
It is seen that two of the FEH catchment descriptors (AREA and SAAR) are transformed
logarithmically. The three coordinates in size-wetness-permeability space are then
standardised by dividing by the standard deviation σ of the relevant PCD. Finally,
differences in catchment size (indexed by ℓnAREA) are down-weighted in importance
relative to differences in catchment wetness (indexed by ℓnSAAR) and in catchment
permeability (indexed by BFIHOST), by inclusion of the ½ multiplier.
Volume II Flood Frequency Estimation
93
estimates on measured rather than modelled data. Gauged BFI has also been found useful in
estimating some hydrograph shape parameters (see Chapter 6 of Volume III).
In cases where there is no gauged BFI value, BFIsoil is to be substituted. The physical
catchment descriptor BFIsoil is detailed in Chapter 5 of Volume IV.
7.4 Growth curve estimation
Growth curves are expressed as algebraic equations for xT in terms of a frequency variable,
typically the return period T. Examples of such equations follow later in Section 7.4.3.
Suppose there are m gauging stations in the pooling group. Following the method of
L-moments, the pooled growth curve is constructed from pooled values of the L-moment
ratios. L-moment ratios were introduced in Chapter 4; see Box 4.2 and Section 4.5 in
particular. Typically, two L-moment ratios – the L-CV and the L-skewness – are used in
fitting the pooled growth curve. Sample values of L-CV and L-skewness are denoted here by
t2 and t3. [Editorial note: Some authors omit the subscript and denote L-CV by t.]
7.4.1 Pooled L-moment ratios
Let t2(i)
and t3(i)
denote sample values of L-CV and L-skewness at the ith
site. Pooled values
of the L-moment ratios are calculated as weighted averages:
2t =
m
1i
i
m
1i
(i)
2i
w
tw
7.3
and
3t =
m
1i
i
m
1i
(i)
3i
w
tw
7.4
where m is the number of stations in the pooling group and wi is a weighting factor.
[Editorial note: Where required, pooled values of L-kurtosis (t4) are calculated by the same
weighting scheme and annotated in similar style.]
Following guidance by Hosking and Wallis (1997), weights are taken proportional to record
length, i.e. wi = ni where ni is the number of annual maxima at station i. wi is then the
number of station-years of data in the pooling group. This weighting scheme is
recommended when the group (or “region”) of stations being pooled is reasonably
homogeneous.
7.4.2 Fitting growth curve distributions by the method of L-moments
In the so-called method of L-moments, a statistical distribution is fitted to a dataset by
equating its theoretical L-moment ratios to the sample values of the L-moment ratios.
In growth curve analysis, AM flood series from different stations are made comparable by
dividing the AM values by the index flood. The FSU adopts QMED as the index variable.
QMED corresponds to the 2-year flood, Q2. In consequence, the flood growth curve is
Volume II Flood Frequency Estimation
94
anchored to take the value x2 = 1.0. This determines one parameter of the distribution used to
model the growth curve.
If a 2-parameter distribution such as the EV1 or LO is being used, the remaining parameter is
determined by equating theoretical and sample values of the L-CV. If a 3-parameter
distribution such as the GEV or GLO is being used, the remaining two parameters are
determined by equating theoretical and sample values of both the L-CV and the L-skewness.
If a 4-parameter distribution such as the kappa is used, theoretical and sample values of the
L-kurtosis are also equated (see Box 4.2).
In pooled growth curve estimation, the values 2t and 3t (and, where required, 4t ) are
equated to expressions for these quantities written in terms of the distribution’s unknown
parameters. The resulting equations are solved for the unknown parameters.
A 3-parameter distribution can be used when pooling data from a group of stations because
the extra information ensures that the resulting standard error is smaller than in the single-site
case. Further, the 3-parameter distribution avoids possible bias resulting from using a
2-parameter distribution when in fact a 3-parameter distribution is more appropriate. The
thinking is summarised in Box 7.2, adapted from Cunnane (1989).
7.4.3 Growth curves for particular distributions
The formulae below define growth curves for the four distributions:
Generalised Extreme Value (GEV) distribution;
Extreme Value Type 1 (EV1) distribution, also known as the Gumbel;
Generalised Logistic (GLO) distribution;
Logistic (LO) distribution
The equations for estimation of their parameters by the method of L-moments are taken from
Hosking and Wallis (1997).
GEV growth curve
In this study, the GEV (and its special case EV1) and GLO (and its special case LO) are the
forms of growth curve chiefly used. The dimensionless GEV growth curve is defined by a
shape factor k and a scale parameter as follows:
k
k
T1T
Tlnln2
k
β1x for k 0 7.5
The parameters k and are estimated from the sample values t2 (of L-CV) and t3 (of
L-skewness):
k = 7.8590c + 2.9554c2 7.6
where
ln3
ln2
t3
2c
3
7.7
and
Volume II Flood Frequency Estimation
95
)2k)(1Γ(1(ln2)k)Γ(1t
ktβ
kk
2
2
7.8
Box 7.2: Qualitative outline of simulation results for flood quantile estimate
FIXED SKEWNESS MODELS (usually with two parameters)
(a) Model Skewness < Parent Skewness Hence negative bias i.e. underestimation at large T
(b) Model Skewness > Parent Skewness Hence positive bias i.e. overestimation at large T
VARIABLE SKEWNESS MODELS (usually with > two parameters)
(c) At-site use (d) Pooled XT estimation + At-site Qmed
Small bias, large se
Small bias and small se
1
0
-1
T
1
0
-1
T
1
0
-1
T
1
0
-1
T
Bias/QT%
[ Bias ± 1.96 se ]/QT %
T
10 100 1000
T
10 100
Fixed skewness models (usually with two parameters)
Variable skewness models (usually with more than two parameters)
FIXED SKEWNESS MODELS (usually with two parameters)
(a) Model Skewness < Parent Skewness Hence negative bias i.e. underestimation at large T
(b) Model Skewness > Parent Skewness Hence positive bias i.e. overestimation at large T
VARIABLE SKEWNESS MODELS (usually with > two parameters)
(c) At-site use (d) Pooled XT estimation + At-site Qmed
Small bias, large se
Small bias and small se
1
0
-1
T
1
0
-1
T
1
0
-1
T
1
0
-1
T
Bias/QT%
[ Bias ± 1.96 se ]/QT %
T
10 100 1000
T
10 100
– – – – – – – Bias/QT
- - - - - - - - - (Bias 1.96 se)/QT
(b) Model skewness > Parent skewness
[leads to +ve bias, with overestimation
at long return period, T]
(c) Single-site analysis
[Small bias and large standard error]
(d) Gauged QMED + pooled xT estimation
[Small bias and small standard error]
1
0
-1
1
0
-1
1
0
-1
1
0
-1
(a) Model skewness < Parent skewness
[leads to -ve bias, with underestimation
at long return period, T]
Cases (a) and (b) illustrate the effect of choosing a distribution having fixed but
incorrect skewness. Cases (c) and (d) illustrate the effect of a more flexible
distribution: namely, low bias and large standard error (SE) when used in single-site
analysis but with much reduced SE (as well as low bias) when used in gauged QMED +
pooled xT estimation.
[Adapted from Figure 5.1 of Cunnane (1989)]
Key
Volume II Flood Frequency Estimation
96
EV1 growth curve
The EV1 growth curve is defined as:
1/T))ln(1ln(ln(ln2)β1xT 7.9
The parameter is estimated from the sample value t2 (of L-CV):
ln(ln2)γtln2
tβ
2
2
7.10
where γ is Euler’s constant = 0.5772.
GLO growth curve
The GLO growth curve is defined by a shape factor k and a scale parameter as follows:
kk
T 1T1k
β1/FF11
k
β1x
for k 0 7.11
The parameters k and are estimated from the sample values t2 (of L-CV) and t3 (of
L-skewness):
k = – t3 7.12
and
kπsinttkπk
kπsintkβ
22
2
7.13
LO growth curve
The LO growth curve is defined as:
1)βln(T1xT 7.14
The parameter is estimated from the sample value t2 (of L-CV):
2tβ 7.15
[Editorial note: Implementation of methods through the FSU Web Portal also provides for
use of the 2-parameter lognormal (LN2) and 3-parameter lognormal (LN3) distributions. For
somewhat obscure reasons, the recommendation is that single-site application of the LN2
distribution is fitted by a scheme not involving L-moments. This variant is shown later in
Section 10.2.4. However, in all applications of the LN3 distribution, and in pooled
applications of the LN2 distribution, the use of L-moment methods is recommended. It
should be noted that different researchers favour different formulations of the lognormal
distribution. The L-moment methods implemented for the LN2 and LN3 distributions are
based directly on Hosking (1990) and on no other source.]
Volume II Flood Frequency Estimation
97
8 Effect of catchment type and period of record on xT and QT
8.1 Data screening
8.1.1 Discordancy
The 88 A1 + A2 grade stations for which BFI values were available were used to explore the
effect of catchment type and period of record on xT and QT.
Very unusual (i.e. discordant) datasets in the set of 88 catchments were first identified using
the Hosking and Wallis discordancy measure:
3
NDi uuAuu ii 1T
8.1
The vector holds the L-moment ratios for station i, namely the L-CV (t2), L-skewness (t3), ui
and L-kurtosis (t4). The superscript T denotes the transpose of the vector. u is the
unweighted regional average and is the matrix of sums of squares and cross-products. A
Further details are given in Hosking and Wallis (1997).
Table 8.1 lists the L-moment ratios and some leading PCDs for the 88 stations. The final
column shows the discordancy value Di when the station is judged relative to the set of 88
stations as a whole. Hosking and Wallis (1997) recommend that any site having Di > 3 be
regarded as discordant.
Table 8.1: L-moment ratios, PCDs and station discordancy within set of 88 A1+A2 stations
Station
#
#
annual
maxima
AREA SAAR BFI FARL URBEXT L-
CV
L-
skew
L-
kurt
Hosking+Wallis
discordancy, Di
06011 48 229.2 1029 0.71 0.87 0.011 0.11 0.09 0.07 0.44
06013 30 309.1 873 0.62 0.97 0.009 0.16 0.02 0.01 0.97
06014 30 270.4 927 0.63 0.93 0.012 0.15 0.22 0.12 0.57
06026 46 148.5 941 0.66 0.92 0.009 0.18 0.24 0.09 1.02
06031 18 46.2 931 0.56 1.00 0.015 0.26 0.39 0.32 1.98
06070 24 162.0 1046 0.73 0.83 0.013 0.14 0.14 0.14 0.02
07006 19 177.4 937 0.55 0.99 0.004 0.12 -0.20 0.07 2.46
07009 29 1658.2 869 0.71 0.99 0.008 0.21 0.21 0.11 0.74
07033 25 124.9 1032 0.44 0.89 0.006 0.13 0.16 0.27 0.72
08002 20 33.4 791 0.60 1.00 0.005 0.11 0.27 0.17 0.86
08005 18 9.2 711 0.52 1.00 0.250 0.38 0.23 0.19 4.71
09001 48 209.6 783 0.51 1.00 0.029 0.24 0.19 0.15 0.83
09002 24 35.0 755 0.67 1.00 0.210 0.42 0.39 0.25 6.26
09010 19 94.3 955 0.56 0.96 0.240 0.42 0.42 0.30 6.86
10021 24 32.5 799 0.65 1.00 0.242 0.21 0.19 0.07 1.10
10022 18 12.9 822 0.60 1.00 0.297 0.23 0.06 0.06 1.27
12001 50 1030.8 1167 0.72 1.00 0.006 0.19 0.18 0.17 0.19
14005 48 405.5 1015 0.50 1.00 0.024 0.15 0.29 0.22 0.54
Volume II Flood Frequency Estimation
98
Station
#
#
annual
maxima
AREA SAAR BFI FARL URBEXT L-
CV
L-
skew
L-
kurt
Hosking+Wallis
discordancy, Di
14006 51 1063.6 899 0.57 1.00 0.017 0.11 0.24 0.21 0.55
14007 25 118.6 814 0.64 1.00 0.007 0.17 0.22 0.07 1.13
14009 25 68.3 831 0.67 1.00 0.001 0.13 0.14 0.24 0.47
14011 26 162.3 807 0.60 1.00 0.028 0.14 0.02 0.16 0.42
14018 51 2419.4 857 0.67 1.00 0.017 0.14 0.04 0.06 0.39
14019 51 1697.3 861 0.62 1.00 0.018 0.14 0.09 0.13 0.05
14029 47 2778.2 877 0.69 1.00 0.016 0.08 0.06 0.07 0.73
15001 42 444.4 935 0.51 1.00 0.002 0.16 0.00 0.07 0.49
15003 50 299.2 934 0.38 1.00 0.006 0.11 -0.15 0.09 1.62
16001 33 135.1 916 0.61 1.00 0.003 0.12 0.01 0.13 0.29
16002 51 485.7 932 0.63 1.00 0.011 0.16 0.17 0.19 0.05
16003 51 243.2 1192 0.55 1.00 0.002 0.10 0.21 0.05 1.85
16004 48 228.7 941 0.58 1.00 0.011 0.11 0.07 0.11 0.19
16005 30 84.0 1154 0.56 1.00 0.003 0.10 0.20 0.13 0.73
16008 51 1090.3 1030 0.64 1.00 0.007 0.07 -0.06 0.06 0.94
16009 52 1582.7 1079 0.63 1.00 0.008 0.10 -0.10 0.05 1.01
16011 52 2143.7 1125 0.67 1.00 0.007 0.17 0.09 0.05 0.62
18004 46 310.3 985 0.68 1.00 0.003 0.09 0.04 0.33 2.89
18005 50 378.5 1190 0.71 1.00 0.004 0.14 0.22 0.18 0.21
19001 48 103.3 1176 0.64 1.00 0.019 0.10 0.10 0.12 0.28
19020 28 74.0 1179 0.66 1.00 0.000 0.20 0.03 0.04 0.99
23001 45 191.7 1084 0.32 1.00 0.003 0.18 0.13 0.18 0.18
23012 18 61.6 1264 0.46 1.00 0.024 0.15 0.39 0.33 1.65
24008 30 806.0 939 0.54 1.00 0.007 0.15 0.06 0.09 0.22
24022 20 41.2 942 0.53 1.00 0.003 0.23 0.12 0.16 0.81
24082 28 762.8 942 0.52 1.00 0.006 0.15 -0.04 0.13 0.80
25006 52 1162.8 932 0.71 0.96 0.019 0.14 0.14 0.18 0.04
25014 54 164.4 1008 0.67 1.00 0.005 0.13 0.10 0.13 0.07
25016 42 275.2 947 0.61 1.00 0.028 0.13 0.08 0.16 0.08
25023 52 113.9 922 0.65 1.00 0.003 0.16 0.14 0.06 0.68
25025 31 161.2 905 0.73 1.00 0.009 0.17 0.08 0.15 0.14
25027 43 118.9 1021 0.65 1.00 0.006 0.16 0.05 0.13 0.22
25029 33 292.7 1109 0.58 1.00 0.013 0.14 -0.02 -0.02 1.37
25030 48 280.0 1184 0.54 0.85 0.001 0.18 0.19 0.13 0.23
25034 24 10.8 969 0.76 1.00 0.000 0.17 -0.08 0.12 1.36
25040 20 28.0 990 0.64 1.00 0.062 0.15 0.20 0.18 0.10
25044 33 92.5 1187 0.58 1.00 0.000 0.19 0.25 0.15 0.48
25124 18 215.4 955 0.87 0.78 0.035 0.20 -0.07 0.23 2.95
26002 53 641.5 1067 0.61 0.98 0.003 0.11 0.22 0.29 0.97
Volume II Flood Frequency Estimation
99
Station
#
#
annual
maxima
AREA SAAR BFI FARL URBEXT L-
CV
L-
skew
L-
kurt
Hosking+Wallis
discordancy, Di
26005 51 1085.4 1054 0.56 0.98 0.002 0.10 0.04 0.14 0.28
26006 53 184.8 1121 0.54 0.97 0.006 0.15 0.40 0.41 2.83
26007 53 1207.2 1046 0.65 0.98 0.002 0.11 0.16 0.16 0.24
26008 49 280.3 1035 0.61 0.86 0.003 0.10 0.19 0.19 0.35
26018 49 119.5 1044 0.72 0.76 0.003 0.11 0.13 0.11 0.25
26019 51 253.0 980 0.54 0.99 0.013 0.14 0.23 0.12 0.71
26021 30 1098.8 945 0.83 0.81 0.004 0.14 -0.11 0.19 2.48
26022 33 61.9 916 0.58 1.00 0.006 0.17 0.09 0.05 0.69
26059 17 256.6 976 0.91 0.73 0.005 0.10 0.11 0.17 0.22
27001 30 46.7 1477 0.28 0.99 0.000 0.11 0.19 0.25 0.56
27002 51 564.3 1336 0.70 0.84 0.001 0.12 0.18 0.22 0.22
29004 32 121.4 1107 0.52 0.99 0.013 0.08 0.14 0.09 0.94
29011 22 354.1 1079 0.63 0.98 0.011 0.14 0.41 0.32 1.81
30007 31 469.9 1115 0.65 0.99 0.005 0.11 0.11 0.20 0.28
30061 33 3136.1 1422 0.78 0.66 0.007 0.15 0.41 0.38 2.38
31002 26 71.3 1530 0.53 0.63 0.000 0.13 0.31 0.18 1.03
32012 24 146.2 1784 0.59 0.84 0.000 0.07 0.00 0.18 0.90
34001 36 1974.8 1323 0.78 0.83 0.008 0.10 0.08 0.21 0.48
34003 29 1802.4 1340 0.80 0.82 0.008 0.09 0.11 0.25 0.85
34009 33 117.1 1257 0.40 1.00 0.011 0.10 0.08 0.16 0.22
34018 27 95.4 1555 0.66 0.73 0.055 0.11 0.18 0.03 1.73
34024 28 127.2 1177 0.52 0.92 0.007 0.07 -0.05 0.14 1.01
35001 29 299.4 1173 0.60 0.92 0.003 0.12 -0.01 0.19 0.92
35002 34 88.8 1381 0.42 0.99 0.000 0.10 0.03 0.08 0.41
35005 55 639.7 1198 0.61 0.90 0.002 0.14 0.20 0.14 0.27
35071 30 247.2 1364 0.77 1.00 0.001 0.10 0.12 0.20 0.28
36015 33 153.1 1091 0.42 0.96 0.000 0.16 0.33 0.30 0.99
36018 50 234.4 950 0.69 0.85 0.007 0.10 0.04 0.10 0.26
36019 47 1491.8 971 0.79 0.76 0.007 0.10 -0.03 0.06 0.58
36021 27 23.4 1570 0.27 1.00 0.000 0.12 0.20 0.22 0.30
36031 30 63.8 910 0.48 0.96 0.060 0.10 0.36 0.39 2.88
Three stations highlighted in red are seen to be discordant. These were excluded from further
analysis in this part of the study, even though it might be argued that this reduces the natural
variability within the dataset being studied. Before proceeding, it is instructive to consider
the characteristics that may explain why these stations are discordant.
8.1.2 Characteristics of the three discordant stations
The three discordant stations have several features in common. All three catchments lie close
to Dublin and are heavily urbanised (URBEXT > 0.20). They all have unusually high values
Volume II Flood Frequency Estimation
100
of L-CV. This is confirmed from the steepness of the EV1 Probability Plots 8.1, where it is
noted that summer events contribute widely to the AM flow series.
EV1 Probability Plots 8.1: Three discordant sites (Stations 08005, 09002 and 09010)
2 5 10 25 50 100 500
0
2
4
6
8
-2 -1 0 1 2 3 4 5 6 7
AM
flo
w (
m3
s-1
)
EV1 reduced variate, y
Station 08005 Sluice at Kinsaley Hall
winter peak
summer peakL-CV = 0.38
2 5 10 25 50 100 500 0
5
10
15
20
25
30
-2 -1 0 1 2 3 4 5 6 7
AM
flo
w (
m3 s
-1)
EV1 reduced variate, y
Station 09002 Griffeen at Lucan
winter peak
summer peakL-CV = 0.42
2 5 10 25 50 100 500 0
50
100
150
200
250
300
-2 -1 0 1 2 3 4 5 6 7
AM
flo
w (
m3 s
-1)
EV1 reduced variate, y
Station 09010 Dodder at Waldron's Bridge
winter peak
summer peakL-CV =0.42
Volume II Flood Frequency Estimation
101
One of the discordant stations is amongst those tested for trend and other non-randomness in
Chapter 3. Station 09002 Griffeen at Lucan showed nothing untoward in five out of the six
tests summarised in Table E.3. However, the parametric test based on samples drawn from
an EV1 distribution indicated a very highly significant trend. The sample CV for this station
is 0.83, making the particular test (which assumes a population CV of 0.30) inappropriate to
judging trend at this site.
No marked trend is apparent in Figure 8.1. The example confirms the value of graphical
display and of using a number of tests when exploring possible non-stationarity.
10
5
0
30
15
0
200520001995199019851980
300
200
100
0
Station 08005 Sluice at Kinsaley Hall
Station 09002 Griffeen at Lucan
Station 09010 Dodder at Waldron's Bridge
Figure 8.1: AM flow series for three discordant sites (Stations 08005, 09002 and 09010)
The discordant Stations 08005 and 09010 are amongst four catchments identified in
Section 2.5.2 for which the PCD model for QMEDrural seriously underestimates QMED. It
was noted there that urbanisation provides a plausible explanation for this.
It would therefore appear that flood characteristics of Station 08005 Sluice at Kinsaley Hall
and Station 09010 Dodder at Waldron’s Bridge are doubly severe:
QMED is unusually large (likely reflects urbanisation);
The growth curve is especially steep reflecting the high value of L-CV.
The high L-CV for Station 09010 is strongly influenced by the unusually large flood arising
from Hurricane Charlie in August 1986.
That greater variability (higher L-CV) is characteristic of some AM series in this part of
Ireland is consistent with the finding of Bruen et al. (2005) that flood growth curves around
Dublin tend to be steeper than prescribed for Ireland in NERC (1975). Table 8.2 provides a
summary of the three discordant stations.
AM
flo
w (
m3s-1
)
Hurricane
Charlie
Volume II Flood Frequency Estimation
102
Table 8.2: Stations showing large discordancy values (in pool of 88 A1 + A2 stations)
Station
number Station name
D
value L-CV CV Possible reason for discordancy
08005 Sluice at Kinsaley
Hall 6.16 0.38 0.68 Urbanised catchment
09002 Griffeen at Lucan 6.26 0.42 0.83 Urbanised catchment
09010 Dodder at
Waldron’s Bridge 8.94 0.42 0.86
Urbanised catchment; outlier arising from
Hurricane Charlie (August 1986)
8.1.3 Other heavily urbanised catchments
The dataset of 88 A1 + A2 stations includes two further heavily urbanised catchments:
Stations 10021 and 10022. These also lie in the area around Dublin. From EV1 Probability
Plots 8.2 it is seen that summer events (indicative of a faster flood response conditioned by
urbanisation) again feature in the AM flow series. However, the L-CV values and flood
growth rates are less exceptional than for the three discordant stations identified above.
EV1 Probability Plots 8.2: Two further heavily urbanised catchments
2 5 10 25 50 100 500
0
2
4
6
8
10
12
14
16
-2 -1 0 1 2 3 4 5 6 7
AM
flo
w (
m3 s
-1)
EV1 reduced variate, y
10021 River Shanganagh at Common's Road
winter peak
summer peak
L-CV = 0.21
2 5 10 25 50 100 500
0
1
2
3
4
5
6
7
8
-2 -1 0 1 2 3 4 5 6 7
AM
flo
w (
m3 s
-1)
EV1 reduced variate, y
10022 River Cabinteely at Carrickmines
winter peak
summer peakL-CV = 0.23
Volume II Flood Frequency Estimation
103
8.2 Effect of catchment type on pooled growth curve estimates
After exclusion of the three discordant stations, 85 catchments remain in the research dataset.
In various experiments, the 85 catchments have been divided into subgroups on the basis of:
Storage attenuation – as indexed by the FARL descriptor;
Peat content – as indexed by the PEAT descriptor;
Size – as measured by AREA;
Geographical location – within one of four regions defined in Map 8.1.
Rather than trying to compare entire growth curves simultaneously, assessments focused
chiefly on the 100-year growth factor x100, with some reference also to assessments of the
50-year growth factor x50.
The growth factor x100 has been estimated at each of the 85 gauging stations from its own
pooling group based on the size-wetness-permeability distance measure defined in
Equation 7.2 and using the 5T rule of Section 7.3.3 to determine the number of stations to be
pooled. Thus, each growth curve analysis pooled a minimum of 5 100 = 500 station-years.
Growth factor estimates were obtained by fitting GEV and GLO distributions to the pooled
data using the method of L-moments (see Section 7.4). Only the results for the GEV are
reported here. The effect of catchment type on pooled growth curve estimates did not differ
qualitatively between the two distributional assumptions.
The effect of catchment type on the 100-year flood growth factor are summarised in box-
plots. See Box 8.1 for an explanation of what these show.
Box 8.1: Box-plots
Box-plots 8.1a summarise the influence of the extent of peat cover (indexed by PEAT) on
values of x100, while Box-plots 8.1b summarise the influence from of the attenuating effect of
reservoirs and lakes within the catchment (indexed by FARL).
Box-plots provide a visual summary of data values
based on five numbers. The 50th
percentile is the
median. The box extends from the 25th
to 75th
percentiles. The part of the box tinted maroon
indicates the range of values between the 25th
and 50th
percentiles. The part tinted pale yellow indicates the
range from the 50th
to the 75th
percentile values.
[*Editorial note: The lines extending above the 75th
percentile and below the 25th
percentile are often
called whiskers. Conventions for drawing these vary.
The illustration is one interpretation which may not
be the one used in the box-plots below. It is possible
that the whiskers have been drawn down to the 9th
percentile and up to the 91st percentile.]
Var
iable
Median
Maximum*
Minimum*
Volume II Flood Frequency Estimation
104
Map 8.1: Location of the 85 stations within four geographical regions
In each case, the leftmost box-plot represents all 85 stations in the dataset. The median value
of x100 – represented by the separating line between the pale yellow and maroon boxes – is
seen to differ only slightly with the extent of peat cover, and without a definite pattern. The
median x100 values show a slight monotonically decreasing trend with increasing amount of
storage attenuation (i.e. for smaller FARL values). This is consistent with the storage action
of reservoirs and lakes typically attenuating flood growth rates.
[Editorial note: Box-plots 8.1 are cumulative. It is therefore puzzling that (e.g.) the x100
values for the 22 stations with PEAT > 0.15 do not span the range indicated for the x100
values for the 14 stations with PEAT > 0.20. As discussed in Box 8.1, it is possible that the
convention used to draw the whiskers was percentile-based. Even so, the diagrams appear
inconsistent. This is not a problem for Box-plots 8.2, which are non-cumulative except in
the LH column. Editing has sought to retain the most important findings, even when the
graphical support is not always clear.]
Shannon
West
South-West
East
Key___
Gauging station
Volume II Flood Frequency Estimation
105
Box-plots 8.1: Pooled GEV estimates of x100 showing effect of (a) PEAT and (b) FARL
The influences of geographical region and catchment size on values of x100 are summarised in
Box-plots 8.2a and b respectively. In terms of geographical region, the South-West has the
lowest median x100 (of just under 1.6) and the East has the highest value (of about 1.9).
Steeper flood growth curves in the East are commensurate with the finding of Bruen et al.
(2005). It should be noted that catchments of all types (e.g. size, degree of storage
attenuation and peat extent) are mixed together in these geographical groupings.
In terms of catchment size, the median x100 value shows a monotonic decrease with
increasing catchment area. In other words, there is some tendency for smaller catchments
(especially AREA < 100 km2) to exhibit steeper flood growth.
Box-plots 8.2: Pooled GEV estimates of x100 showing effect of (a) location and (b) AREA
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All st(85) ≥5%
peat(44)
≥10%
peat(30)
≥15%
peat(22)
≥20%
peat(14)
farl
≤.98(30)
farl
≤.95(23)
farl
≤.90(19)
farl
≤.85(15)
X1
00
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All st(85) West(17) South-
West(18)
East(24) Shannon
(26)
A≤100(19) A~101-
200(20)
A~201-
500(24)
A≥500(22)
X1
00
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All st(85) West(17) South-
West(18)
East(24) Shannon
(26)
A≤100(19) A~101-
200(20)
A~201-
500(24)
A≥500(22)
X1
00
24 stns
in
East
26 stns
in
Shannon
All
85
stns
19 stns
with
AREA
≤100
20 stns
with
AREA
101-200
24 stns
with
AREA
201-500
22 stns
with
AREA
>500
Flo
od
gro
wth
fac
tor,
x1
00
All
85
stns
17 stns
in
West
18 stns
in
SW
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All st(85) West(17) South-
West(18)
East(24) Shannon
(26)
A≤100(19) A~101-
200(20)
A~201-
500(24)
A≥500(22)
X1
00
All
85
stns
44 stns
with
PEAT
≥0.05
30 stns
with
PEAT
≥0.10
22 stns
with
PEAT
≥0.15
14 stns
with
PEAT
≥0.20
All
85
stns
30 stns
with
FARL
≤0.98
23 stns
with
FARL
≤0.95
19 stns
with
FARL
≤0.90
15 stns
with
FARL
≤0.85
Flo
od
gro
wth
fac
tor,
x100
Volume II Flood Frequency Estimation
106
8.3 Temporal effect on pooled growth curve estimates
8.3.1 Periods of record considered
Pooled growth curves have been derived from different periods of record in four formats:
Distinguishing early and late halves of each record (for 90 stations);
Distinguishing three decades: 1972-1981, 1982-1991 and 1992-2001 (for 50 stations);
Distinguishing five decades: 1952-1961, 1962-1971, 1972-1981, 1982-1991 and
1992-2001 (for 26 stations);
Distinguishing five offset decades 1957-66, 1967-76, 1977-86, 1987-96 and 1997-
2006 (34 stations for all five decades and 68 stations for the last three decades).
The periods detailed above refer to water-years, e.g. the decade 1972-1981 corresponds to the
period 1 Oct 1972 to 30 Sep 1982.
8.3.2 Summary results for 90 stations
Some 90 gauging stations were used in this part of the study, which was completed before the
data provider indicated that some stations were unsuitable for detailed analysis. It is believed
that the general nature of the findings would not be greatly altered had the study been
repeated with the five rejected stations excluded.
The effects of period of record (1st half of each record compared with 2
nd half) and of
successive decades are shown in Box-plots 8.3. Median x100 shows a decrease between the
earlier and later halves of records, while the decadal pattern of values of x50 is up and down.
Box-plots 8.3: Period-of-record and decadal effects on pooled flood growth by GEV
8.3.3 Individual results for three decades (50 stations)
There are 50 stations for which estimates are available for all three decades. The decadal
values of x50 for each station are displayed in Figure 8.2a. It can be seen that the largest x50
values occur in the middle decade (1982-91) in every case.
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All record(90) 1st half
record(90)
2nd half
record(90)
Decade70(52) Decade80(83) Decade90(82)
X100 (
X50 in c
ase o
f decade a
naly
sis
)
90
records
in full
1st half of
90 records 2
nd half of
90 records
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All record(90) 1st half
record(90)
2nd half
record(90)
Decade70(52) Decade80(83) Decade90(82)
X100 (
X50 in c
ase o
f decade a
naly
sis
)
Pooling Growth Curve(GEV)
1
1.2
1.4
1.6
1.8
2
2.2
2.4
All record(90) 1st half
record(90)
2nd half
record(90)
Decade70(52) Decade80(83) Decade90(82)
X100 (
X50 in c
ase o
f decade a
naly
sis
)
Flo
od
gro
wth
fac
tor,
x50
Flo
od
gro
wth
fac
tor,
x1
00
1972-
1981
52 stns
1982-
1991
83 stns
1992-
2001
82 stns
90
records
in full
[Values are pooled estimates at individual sites based on
the size-wetness-permeability pooling system of Equation 7.2 and the 5T rule]
Volume II Flood Frequency Estimation
107
The corresponding QMED values, standardised at each station by the 1972-81 value of
QMED, are shown in Figure 8.2b. In most cases, the QMED values for the 2nd
and 3rd
decade
exceed the 1st decade value. Although the 3
rd decade displays a number of remarkably high
QMED values, only the expected one third of stations have their largest QMED in that
decade. Figure 8.2c shows the estimated Q50 values, standardised at each station by the
1972-81 value. The majority of Q50 values in the 2nd
and 3rd
decades exceed those of the
1st decade, with more than half of these exceedances occurring during the 2
nd decade.
Figure 8.2: Decadal estimates for 50 stations (arranged in station-number order)
1972-81 1982-91 1992-2002 (a) Pooled growth factor, x50
(b) QMED (standardised to 1st decade)
(c) Q50 (standardised to 1st decade)
1972-81 1982-91 1992-2002
1972-81 1982-91 1992-2002
Volume II Flood Frequency Estimation
108
8.3.4 Individual results for five decades (26 stations)
There are 26 stations for which decadal estimates are possible for five decades. At some
stations, there are not a full ten years of data available, especially for the 1st decade. Only
stations with seven or more years of data available for the 1st decade are included. The
analyses were executed as for the 3-decade case above, with results presented in Figure 8.3.
Figure 8.3: Decadal estimates for 26 stations (arranged in station-number order)
(a) Pooled growth factor, x50
(b) QMED (standardised to 1st decade)
(c) Q50 (standardised to 1st decade)
1952-61 1962-71 1972-81 1982-91 1992-02
1952-61 1962-71 1972-81 1982-91 1992-02
1952-61 1962-71 1972-81 1982-91 1992-02
Volume II Flood Frequency Estimation
109
The 2nd
(1962-71) decadal value of x50 is larger than those of other decades at every station,
with an average value of approximately 1.7. This contrasts strongly with the results for the
5th
decade (1992-2001) where the average value is approximately 1.4. Figure 8.3b shows the
decadal standardised QMED values for which the largest values tend to be from the 2nd
(1962-71) and 5th
(1992-2001) decades. The Q50 values in Figure 8.3c also convey the
impression that the largest Q50 values occur for the 2nd
(1962-71) and 5th
(1992-2001)
decades.
8.3.5 Individual results for a different set of five decades
For 34 stations, it was possible to examine a different set of five decades, offset by five years.
Results for decades ending in 1966, 1976, 1986, 1996 and 2006 were found to be broadly
similar to those described in Section 8.3.4.
8.3.6 Outcome
Natural variability in climate can lead to flood series exhibiting so-called flood-rich and/or
flood-poor periods. The period over which flood data are available has been confirmed to
have an effect on flood frequency estimation. For example, there is some limited evidence –
by combination of Figure 8.3c and the equivalent result from offset decades (Section 8.3.5) –
that the period 1962-1976 was flood-rich.
This finding emphasises the importance of analysing the longest flood records available and,
of course, incorporating the most recent flood data. The incorporation of historical flood
information can also be important (see Box 8.2).
Box 8.2: Incorporation of historical flood data
8.4 Arterial drainage effect on pooled growth curve estimates
Pooled growth curves have been compared for pre and post-drainage records of 16 gauging
stations, eight of which are in the Boyne river basin (Hydrometric Area 07).
The recommended pooling procedure of Section 7.3.4 requires BFI, the baseflow index (see
Section 5.1 of Volume IV). For historical or hydrometric reason, some gauging stations are
unable to measure flow throughout the full range of low, medium, high and flood flows. At
the time of study, BFI values were not available for three of the 16 stations. Values of BFI
from nearby gauged stations were therefore substituted. [Editorial note: The recommended
procedure in such cases is to use the physical catchment descriptor BFIsoil in place of BFI.
However, the study reported was undertaken before completion of the BFIsoil model. It
Editorial note: Historical flood data was not central to the FSU project, in large measure
because of the extensive work done under other initiatives, most notably the National
Flood Hazard Mapping website (http://www.floodmaps.ie/). Cawley et al. (2005) discuss
historical floods in the Brosna, Corrib, Deel, Dodder, Griffeen and Lee catchments.
Guidance on procedures for incorporating historical flood information into the statistical
analysis of peak flows is nevertheless required. Bayliss and Reed (2001) advocate the use
of graphical methods. More theoretical methods are available in the literature, including:
Guo (1990), Guo and Cunnane (1991), Cohn et al. (1997) and Payrastre et al. (2011).
Volume II Flood Frequency Estimation
110
transpires that the BFI substitutions made for Stations 07011 and 30004 were somewhat less
appropriate than that made for Station 07007. BFI is itself influenced by arterial drainage.]
The results of the pooled estimation of x50, QMED and Q50 for pre-drainage and post-
drainage records at the 16 stations are summarised in Figure 8.4.
Figure 8.4: Comparison of pre and post-drainage flood statistics at 16 stations
The post-drainage growth factor (Figure 8.4a) is seen to be smaller than the pre-drainage
growth factor in every case. This reflects that the index flood QMED is itself greater after
Pre-drainage Post-drainage
Pre-drainage Post-drainage
(a) Pooled growth factor, x50
(b) QMED (standardised to value for pre-drainage series)
(c) Q50 (standardised to value for pre-drainage series)
Pre-drainage Post-drainage
Volume II Flood Frequency Estimation
111
drainage (see Figure 8.4b and also Section 2.4). The estimated Q50 values for the post-
drainage period exceed the pre-drainage values in ten of the 16 cases, although Station 24001
shows only a minor increase.
It is seen in Figure 8.4b and Figure 8.4c that one station (Station 30001 Aille at Cartonbower)
shows a contrary result, with QMED and Q50 appreciably reduced after drainage. This station
may warrant special investigation.
Wider effects of drainage on river flows have been extensively studied by (amongst others)
Lynn (1981), Robinson (1990) and Bhattarai and O’Connor (2004).
8.5 Implications for flood frequency estimation
8.5.1 Implications for pooling-group formation
Storage attenuation (indexed by FARL), catchment size and geographical location appear to
have a noticeable effect on the x100 growth factor, whereas the extent of peat cover does not.
It is therefore reassuring that differences in AREA appear in the recommended distance
metric (Equation 7.2) used to form pooling groups by the region-of-influence method (see
Section 7.3.2). The non-appearance of FARL and geographical location in the recommended
distance metric is less comforting.
Geographical location may be partially taken into account by SAAR but not uniquely so. It is
likely that users will want to take geographical location explicitly into account in some cases,
especially for locations in the East, not least around Dublin.
The lack of representation of FARL in the distance metric is a reminder of the weakness of
purely rule-based methods of flood frequency estimation. The practical recommendation is
that users should take particular heed of FARL values when diagnosing pooling groups that
appear heterogeneous.
Users need to be alert to the trade-off between tailoring the pooling-group constituency to the
particular characteristics of the subject catchment and retaining a sufficient number of
station-years in the pooling group. Adherence to the 5T rule of Section 7.3.3 is generally
recommended.
8.5.2 Respecting recent flood data
With regard to temporal effects, there is no clear evidence that respect for recent flood data
should be prioritised over respect for older flood data. This does not sit well with
engineering guidance that – with on-going global and land-use change – “The past is no
longer the key to the future, and the future is uncertain” (Irish Academy of Engineering,
2007). However, there is a one-sided nature to extreme events. Once they have been
observed, they are in the record forever. Until they have been observed, they are entirely
missing.
A theme from Ireland at Risk (Irish Academy of Engineering, 2007) is that we should not
wait for extreme events to be gauged before taking action. Although in part paradoxical, one
reaction could be to reinforce the practice of incorporating historical flood information (see
Section 1.3 and Box 8.2) into flood frequency estimation. “The past may not be the key to
the future. But neither is neglect of the past.”
Volume II Flood Frequency Estimation
112
9 Uncertainty estimation
9.1 Standard errors – an introduction
The standard error (SE) of an estimate of QT is an indication of how reliable that estimate is.
It is based on the assumption that the data upon which the estimate is based are randomly
drawn from a single population – an assumption that cannot easily be proven. If an infinite
number of similarly sized datasets were to be drawn from the same population and the value
of QT obtained from each set by the same procedure then the SE is defined as:
SE(QT) = Standard deviation of all the possible set of QT values.
This measure represents only the degree of scatter of the several estimates and does not refer
to whether the mean of these equals the true value in the population. If this equality holds,
the procedure whereby QT is calculated is said to be unbiased; otherwise it is said to be
biased.
In flood hydrology, randomness of annual maximum floods and lack of trend with time are
generally assumed. Likewise it is assumed that a single form of statistical distribution
describes all the AM flood series in a region or country. Such assumptions cannot be proven
and there is some evidence from Chapter 3 and Appendix E to suggest that Irish flood data
are not entirely trend free. The presence of low or high outliers (see Section 4.6) in some
datasets also makes interpretation difficult. If in reality the data are not drawn from a unique
homogeneous parent population then the SE cannot strictly be defined.
However, the SE concept is widely used in flood hydrology. Its value is determined on the
assumption of a unique homogeneous parent and the value of SE so obtained may be
considered as a lower bound on the true value, were it derivable. Hence the SE provides a
useful guide to the precision of the QT value obtained in any situation.
The aim in this chapter is to provide expressions or graphs which give an indication of the
order of magnitude of SE(QT) in single-site and pooled estimations of QT. The discussion is
restricted to the EV1 and GEV cases on the assumption that the LO and GLO (and LN2 and
LN3) cases would yield standard errors of the same order of magnitude.
9.2 Standard error of QMED estimation from gauged flood data
Let there be N annual maximum values at a gauged site. QMED is obtained as the median of
the annual maximum flow series:
QMED ≡ Median (Q1, Q2, … QN)
The Qi values are re-ordered from smallest to largest so that Q(1) < Q(2) < … < Q(N). If N is
odd, the single mid-value is taken as QMED. If N is even, QMED is taken as the average of
the two middle-most values.
The standard error (SE) of the mean in a random sample from a Normal distribution is:
Nσmean
XSE 9.1
Volume II Flood Frequency Estimation
113
The SE of the median in such a sample is about 25% greater:
Nσ1.253med
XSE
9.2
That the standard error of the median is 25% larger than that of the mean is less of a
drawback when applied to AM flood data, where it is helpful that QMED is unaffected by
low/high outliers in the data sample and has a return period precisely anchored at 2 years.
Since Irish flood data are slightly more skewed than the Normal distribution, SE(QMED) for
flood data will be slightly greater than by Equation 9.2. Simulation indicates that the
multiplier is increased from 1.253 to about 1.28 to 1.30, when N = 5.
Adopting the larger value of 1.30, Equation 9.2 can be written:
Nmean
QCV1.30med
QSE
9.3
and:
Nmed
Q
medQ
meanQ
CV1.30med
QSE
9.4
Taking average values of CV and of the ratio Qmed/Qmean for 110 Irish A1 + A2 stations into
account – from Table 4.2 these are seen to be 0.273 and 0.963 respectively – an approximate
value of the multiplier of QMED is reached as:
Nmed
Q963.0
1273.01.30
medQSE
9.5
Thus:
NQMED369.0QMEDSE 9.6
Here, Qmed has been written in the preferred FSU notation of QMED.
9.3 A comparison with FEH methods
The standard errors above are seen in Table 9.1 to be somewhat smaller than those quoted for
UK flood data in the FEH (Robson, 1999a) for records shorter than about 20 years. This
difference likely reflects the smaller CV and lower skewness of Irish data. But see Box 9.1.
Table 9.1: Typical standard errors when estimating QMED from annual maxima
Number of annual
maxima, N
Approx. standard error,
by Equation 9.6
Equivalent value inferred from AM
column of Table 12.3 of Robson (1999a)
5 0.165 QMED 0.218 QMED
10 0.117 QMED 0.138 QMED
15 0.095 QMED 0.109 QMED
20 0.083 QMED 0.079 QMED
50 0.052 QMED
Volume II Flood Frequency Estimation
114
Box 9.1: Comparison with uncertainty of FEH methods
One method of reducing the standard error of estimate is by “record extension” whereby a
short record at the subject site is extended by exploiting the correlation between the AM
series at the subject site and the AM series of one or more neighbouring sites (Fiering, 1963).
A variation on this approach to QMED refinement is used in Section 2.1.2.
9.4 Standard error of QT in single-site estimation
9.4.1 Method
Theoretical expressions for L-moment based estimates of QT in both EV1 and GEV cases
have been given by Lu and Stedinger (1992) as follows. For samples drawn from the EV1
distribution:
2
T 0.8046y0.4574y1.1128n
αQSE 9.7
where TQ is the estimate for the T-year flow event; α is the EV1 scale parameter;
1/T1lnlnyy T 9.8
is the EV1 (or Gumbel) reduced variate; and n is the number of observations in the sample.
For samples drawn from the GEV distribution:
1/23
3
2
210T kTakTakexpTaTaexpn
αQSE 9.9
where TQ is the estimate for the T-year flow event; a0(T), a1(T), a2(T) and a3(T) are
coefficients that depend on the return period T; α and k are scale and shape parameters
respectively; and n is the number of observations in the sample. The values for the
coefficients for different return periods are tabulated in Lu and Stedinger (1992). For
example:
For T=10, a0= -2.667, a1= 4.491, a2= -2.207, a3= 1.802;
For T=100, a0= -4.147, a1= 8.216, a2= -2.033, a3= 4.780.
Editorial notes: The comparison with the FEH is not straightforward because of the
different approaches taken to assessing the uncertainty of QMED. The values in the RH
column of Table 9.1 may not provide a fair reflection. The cited FEH table presents
estimates of the factorial standard error (FSE). To subtract 1.0 from these numbers and
present them as equivalent values of the relative SE appears inflationary. Use of
resampling methods – as in Robson (1999a) – would allow an authoritative comparison.
Robson found that, when N < 14, QMED can be estimated with slightly smaller standard
error by using the peak-over-threshold (POT) flood series. This gain may not transfer to
Ireland where – for historical and hydrometric reasons – AM flow data are curated to a
higher standard than POT data. Nevertheless, there may be merit in considering the POT
approach for short flood series (say N < 10) of high quality.
Volume II Flood Frequency Estimation
115
9.4.1 Relative standard error
The quantity chiefly discussed below is the relative standard error: SE(QT)/QT. In the EV1
case, this can be expressed as a function of the ratio /u and yT or equivalently as a function
of CV and yT or of L-CV and yT. Note that L-CV ≈ 0.5 CV in the EV1 case.
Equation 9.7 has been applied to the data of 85 A1 + A2 stations in which the values of u,
and QT were estimated from each site's data separately and the ratio SE(QT)/QT formed. The
true value of SE(QT) is not obtained by using estimated values of parameters in these
expressions. However, it is considered useful to examine the range of values so obtained
because part of their scatter is caused by inter-site heterogeneity, assuming such exists, as
well as by random sampling effects and unequal record lengths.
9.4.2 Relative standard errors under the EV1 assumption
Relative standard errors across the 85 sites are summarised in Box-plots 9.1 for six different
return periods. It is suggested that the interquartile range – i.e. the range of the middle 50%
of values – ought to give a good indication of the range within which the true SE values fall.
The extreme high and low values can be considered unrepresentative of the whole, and the
suggestion is therefore to focus on the box (rather than whisker) part of the box-plot (see Box
8.1 for an introduction to box-plots).
The results are presented a second time in Box-plots 9.2, along with theoretical EV1 relative
standard errors calculated from Equation 9.7, using the average sample size of 37 years and
for a range of L-CV values that span the values experienced among Irish AM flood data.
[Editorial note: The different style of the boxes in Box-plots 9.2 arises from a technical
difficulty in graphing.]
What these show – under the EV1 assumption – is that the relative standard error can be
taken to be between 5% and 7% for the 10-year quantile and between 6.5% and 10% for the
100-year quantile.
Box-plots 9.1: Relative SE of single-site quantile estimates – EV1 assumed
SE
(QT)/
QT a
s %
Based on 85 A1 + A2 stations
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Relative Standard Error (At site/EV1)
0
2
4
6
8
10
12
14
16
18
T=5 T=10 T=25 T=50 T=100 T=500
se(Q
^T)/
QT %
Volume II Flood Frequency Estimation
116
Box-plots 9.2: As above but with lines superposed to show theoretical values of relative SE
(for N = 37 and a range of L-CV)
9.4.3 Relative standard errors under the GEV assumption
Corresponding results for the GEV case are presented in Box-plots 9.3 and Box-plots 9.4.
The theoretical values shown in Box-plots 9.4 are for a single value of the shape parameter
k = -0.1, which it is felt ought to cover the most extreme underlying population case arising
in Irish conditions. From the box-plots we can conclude, using the arguments as above, that
under the GEV assumption, the relative standard error can be assumed to be between 6% and
8% for the 10-year quantile and between 8% and 15% for the 100-year quantile. Although it
can be seen that the most extreme values are extremely large, these are not representative of
the true relative standard error.
Box-plots 9.3: Relative SE of single-site quantile estimates – GEV assumed
SE
(QT)/
QT a
s %
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Relative Standard Error (Atsite/EV1)
LCV=0.15
LCV=0.3
LCV=0.25
LCV=0.2
LCV=0.1
0
2
4
6
8
10
12
14
16
18
T=5 T=10 T=25 T=50 T=100 T=500
se(Q
T)/
QT%
SE
(QT)/
QT a
s %
Based on 85 A1 + A2 stations
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Relative Standard Error (At site/GEV)
0
10
20
30
40
50
60
70
80
90
100
T=5 T=10 T=25 T=50 T=100 T=500
se(Q
^T)/
QT %
Volume II Flood Frequency Estimation
117
Box-plots 9.4: As above but with lines superposed to show theoretical values of relative SE
(for k = -0.1, N = 37 and a range of L-CV)
9.5 Standard error of pooled estimate of xT and of QT
9.5.1 Simulation method
An estimate of the order of magnitude of the pooled estimate of xT, SE(xT), for selected
values of T, was obtained by simulation for each of the 85 stations. Each gauging station was
selected in turn as the subject site and the following procedure applied:
Step 1 Identify the gauging stations in the subject site’s pooling group using the
recommended distance metric (i.e. dij values given by Equation 7.2) to achieve a
minimum of 500 station-years of data in the pooling group (so as to comply with the
5T rule for the 100-year quantile).
Step 2 Random samples are drawn from EV1 populations for the subject site and for each
site in the pooling group. For each site, the sample size is taken as equal to the length
of the observed record at the site and the parameter values u and are those estimated
from the observed record by L-Moments.
Step 3 The sample QMED is obtained for the subject site.
Step 4 The L-CV value is obtained for each sample in the pooling group and the weighted
average of these is calculated (using Equation 7.3) to yield the pooled L-CV.
Step 5 The pooled L-CV is used to determine the pooling group’s EV1 growth curve
parameter β (using Equation 7.10)
Step 6 The subject site’s ))/11ln(ln()2ln(ln1 TxT is calculated for T = 5, 10,
25, …, 500 years
Step 7 The subject site’s QT = QMED × xT is calculated for T = 5, 10, 25, …, 500 years.
Step 8 Steps 2 to 7 are repeated 1000 times to provide 1000 values of xT and QT at the
subject site and the standard errors SE(xT) and SE(QT) are calculated for the subject
site by the following equations:
SE
(QT)/
QT a
s %
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Volume II Flood Frequency Estimation
118
N
1i
M
1m
2
T
i
T
im
T
i,
TX
xx
M
1
N
1)SE(x 9.10
N
1i
M
1m
2
T
i
T
im
T
i,
TQ
M
1
N
1)SE(Q 9.11
where m
T
i,x and m
T
i,Q are the estimated T-year growth factor and T-year flood quantile
respectively at site i at the mth
repetition; T
ix and T
iQ are the mean of these estimated
measures; T
ix and T
iQ are the assumed true T-year growth factor and T-year quantile
at site i; N is the number of sites in the pooling group and M is the number of
repetitions. These expressions average the standard error over all the sites in the
pooling group to take heterogeneity in the pooling group into account. It should be
noted that the actual standard error for any individual site could be smaller or larger
than the calculated value depending on the values of L-CV and L-skewness at the
particular site; the extreme values in the box-plots below illustrate the range that
could occur.
9.5.2 Results based on EV1 simulations
Step 1 to Step 8 were repeated for each of the 85 stations. The resultant relative standard
errors for xT and QT are presented in Box-plots 9.5 Box-plots 9.6 respectively. The same
remarks apply to the interpretation of these plots as to those for the single-site quantile
estimates presented in Section 9.4.
The relative standard errors for xT range from 1.1% to 1.2% at T = 10 to 1.8% to 2.0% at
T = 100 years. The corresponding relative standard errors for QT differ little with return
period, ranging from 3.7% to 8.5% at T = 10 to 4.0% to 9.0% at T = 100 years.
Box-plots 9.5: Relative standard error in xT – pooled EV1 simulations
Relative Standard Error of Growth Curve (Pooled/EV1-Simulation)
0.0
0.5
1.0
1.5
2.0
2.5
3.0
T=5 T=10 T=25 T=50 T=100 T=500
se(X
^T)/
XT %
SE
(xT)/
xT a
s %
Based on 85 A1 + A2 stations
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Volume II Flood Frequency Estimation
119
Box-plots 9.6: Relative standard error in QT – pooled EV1 simulations
9.5.3 Results based on GEV simulations with k = -0.1
The simulation procedure described above was also applied with the GEV distribution used
instead of the EV1 in Step 2 to Step 6. Because the shape parameter k can be very variable –
and unreasonably large or small when estimated from single-site records – a single value of
k = -0.1 was adopted in all simulations. This is considered a conservative choice in the
context of Irish flood data.
In Step 4, L-skewness and average L-skewness are calculated as well as L-CV. In Step 5, the
GEV growth curve parameters and k are calculated by Equations 7.8 and 7.6. In Step 6, the
expression for xT given in Equation 7.5 is used.
The results are presented in Box-plots 9.7 and Box-plots 9.8. The standard error for xT
ranges from 1.4% to 1.8% at T = 10 to 3.7% to 5.0% at T = 100, while the corresponding
standard errors for QT show an increasing trend, from 3.8% to 9.0% at T = 10 to 4.6% to
10.6% at T = 100.
Box-plots 9.7: Relative standard error in xT – pooled GEV simulations for k = -0.1
Relative Standard Error of Quantile Estimate (Pooled/EV1-Simulation)
0
5
10
15
20
25
30
35
T=5 T=10 T=25 T=50 T=100 T=500
se(Q
^T)/
QT %
SE
(xT)/
xT a
s %
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Based on 85 A1 + A2 stations S
E(x
T)/
xT a
s %
Based on 85 A1 + A2 stations
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Volume II Flood Frequency Estimation
120
Box-plots 9.8: Relative standard error in QT – pooled GEV simulations for k = -0.1
9.5.4 Summary
The ranges for six return periods and both EV1 and GEV distributions are summarised in
Table 9.2. Since the 100-year flood is used for design purposes for many major projects in
Ireland, it is worth noting that its standard error can be considered to be approximately 10%
of Q100.
Table 9.2: Relative standard errors for growth factors xT and quantile estimates QT
Pooled EV1 case Pooled GEV case (k = -0.1)
Return
period, T
Relative SE (%)
in xT
Relative SE (%)
in QT
Relative SE (%)
in xT
Relative SE (%)
in QT
5 0.7 to 0.8 3.7 to 8.4 0.8 to 1.0 3.7 to 9.0
10 1.1 to 1.2 3.7 to 8.5 1.4 to 1.8 3.8 to 9.0
25 1.4 to 1.6 3.8 to 8.7 2.3 to 3.0 4.1 to 8.9
50 1.6 to 1.8 3.9 to 8.9 3.0 to 3.9 4.3 to 9.6
100 1.8 to 2.0 4.0 to 9.0 3.7 to 5.0 4.6 to 10.6
500 2.1 to 2.3 4.1 to 9.3 5.4 to 7.3 5.7 to 12.2
[Editorial note: These standard errors seem surprisingly small. Perhaps the assumptions
made are unduly restrictive. The scope for underestimation of flood quantiles is intrinsically
much greater than the scope for overestimation. This is evident in the lop-sided whiskers in
the standard errors of the quantile estimates shown in Box-plots 9.6Box-plots 9.8.]
9.6 Standard error of QT based on PCD estimate of QMED and pooled xT
Using the index flood method:
TmedT xQQ 9.12
SE
(QT)/
QT a
s %
Based on 85 A1 + A2 stations
T = 5 T = 10 T = 25 T = 50 T = 100 T = 500
Volume II Flood Frequency Estimation
121
The variance of the T-year event is approximated using Taylor series expansion as:
)xQVar()QVar( TmedT
9.13
)x,QCov()xE()QE(2)xVar()QE()QVar()xE( TmedTmedT
2
medmed
2
T
Ignoring effects arising from:
Bias in the sample median
Bias in the sample L-moment ratio
Covariance between the sample median and the estimated regional growth curve
leads to:
)xVar(|)(Q)QVar(|)(x)QVar( Tm
2
medmedm
2
TT 9.14
The expression for Var(QT) is dominated by the Var(QMED) term as Var(xT) affects only the
3rd
or 4th
decimal point in the value of the expression. Hence the expression effectively
reduces to:
)SE(Qx)SE(Q medTT
)Q(x)SE(QxQ)SE(Q medTmedTTT
medmedTT Q)SE(QQ)SE(Q
Applying the approximate SE in QMED from Equation 9.6 yields:
med
medTT
Q
Q
N
0.369Q)SE(Q
i.e. N
0.37Q)SE(Q TT 9.15
Therefore, SE(QT)/QT is independent of T but dependent on the value of N associated with
the worth of a PCD-based estimate of QMED.
Experience suggests that the worth of a PCD-based estimate of the index flood is typically
equivalent to about one year of AM data, e.g. Nash and Shaw (1965), p I.342 of NERC
(1975) and Hebson and Cunnane (1987). [Editorial note: Experience in the FEH was much
the same, with the FSE of 1.58 quoted for the recommended PCD model of QMEDrural
slightly inferior to the FSE of 1.52 quoted for estimation of QMED from one year of AM
flow data.] Inserting N=1 into Approximation 9.15, the relative standard error of QT is
estimated to be 0.37. This value is consistent with FSE = 1.37 for the PCD-based estimate of
QMED on rural catchments (see Equation 2.8 in Section 2.2.5). [Editorial note: As
discussed in Box 9.1, comparison of the relative standard error of X with the FSE of X minus
1.0 does not appear to be even-handed. The Equation 2.8 PCD-based model is especially
good of its kind. It is likely worth a little more than one year of AM flow data.]
The finding – in derivation of Approximation 9.15 – that the uncertainty in QT is dominated
by the uncertainty in QMED and is independent of return period may seem counterintuitive.
The result applies regardless of whether the growth curve is obtained using EV1 or GEV.
Volume II Flood Frequency Estimation
122
[Editorial note: This finding does not sit comfortably with the experience of many applied
hydrologists that growth-curve estimation is a skilled task that plays a crucial role in
estimating the T-year flood, QT. It is particularly helpful that Section 10.4 includes a number
of worked examples demonstrating the subtlety of flood estimation problems and their
resolution.]
[Editorial note: While it is possible to judge the relative uncertainty of different methods of
single-site analysis and the relative uncertainty of different methods of pooled analysis, the
inter-comparison of pooled and single-site methods is exceptionally challenging. A practical
reason why pooled methods are often to be preferred is the desire to promote consistency in
flood estimates obtained for different sites along a river and on different catchments that are
hydrologically similar.]
Volume II Flood Frequency Estimation
123
10 Guidelines for determining QT
10.1 Introduction
10.1.1 Single-site or pooled analysis?
QT is the T-year return period flood, i.e. the flood peak with an annual exceedance probability
of 1/T. If sufficient data exist and the required return period is short, QT may be determined
by single-site analysis. Although adequate AM flow data may exist, interpretation can still be
problematic if the probability plot is of unusual shape. In such a case, it may be necessary to
combine single-site and pooled approaches.
For estimating floods of long return period, a combined single-site and pooled approach is
recommended. Where the subject site is ungauged, QMED must be estimated from the
catchment descriptor model of Section 2.2.5 and enhanced where possible by transferring
information from a nearby or similar gauged site (see Sections 2.6 and 2.7, and Chapter 11).
10.1.2 Probability plots
It is essential to examine a probability plot when analysing AM flow data. When choosing
the statistical distribution or estimation procedure to adopt, practitioners are often greatly
influenced by the appearance of single-site data on probability plots. However, it must be
borne in mind that random samples from a particular statistical population show considerable
inter-sample variation when displayed on probability plots. Some samples do not display
convincing straight-line behaviour even when the parent distribution is known to be a straight
line on such a plot.
Plots in Section 5.1.2 show nine random samples from an EV1 population with CV=0.33, a
value slightly higher than the average CV for Irish Grade A1 station data. Sample sizes are
25 in EV1 Probability Plot 5.1 and 50 in EV1 Probability Plot 5.2. It can be seen that
departures from straight-line behaviour are more striking in the smaller samples than in the
larger samples. Some of the samples of size 25 would have been given a low linearity score
in the assessment of probability plots described in Chapter 5. However, relatively few of the
samples of size 50 depart too markedly from straight-line behaviour.
10.1.3 Factors to be borne in mind
It can be convenient to refer to the estimate of QMED made from a site’s own flood data as
the at-site QMED. Factors to be borne in mind when selecting a design flood magnitude
include:
i The site or pivotal station’s AM probability plot on EV1 and LN scales. [Editorial
note: There can be a trade-off for the practitioner between adopting a type of
probability plot in which they have gained experience and not pre-judging the most
appropriate probability distribution.]
ii The range of variation that can occur between random samples drawn from a single
population i.e. recognising that a single-site estimate may differ from the true value;
iii One or more pooled growth curves;
iv Preference in single-site analysis for a 2-parameter distribution (rather than a
3-parameter distribution);
Volume II Flood Frequency Estimation
124
v In homogenous pooling groups, the overall statistical superiority of “At-site QMED
plus pooled growth curve” estimation (see Box 7.2 in particular);
vi The robustness of “At-site QMED plus pooled growth curve” estimation to small
departures from pooling-group homogeneity;
vii Consideration of straight-line versus concave-downwards curves on probability plots,
and whether an implied upper bound on flood magnitude is physically meaningful;
viii Comparison of design water levels (corresponding to the proposed estimate of QT)
with “ground truth”, i.e. does the flood estimate make sense in terms of what has been
observed (or not observed) locally?
ix Consideration of the scope for underestimation when the “At-site QMED plus pooled
growth curve” estimate of QT is smaller than some of the observed floods at the
station;
x The credibility of outliers (i.e. unusually large or small flood values) and how they
might be verified, assessed and accommodated within the analysis.
10.2 Determining QT by single-site analysis
10.2.1 General guidance
QT may be estimated by single-site analysis provided that both:
The record length N is ten years or longer;
The required return period T is less than N, or not appreciably larger than N and
certainly no greater than 2N.
Ordinarily, a 2-parameter distribution should be used in single-site estimation: either the EV1
or LN2 distribution. In the case of EV1, the parameters should be estimated by L-moments
(see Section 10.2.3 below). In the case of LN2, the parameters can be estimated from the
mean and standard deviation of the logarithms of the AM data (see Section 10.2.4 below).
Use of L-moments on the untransformed AM data is also statistically efficient for fitting the
LN2 distribution, should the user prefer to use L-moment methods throughout.
It is assumed that the basic statistics of the station’s AM flow data have been computed and
that summary statistics (see Chapter 4) have been examined to see where they fall in the
range of observed values among all gauged catchments. In other words, it is assumed that the
user has carried out a thorough check of the station’s data: (i) as they stand, (ii) in the context
of data from similar types of catchment and (iii) in the context of the overall national dataset.
A starting point for consideration of catchment similarity is the distance metric dij used in
pooling-group construction, i.e. Equation 7.2. However, similarity in terms of other PCDs
may be important if the subject catchment has unusual features.
10.2.2 Plotting positions
The station’s AM flow values should be displayed on an EV1 based probability plot using
Gringorten plotting positions, and (using logged AM values) on a lognormal probability plot
using Blom plotting positions. These plotting positions are defined in Box 10.1.
Volume II Flood Frequency Estimation
125
Box 10.1: Plotting positions
10.2.3 Parameter estimation for EV1 distribution
EV1 parameters u and are estimated by the method of L-moments (see Box 4.2) as:
n2
λ
n2
M2Mα 2100110
10.1
and
α0.5772λα0.5772Mu 1100 10.2
Then:
TT yαuQ 10.3
where:
))T
1n(1n(yT 10.4
is the EV1 (or Gumbel) reduced variate.
EV1 (or Gumbel) plotting positions are defined by the EV1 reduced variate:
))n(Fn(y ii
where:
N,...2,1,i,0.12N
0.44iFi
is the Gringorten plotting-position formula.
These positions are suitable for use with the EV1 and GEV distributions.
Plotting positions for the Logistic (LO) and Generalised Logistic (GLO) distributions also
apply the Gringorten plotting-position formula but with the Logistic reduced variate:
i
iL
iF
F1ny
Plotting positions for the LN2 and LN3 distributions typically apply the Blom plotting-
position formula:
N...,2,1,i,1/4N
3/8iFi
with the standardised Normal variate:
))(FΦy i
1N
i
where:
Φ(y) is the standardised Normal distribution function.
Volume II Flood Frequency Estimation
126
10.2.4 Parameter estimation for LN2 distribution
Writing z = ℓnQ, LN2 parameters are estimated by the method of moments (see Box 4.1) as:
n
1i
iz zn
1μ 10.5
and
21
n
1i
2
iz zz1-n
1σ
10.6
Then
TZ
T eQ 10.7
where )T
1(1ΦZ 1
T and (F)Φ 1is the standardised inverse Normal distribution.
10.3 Determining QT from pooled data
10.3.1 When flood data are available at the subject site
Determining QMED
If a gauging station exists at the subject site then QMED is obtained directly from the AM
series of floods at that site. If the record is long, the standard error in QMED will be
relatively small (see Section 9.2). Since gauged data are so much better than information
obtained from a PCD-based formula, it is recommended to base the QMED value on the
observed data, even if the record is short.
An important proviso to this guidance is that flood data from upstream/downstream gauging
stations – as well as from other neighbouring stations – should be checked to see if the
QMED values obtained from the same period of record as those at the subject site are smaller
or larger than the long-term QMED at these sites. The ratio of long-term QMED to short-
term QMED can then be applied to the subject site’s QMED as an adjustment factor. Such
adjustments are discussed in Section 2.1.2.
Data transfer techniques (from gauged to subject site) are discussed in Chapter 11, in addition
to the detailed worked example of Section 2.7.
Determining xT
It is recommended that the growth factor be determined from pooled data in a pooling group
selected using the distance metric dij of Equation 7.2 and the 5T rule of Section 7.3.3. When
applied to estimation of the 100-year flood, the pooling group will be extended until a
minimum of 500 station-years of data are incorporated. It is suggested that such a pooling
group be used for all return periods rather than constructing a separate group for each return
period of interest.
If the hydrologist wishes to depart from this advice and to use geographical regions or
exclude certain types of catchment, the reasons for doing so will need to be documented.
Some diagnosis and amendment of the pooling group may be warranted if the pooled flood
Volume II Flood Frequency Estimation
127
data are strongly heterogeneous. However, it is important that a station with exceptional
flood data is excluded from the pooling group only if its physical characteristics are
fundamentally dissimilar to those of the subject catchment.
Consideration must also be given to the choice of a 2-parameter or 3-parameter distribution in
line with the discussion in Section 10.5 below. In the event that the single-site estimate of the
flood growth curve is steeper than the pooled one, consideration should be given to using a
combination of the single-site and pooled estimates of the flood growth curve or of the flood
frequency curve. While there is a possibility that this might lead to over-design it avoids (or
at least moderates) the less desirable outcome of under-design.
10.3.2 When the subject site is ungauged
The procedure for determining xT is as above. However, QMED has to be estimated from
PCDs using the methods set out in Chapter 2. Where appropriate, an urban adjustment factor
is to be applied. Indeed, it is recommended that the urban adjustment factor is habitually
applied, even when the catchment is essentially rural.
As described in Section 2.7, data transfers are a crucial part of QMED estimation at an
ungauged site. The recommended procedure requires the user to assess the most appropriate
data transfer by making a reasoned selection of the pivotal catchment. This is the gauged
catchment judged to be most relevant to the specific flood estimation problem.
Unless the subject and pivotal catchments are both rural, the recommended procedure
requires a clear head. It is the estimate of QMEDrural that is transferred from the pivotal site
to the subject site. Thus, the procedure is:
A value of QMEDrural is back-calculated from the gauged value of QMED at the
pivotal site by applying the urban adjustment factor of Equation 2.9 in reverse;
The data transfer is made from the pivotal site to estimate QMEDrural at the subject
site;
The relevant urban adjustment is made to obtain QMED at the subject site (i.e
applying Equation 2.9 in the normal manner).
The urban adjustment factor (UAF) is determined from the urban extent (URBEXT)
according to Equation 2.13. The procedure is further complicated if the user decides that
two or more QMED values at gauged sites are equally relevant, and a dual or multiple
transfer is required.
10.4 Characteristic examples of probability plot behaviour
10.4.1 Good straight-line behaviour
A first characteristic example is provided by Station 25016 Clodiagh at Rahan, which has 42
years of AM flow data. EV1 Probability Plot 10.1 shows near straight-line behaviour. In
Appendix H, it is assigned a score of 4 (meaning “good agreement with straight line”) and a
curve pattern of L2 (meaning “little deviation from straight line”).
Volume II Flood Frequency Estimation
128
EV1 Probability Plot 10.1: Single-site and pooled estimates, Station 25016
The Hazen skewness for this flood series is 0.62: considerably less than the theoretical value
of 1.14 in the EV1 distribution. Nevertheless, it seems reasonable to suggest that the 50-year
flood (or possibly even the 100-year flood) could be estimated by the EV1 distribution fitted
to the site AM flows in this case. From Box-plots 9.1, the standard error of Q100 in such a
case is expected to be approximately 6.5% to 10%, on the assumption that the underlying
parent distribution is indeed EV1.
The CV of the site AM flows is (at 0.22) rather lower than the national average of 0.27 for
110 A1 + A2 stations shown in Table 4.2. The pooled growth curves shown in EV1
Probability Plot 10.1 are nevertheless in general agreement with the site AM flows.
Despite the good behaviour of the single-site frequency analysis, it is reasonable to adhere to
the default recommendation to adopt the pooled estimate when estimating a design flood of
return period longer than the gauged record length of 42 years.
10.4.2 Good straight-line behaviour but single-site and pooled disagree
A second important example is provided by Station 09001 Ryewater at Leixlip which has 48
years of AM flow data. EV1 Probability Plot 10.2 shows excellent straight-line behaviour.
In Appendix H, it is assigned a score of 5 (meaning “very good agreement with straight line”)
and a curve pattern of L1 (meaning “perfect straight line”).
At 1.17, the Hazen skewness is practically equal to the theoretical value of 1.14 in the EV1
distribution, confirming the good fit of this distribution to the data. It seems reasonable to
suggest that the 50-year flood (or possibly even the 100-year flood) could be estimated by the
EV1 distribution fitted to the station’s AM data. From Table 9.2, the relative standard error
of Q100 in such a case is expected to be approximately 4% to 9%, on the assumption that the
underlying parent distribution is indeed EV1.
The CV of the site AM flows is (at 0.44) appreciably higher than the national average of 0.27
for 110 A1 + A2 stations shown in Table 4.2, and higher than that of other stations in the
hinterland. In consequence, the pooled growth curves shown in EV1 Probability Plot 10.2
are appreciably less steep than the single-site curve. Their adoption would lead to
underestimation of design floods.
25016 RIVER CLODIAGH @ RAHAN
EV1
2 5 10 25 50 100 5000
10
20
30
40
50
60
-2 -1 0 1 2 3 4 5 6 7 8EV1 y
AM
F(m
3/s
)AMF data
At-site EV1 fit
Pooled EV1 fit
pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
+ AM flow data
–– Single-site EV1 fit
–– Pooled EV1 fit
–– Pooled GEV fit
Volume II Flood Frequency Estimation
129
EV1 Probability Plot 10.2: Single-site and pooled estimates, Station 09001
However much faith is placed in the superiority of the pooled approach in general, it is a fact
that the pooled EV1 estimate of Q50 in this case has been exceeded three times in 48 years.
Indeed, the pooled GEV estimate of Q50 has been exceeded six times in 48 years! This
evidence cannot be overlooked when a final design flood is selected, even for long return
periods. It could be that the observed sample displays a gradient on the probability plot that
is at the upper end of the scale of steepness, and that the sample has come from a population
that actually has a less steep growth curve, i.e. with lower CV. However, no practical design
project would ignore the single-site analysis in such a case.
10.4.3 Concave upwards behaviour with outlier
Another important case is exemplified by Station 09010 Dodder at Waldron’s Bridge, which
has a relatively short record with just 19 years of AM flow data. Even if the underlying
parent growth curve is a straight line, the short record makes it more likely to exhibit
departure from linearity on the probability plot (see examples in EV1 Probability Plot 5.1).
EV1 Probability Plot 10.3 shows marked concave upwards behaviour. In Appendix H, it is
assigned a score of 2 (meaning “poor agreement with straight line”) and a curve pattern of U2
(meaning “severe concave upwards”).
EV1 Probability Plot 10.3: Single-site and pooled estimates, Station 09010
9001 RIVER RYEWATER @ LEIXLIP
2 5 10 25 50 100 500
0
20
40
60
80
100
120
140
-2 -1 0 1 2 3 4 5 6 7 8EV1 y
AM
F(m
3/s
)AMF data
At-site EV1 fit
Pooled EV1 fit
Pooled GEV fit
9010 RIVER DODDER @ WALDRON'S BRIDGE
500100502510520
50
100
150
200
250
300
350
400
-2 -1 0 1 2 3 4 5 6 7 8EV1 y
AM
F(m
3/s
)
AMF data
Atsite GEV fit
Pooled GEV fit
Pooled EV1 fit
+ AM flow data
–– Single-site EV1 fit
–– Pooled EV1 fit
–– Pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
+ AM flow data
–– Single-site GEV fit
–– Pooled EV1 fit
–– Pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
Hurricane Charlie
25/26 Aug 1986
Volume II Flood Frequency Estimation
130
This gauged record in South Dublin has attracted much comment already. In Section 2.5.2, it
is noted that the gauged value of QMED is much greater than indicated by the PCD model of
QMEDrural. Heavy urbanisation provides only a partial explanation of this and Bruen et al.
(2005) suggest that flood growth curves in this mid-Eastern region may be characteristically
steeper than in other parts of Ireland.
In Section 8.1, the station’s flood growth behaviour was noted to be the most discordant
amongst 88 A1 + A2 stations across Ireland. The variability and skewness of the AM flows
are both exceptionally high (CV = 0.86 and H-skew = 3.24; L-CV = 0.42 and L-skew = 0.42).
These are amongst the highest such summary statistics across the entire FSU dataset.
It was noted in Section 8.1.2 that the three most discordant stations (one of them Station
09010 Dodder at Waldron’s Bridge) are all close to Dublin and are all heavily urbanised.
Thus, flood behaviour at this site is exceptional both for the typical size of AM floods (high
value of QMED) and for their great variability (yielding an unusually steep growth curve). In
terms of flood risk, this might be viewed as the worst of all worlds. In mitigation, it is noted
that the incidence of Hurricane Charlie in August 1986 accounts for some of the high
curvature in EV1 Probability Plot 10.3.
While a 3-parameter GEV distribution provides a good graphical fit in this case, it has to be
borne in mind, from a theoretical point of view, that GEV single-site estimates have high
standard error especially with short datasets. That means that other similar sized later records
might not show the same pattern nor be so steep. Hence, notwithstanding the good fit of the
single-site GEV, it would normally be recommended that a pooled estimate be used for flood
estimation at this site except for estimates at short return period (T < 25 years). Because
other stations included in the pooling group inevitably have much smaller CV and skewness
than the Station 09010 data, the pooled growth curve is much lower than the single-site
growth curve. However, the pooled growth curve looks exceptionally low in comparison
with the observed data.
[Editorial note: This is an exceptionally difficult catchment to handle. It should be noted
that the three most discordant of 88 A1 + A2 catchments – all close to Dublin and all heavily
urbanised – were excluded from the simulation experiments reported in Chapter 9. This may
in part account for the relatively modest standard errors presented in Table 9.2. Whether the
three discordant catchments – Stations 08005 and 09002 and Station 09010 itself – were
made available for admission to the Station 09010 pooling group is unclear. The decision of
many analysts to exclude records from urbanised catchments when estimating flood growth
on a rural catchment is understandable. Important flood estimation problems in urbanised
catchments must, however, find a way of incorporating all available data.]
The hydrologist has to balance the weight of published evidence – indicating the advantages
of pooled over single-site methods – and their own beliefs about the representative nature of
the observed data and whether the behaviour shown could be repeated in another similar
length of record. In the present example for Station 09010, one has to ask whether the
single-site GEV estimate of Q100 ≈ 350 m3s
-1 is credible. The largest flood on record
(269 m3s
-1) was caused by Hurricane Charlie in August 1986. Its counterpart on the
neighbouring Dargle catchment had earlier been equalled twice in the previous 80 years.
Undoubtedly therefore, the flow of 269 m3s
-1 on the Dodder could occur again. But this does
not necessarily support the figure of Q100 = 350 m3s
-1. On the other hand, it is manifest that
Volume II Flood Frequency Estimation
131
the pooled curve is definitely too low. The pooled estimate of Q100 has been exceeded four
times in only 19 years!
This sort of dilemma presents itself whenever the subject site record has a CV and skewness
which are considerably larger than average (see also the example in Section 10.4.2). In such
cases, a prudent approach is to seek out and interpret historical information about flood
frequency on the particular river. In other words, it is necessary to reinforce the single-site
analysis in some way before crossing out the pooled analysis. Some compromise approach
that respects the single-site analysis at short return period but tempers the steepness and
curvature of its growth curve is another possibility. Regardless of the method adopted, it is
advisable in such cases not to place too much faith in estimates of QT for large T.
Somewhat similar behaviour is found for Station 29011 River Dunkellin at Kilcolgan Bridge
(N = 22 years, CV = 0.30, H-skew = 3.37, one notably large flood) and Station 36015 Finn at
Anlore (N = 33 years, CV = 0.32, H-skew = 2.62, two notably large floods). But the CV
values for these stations are not exceptionally large. It is the combination of high variability
and high skewness that makes Station 09010 particularly problematic. [Editorial note: The
FSU recommendation is to favour L-moment ratios (L-CV and L-skew) over conventional
moment ratios (CV and H-skew). Use of a dual currency risks confusion.]
10.4.4 Convex behaviour
Convex (or concave downwards) probability plots typically arise in cases where either there
are a number of high values which are almost equal or where there are one or more
exceptionally low annual maxima in the series. A fine example of a data series exhibiting
such behaviour is Station 25002 Newport at Barrington’s Bridge with 51 years of record (see
EV1 Probability Plot 10.4).
EV1 Probability Plot 10.4: Single-site and pooled estimates, Station 25002
The CV of 0.16 and Hazen skewness of -0.65 are both amongst the lowest values recorded in
the entire FSU dataset. While a 3-parameter GEV distribution is seen to provide a very good
graphical fit, it has to be borne in mind from a theoretical point of view that GEV single-site
estimates have high standard error with short datasets (although 51 years is not considered
short). More particularly, the positive value of the curvature parameter (k = 0.55) of the
fitted GEV distribution implies an upper bound to floods on this catchment. The upper bound
+ AM flow data
–– Single-site GEV fit
–– Pooled EV1 fit
–– Pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
140
120
100
80
60
40
20
0
Volume II Flood Frequency Estimation
132
of 78.39 m3s
-1 is only 5% greater than the three largest AM floods (each nominally assigned a
peak flow of 74.64 m3s
-1) on 3 December 1960, 12 December 1964 and 7 October 1967.
The closeness of the implied upper bound is not reasonable from a hydrological point of
view. While it might be satisfactory to use the fitted distribution up to a return period of
25 years, use of a pooled estimate is to be recommended for flood estimation at this site.
Because most other stations included in the pooling group will more than likely have
skewness greater than that of the site data, the pooled growth curve is less convex than the
single-site curve. Judgement of 50 and 100-year flood estimates on this catchment may
require that the pooled estimates are interrogated further in the context of historical
information about flooding on the river. It is known that this river – in common with the
neighbouring Mulkear – has been embanked since the 1920’s and that these embankments are
overtopped in about one out of every five years. This may go some way to explaining why
the three largest floods on record are exactly the same: each nominally assigned a peak flow
of 74.64 m3s
-1. [Editorial note: Station 25003 was amongst stations rejected from most
analyses in this volume, for reasons summarised in Appendix B. See also the further
discussion of upper bounds in Section 10.5.4.]
Other examples of this kind of behaviour are provided by Station 25021 Little Brosna at
Croghan (N = 44 years, CV = 0.14, H-skew = -0.13) and Station 34024 Pollagh at Kiltimagh
(N = 28 years, CV = 0.12, H-skew = -0.39). It is noted that two of these three examples are
based on records which are considered to be among the longest in the study.
10.4.5 Unclear behaviour with extreme outlier
An extreme example of a series with a high outlier is provided by Station 08009 Ward at
Balheary (N = 11 years, CV = 1.41, H-skew= 5.43) where ten of the 11 AM flows available
are less than 12 m3s
-1 but the largest is recorded as 53.6 m
3s
-1 on 12 June 1993. The single-
site and pooled estimates are shown in EV1 Probability Plot 10.5. The small sample-size
exacerbates the problem of trying to estimate a design flood at this location. This station has
by far the highest recorded skewness in the whole dataset, though it has to be acknowledged
that skewness measured from such a short record is unreliable.
EV1 Probability Plot 10.5: Single-site and pooled estimates, Station 08009
8009 RIVER WARD @ BALHEARY
2 5 10 25 50 100 500
0
10
20
30
40
50
60
70
80
90
-2 -1 0 1 2 3 4 5 6 7 8EV1 y
AM
F(m
3/s
)
AMF data
Atsite EV1 fit
Pooled EV1 fit
Pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
+ AM flow data
–– Single-site EV1 fit
–– Pooled EV1 fit
–– Pooled GEV fit
Volume II Flood Frequency Estimation
133
[Editorial notes: EV1 Probability Plot 10.5 shows data for 15 rather than 11 years. Four
annual maxima (for the 1979, 1981, 1982 and 1983 water-years) were omitted in the main
FSU analysis because they derived from notably incomplete records. In any case-specific
application, it is good practice (as here) to investigate exclusions and consider their
incorporation. As noted in Box 4.4, a recent study has queried both the completeness of
records at Station 08009 and the validity of the peak flow recorded on 12 June 1993.]
The single-site and pooled estimates for this station are fundamentally incompatible, and no
unique recommendation can be made. Further work might consider the following points:
The meteorological conditions leading to the June 1993 storm should be examined
and the likelihood of these being repeated anywhere in the region (including the
subject site) should be assessed.
Even if the rating curve or other features of the measurement are less than
satisfactory, the water level achieved locally will be known. This could be used as the
basis of the design of important works and the floor levels of new dwellings.
The hydrologist has to balance the competing beliefs that (i) the June 1993 flood was
so large that it could never happen again and (ii) if it happened once, it could occur
again. It is relevant to note that the Hurricane Charlie (August 1986) flooding in Bray
was so large that many believed it unrepeatable until it was discovered that almost
exactly similar events had previously occurred in 1904 and 1932. This underscores
the importance of historical review.
Further examples of high outliers – but not of such an extreme nature – are provided by
Station 36021 Yellow at Kiltybarden, Station 36031 Cavan at Lisdarn and Station 26006
Suck at Willsbrook. See also Table 4.3 in Section 4.6. Low outliers (e.g. Table 4.4) may
also be unduly influential in single-site estimation.
10.4.6 Irregular behaviour
Other kinds of irregular behaviour on probability plots include an elongated S shape or a
noticeable change of slope between lower and upper segments of the graph. Examples
include Station 09002 Griffeen at Lucan (N = 25 years) and Station 26008 Rinn at Johnston's
Bridge (N = 50 years). The single-site and pooled estimates for these stations are displayed
in EV1 Probability Plot 10.6EV1 Probability Plot 10.7 respectively.
EV1 Probability Plot 10.6: Single-site and pooled estimates, Station 09002
9002 RIVER GRIFFEEN @ LUCAN
2 5 10 25 50 100 5000
5
10
15
20
25
30
35
40
-2 -1 0 1 2 3 4 5 6 7 8EV1 y
AM
F(m
3/s
)
AMF data
Atsite EV1 fit
Pooled EV1 fit
Pooled GEV fit
+ AM flow data
–– Single-site EV1 fit
–– Pooled EV1 fit
–– Pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
Volume II Flood Frequency Estimation
134
Station 09002 is particularly problematical because there is a noticeable increase in slope
caused by three large floods in the record. This catchment has experienced an increase in its
urban fraction, especially in recent decades although no strong trend is evident in the time
series shown for this station in Figure 8.1. QMED is 5.25 m3s
-1. Adopting a relatively large
(for Ireland) value of x100 = 2.5 yields a Q100 of 13 m3s
-1. Yet this value has been amply
exceeded three times in 25 years. Additional local knowledge must be gained about the
meteorological and physical conditions leading to these floods. [Editorial note: Some
analyses of Station 09002 Griffeen at Lucan presented in this volume consider only 24 AM
flows, omitting the final modest AM value because it was preceded by two missing years.
See Figure 8.1.]
EV1 Probability Plot 10.7 shows an elongated S shape for Station 26008 Rinn at Johnston's
Bridge. While no single distribution could describe the probability plot, the upper end is not
too dissimilar to some of the random samples in EV1 Probability Plot 5.1. While application
of a pooled growth curve should provide satisfactory estimates of QT, it would be prudent to
check the sensitivity of analyses to inclusion/exclusion of the low outlier seen at this station.
EV1 Probability Plot 10.7: Single-site and pooled estimates, Station 26008
10.5 Additional notes on the choice of distribution and method
10.5.1 Problems in the use of 3-parameter distributions for single-site analysis
In ordinary circumstances a 3-parameter distribution should not be used with single-site data.
An exception could be made if the data series is very long, say > 50 years, and the required
return period is short, say ≤ 25 years.
A 3-parameter distribution is more flexible and may give a better fit visually on a probability
plot. However, the estimation of a third parameter has the effect of increasing the standard
error of the estimated quantile. See diagram (c) of Box 7.2. The situation can also occur
where use of a 3-parameter distribution in single-site analysis leads to a Q-T relation that is
not intuitively acceptable because of extreme upwards or downwards curvature.
26008 RIVER RINN @ JOHNSTON'S BRIDGE
500100502510520
10
20
30
40
50
60
-2 -1 0 1 2 3 4 5 6 7 8EV1 y
AM
F(m
3/s
)
AMF data
At-site EV1 fit
Pooled EV1 fit
Pooled GEV fit
+ AM flow data
–– Single-site EV1 fit
–– Pooled EV1 fit
–– Pooled GEV fit
AM
flo
w (
m3 s
-1)
-2 -1 0 1 2 3 4 5 6 7
EV1 reduced variate, y
Volume II Flood Frequency Estimation
135
10.5.2 Reconciling single-site and pooled analyses
In the event that the probability plot shows a marked departure from straight-line behaviour –
either by way of upwards or downwards curvature or because of some form of an elongated
S shape – consideration must be given to applying a pooled growth curve xT to the gauged
value of QMED. Even this may leave doubt about the suitability of the method chosen if the
pooled growth curve differs substantially from whatever general pattern is shown by the site
data e.g. as in EV1 Probability Plot 10.2.
A decision may have to be made to trust the pooled growth curve as a matter of good
practice: based on the well-documented reduced standard error of estimate and robustness of
the pooling method. See diagram (d) of Box 7.2. When AM flows at the site display CV and
skewness greatly in excess of the pooled values – as in EV1 Probability Plot 10.3 – it can,
however, be difficult to trust the pooled growth curve. If a very large flood is observed
during the period of record the question arises as to whether it should over-ride any more
modest estimate of QT obtained by pooled analysis.
A progressive approach can be to use a weighted combination of the pooled and single-site
estimates (see Box 10.2). However, it should be noted that the relative weight given to each
component cannot presently be specified by any rule based on scientific evidence. The
weight must be chosen subjectively and supported by rational argument alone.
Box 10.2: Combined use of single-site and pooled estimates
10.5.3 Probability associated with a very large recorded flood
Where a very large observed flood has occurred, it is possible – under certain assumptions –
to calculate the probability that such a large flood could occur. For instance, if a sample of
50 floods are drawn randomly from an EV1 distribution with CV = 0.3, the probability that
the largest flood would exceed 1.25 Q100 – where Q100 is the population value of the 100-year
flood – is less than 7% and the probability that it would exceed 1.5 Q100 is less than 1%.
In practice, the true value of Q100 is unknown and has to be replaced by an estimate which
makes these percentage probabilities less reliable. Nevertheless, such calculations can be
adapted to provide an informal test of the hypothesis that the observed outlier is consistent
with the assumed parent population. The probabilities quoted above are based on the fact that
if Q is distributed as an EV1 with parameters u and , then the largest flood in N years (Qmax)
is distributed as an EV1 with parameters u + ℓnN and .
Editorial note: The FSU Web Portal supports the combined use of single-site and pooled
estimates. The site and pooled flood growth curves are combined by taking weighted-
averages of the site and pooled L-Moment ratios, specifically:
L-CVcomb = ω L-CVsite + (1 - ω) L-CVpooled 10.8
L-skewnesscomb = ω L-skewnesssite + (1 - ω) L-skewnesspooled 10.9
L-kurtosiscomb = ω L-kurtosissite + (1 - ω) L-kurtosispooled 10.10
where ω takes a value between 0 and 1.
Volume II Flood Frequency Estimation
136
10.5.4 Flood growth curves with an upper bound
If the AM flow series shows negative skewness, and the three or four largest floods differ
from one another by only a very small amount, the data series will usually exhibit convex
curvature (e.g. EV1 Probability Plot 10.4). In such situations, any 3-parameter distribution
fitted to the data will usually have an upper bound which is not very much larger than the
largest recorded flood. The GEV and GLO have an upper bound if the shape parameter k is
greater than zero.
While it is unwise to be prescriptive, it is seldom hydrologically realistic if the upper bound
to flooding (implied by the fitted distribution) is only a little greater than one or more floods
already experienced at the station. This objection applies regardless of whether the flood
frequency curve has been derived by single-site or pooled analysis. The experienced
hydrologist knows that some unprecedented rainfall could occur in the future which leads to a
runoff or routing mechanism different to all previous floods and which could therefore
produce a flood much in excess of the largest flood on record.
The general recommendation in such cases is to consider instead use of a 2-parameter
distribution growth curve that yields a straight-line growth curve on the relevant probability
plot. Nevertheless, judgement is required before a design flood can be specified. There is the
possibility that a straight-line 2-parameter distribution, when extrapolated to estimate very
rare floods, may produce flood estimates that are implausibly large in the context of the
relevant physical factors and the known water levels reached during previous floods. This is
an example where the skewness of the model is potentially greater than the skewness of the
parent data. See diagram (b) in Box 7.2. The hydrologist will need to apply judgement to the
estimation process in such a case.
Volume II Flood Frequency Estimation
137
11 Data transfers revisited
Many flood estimation problems arise at subject sites that are ungauged. The transfer of
information from gauged to subject site is a fundamental matter to be addressed by
practitioners.
The topic has been touched on elsewhere, not least in the detailed example of QMED
estimation at an ungauged site in Section 2.7. However, data transfer is of such importance
that some final remarks are made in Section 11.2, after reporting an interim assessment
undertaken as part of the FSU research.
11.1 Interim assessment of QMED data transfers
11.1.1 Subject sites used in the interim assessment
The assessment is made by treating gauged sites as if they were ungauged. Some 184 gauged
sites were identified as potentially suitable candidates on which to base the assessment.
It was decided to focus on rivers with at least three gauging stations, each with at least 20
years of data. The 38 gauging stations thus identified lie in ten different river systems. The
stations used are identified in Table 11.1.
Table 11.1: Rivers having three or more gauging stations for assessment
Hydrometric
Area (HA) River
name
#
gauges
Station numbers within HA
No. Name Most
u/s
Most
d/s
06 Glyde Glyde 3 26 14 21
14 Barrow Barrow 6 5 6 19 34 18 29
16 Suir Suir 5 4 2 8 9 11
18 Blackwater Blackwater 5 16 50 48 6 3
24 Shannon Estuary Deel 3 30 11 12
24 Shannon Estuary Maigue 3 4 82 8
25 Lower Shannon Brosna 3 124 6 11
26 Upper Shannon Suck 4 6 2 5 7
26 Upper Shannon Inny 3 58 59 21
36 Erne Erne 3 12 11 19
To the extent possible, the record-length criterion insulates the assessment from uncertainty
in observing QMED itself (see Section 9.2). The selection of rivers with three such gauges
allowed assessment of the relative merits of data transfer from upstream and downstream
sites. Comparisons are also made with data transfers from more distant (but possibly
hydrologically-similar) gauged catchments i.e. from gauged stations that do not lie either
upstream or downstream of the subject site.
Volume II Flood Frequency Estimation
138
11.1.2 Methods used in the interim assessment
The approach used in the interim assessment of data transfers loosely resembles the final data
transfer procedure recommended and illustrated in Section 2.7. The procedure for estimating
QMED at an ungauged site has three main steps:
Selection of the pivotal catchment;
Estimation of QMED from PCDs at the subject site and at the pivotal site;
Adjusting QMED at the subject site by reference to performance of the PCD model in
estimating QMED at the (pivotal) gauged site.
Because a number of different data-transfer strategies were being compared in the interim
assessment, use of the term donor catchment is appropriate here. [Editorial note: Use of the
term pivotal catchment would give the impression that a prior judgement had been made.]
The interim assessment compared a number of data transfers to each of the 38 subject sites
studied. These experimental subject sites were the 38 gauged sites identified in Table 11.1.
Each data transfer used only one donor catchment at a time.
Selection of the donor catchment
The donor site selection chose the nearest gauge on the same stream as the subject site. The
2nd
nearest, 3rd
nearest etc. in the upstream and downstream directions are also examined
where such sites exist. Use of a data transfer from an upstream or downstream site is good
practice and the donor site is often self-selecting. One would normally use the nearest such
gauge measured along the river network or the one whose AREA ratio (compared to the
AREA to the subject site) is closest to 1.0.
The interim assessment also tested the effectiveness of a data transfer from outside the
catchment. For historical reasons largely related to terminology introduced by the FEH, a
distant donor catchment is termed an analogue catchment. Whereas selection of an upstream
or downstream site as the pivotal catchment is relatively straightforward, selection of an
analogue catchment is inherently controversial. [Editorial note: For the transfer to be
justifiable, the analyst has to be convinced that the donor catchment is both (i) hydrologically
similar to the subject catchment, and (ii) the most hydrologically similar (or otherwise most
appropriate) of all such gauged catchments available to act as a donor.] The selection of an
analogue catchment is never straightforward.
For the interim assessment, the analogue catchment was selected as the “nearest” (i.e. least
dissimilar) catchment according to the dissimilarity metric of Equation 7.2. For reasons
explained in Section 2.6.4 (see also Section 11.2), the FSU recommendation is opposed to
any automated selection of the pivotal catchment. However, the strategy had the merit of
allowing the interim assessment to proceed in a straightforward manner.
Estimation of QMED from PCDs at the subject site and at the donor site
For logistical reasons, the assessment used a different model for estimating QMED from
PCDs than that developed and recommended in Chapter 2. The model uses just three PCDs:
-1.5390.8980.829 BFISAARAREA0.000302QMED 11.1
Volume II Flood Frequency Estimation
139
In terms of the particular dataset of 164 catchments used in its calibration, the model explains
82% of the total variance in ℓnQMED (i.e. r2 = 0.82). It yields a root mean standard error
(RMSE) of 0.472 in the ℓn domain, so that the factorial standard error is FSE = e 0.472
= 1.60.
As is to be expected, the model is inferior to the best models reported in Section 2.2. For
example, Table 2.6 reports a 3-variable model having r2 = 0.843 and RMSE = 0.407.
Use of the baseflow index in Equation 11.1 is somewhat controversial. BFI is formally
known only at gauged sites with daily mean flow data. The results that follow should
therefore be treated as interim rather than authoritative.
Adjusting QMED at subject site by reference to performance of PCD model at donor site
The method of data transfer used in the interim assessment was a full (or “hard”) data
transfer. Thus, if the PCD model is found to underestimate QMED by 33% at the donor site,
the PCD estimate of QMED at the subject site is increased by factor of 100/67, i.e. by 49%.
11.1.3 Results
The performances of various data transfer methods at five of the 38 experimental sites are
summarised in the figures below. The number above each adjusted estimate denotes the area
of the donor catchment in km2. A perfect data transfer would yield an adjusted estimate
equal to 100% of the gauged QMED at the site. In the data transfer, each of the five sites is
treated as if it were ungauged.
Figure 11.1: Performance at Station 16002 Suir at Beakstown (512 km
2)
Figure 11.2: Performance at Station 16008 Suir at Newbridge (1120 km2)
236 1120
1602 2173
309
0
100
236
512 1602 2173
1207
0
100
1st u/s 1
st d/s 2
nd d/s 3
rd d/s Analogue PCDs
% o
f obse
rved
QM
ED
%
of
ob
serv
ed Q
ME
D
2nd
u/s 1st u/s 1
st d/s 2
nd d/s Analogue PCDs
Volume II Flood Frequency Estimation
140
182 1050 1184
152
0
100
113 881
1058 233
293
0
100
Figure 11.3: Performance at Station 18050 Blackwater at Duarrigle (244.6 km2)
Figure 11.4: Performance at Station 24008 Maigue at Castleroberts (805 km2)
Figure 11.5: Performance at Station 26002 Suck at Rookwood (626 km2)
Consideration of these and other cases led to the interim conclusions that:
Data transfers from donor stations located upstream or downstream of the subject site
tend to perform better than a data transfer from an analogue site;
Some limited evidence was found to suggest that data transfer from a gauge sited
downstream of the subject site typically performs a little more strongly than data
transfer from a gauge sited upstream of the subject site.
Most data transfers (even from an analogue site) perform better than estimating
QMED from PCDs alone.
It must be recalled that the above assessment was made with an interim method of estimating
QMED from PCDs, not with the recommended method developed in Chapter 2.
246 764
1063
0
100
% o
f ob
serv
ed Q
ME
D
2nd
u/s 1st u/s 1
st d/s 2
nd d/s Analogue PCDs %
of
ob
serv
ed Q
ME
D
1st u/s 1
st d/s Analogue PCDs
% o
f obse
rved
QM
ED
1st u/s 1
st d/s 2
nd d/s Analogue PCDs
Volume II Flood Frequency Estimation
141
11.1.4 Remarks
The user is reminded that the estimation of QMED plays a particularly important role in flood
estimation in many Irish rivers, especially those of low gradient and a sluggish flood
response. Such rivers typically have flood growth curves of low gradient, making estimation
of the index flood especially influential on design flood magnitudes. This view is
underscored by the Chapter 9 discussion of uncertainty which ascribes – under certain rather
large assumptions – that most of the uncertainty in estimating QT in Irish rivers derives from
uncertainty in estimating QMED.
It is anticipated that the wider application of data transfer techniques will promote a healthy
discussion of many factors, including:
Methods of assessing hydrological similarity;
Strengths and weaknesses of the national hydrometric network in relation to sites at
which flood estimates are required;
The value of installing a flood gauge locally when flood frequency estimation at a
critical subject site is not well served by gauged networks.
11.2 Further guidance on pivotal catchments and data transfers
Implementation of the FSU research requires the user to make some important judgements.
The pivotal catchment is the user’s assessment of the most relevant catchment on which to
base a data transfer. Where flood data are available from a gauge sited upstream or
downstream of the subject site, this will often be readily selected as the pivotal catchment. In
other cases, the selection is likely to be more precarious and to hinge on the user’s judgement
of catchment similarity.
An automated judgement of catchment similarity is likely to give weight to differences in a
few leading factors – e.g. catchment size (represented by AREA), catchment wetness
(indexed by SAAR) and catchment permeability (indexed by BFI or BFIsoil) – and to neglect
all other factors. This is not a safe approach.
A particular feature present on one catchment and absent on another may lead to strong
differences in their flood behaviour. Arterial drainage (indexed by ARTDRAIN and
ARTDRAIN2) is perhaps the most notable such feature. Section D4 of Appendix D finds
evidence that BFIsoil and ARTDRAIN2 are important in characterising the post-drainage
flood response of a catchment, whilst the descriptors DRAIND and S1085 are more important
in characterising the response of undrained catchments. These findings may help the
experienced user to judge which PCDs to examine closely when judging hydrological
similarity as a means of selecting the pivotal catchment.
Other notable features to consider when assessing catchment similarity are the extent of
urbanisation (indexed by URBEXT) and the presence of large lakes (indexed by FARL).
Research reported in Section 2.3.4 endorses the recommendation to favour geographical
closeness – as well as similarity in key PCDs such as FARL – when selecting a pivotal
catchment for use in a particular flood estimation problem.
Volume II Flood Frequency Estimation
142
Selection of the pivotal catchment is a demanding task that calls for reasoned judgements.
Some judgements will be unsettling because of the impact of the data transfer on the final
flood estimates. In some cases – not least on small catchments – the pivotal catchment
selected will not be wholly convincing. This is not a reason to abandon making a data
transfer. However, it may be good reason to consider making only a partial transfer. An
approach to making a partial transfer is included in Step 5 of the procedure illustrated in
Section 2.7.1.
Two final remarks concern the treatment of urbanised catchments in data transfers. It is
recommended that data transfers are applied to the rural element of QMED. The relevant
procedure is explained in Box 2.3 in Section 2.7. Second, it is recommended that the urban
adjustment factor UAF is applied in all cases, even to subject catchments that are almost
entirely rural. With ever-increasing automation, and a persistent interest in making flood
estimates throughout a river basin, it is unhelpful to have a procedure in which an urban
adjustment is incorporated at some sites but not others.
Volume II Flood Frequency Estimation
143
12 Summary and conclusions
12.1 Data
Data from some 200 gauging stations in the Republic of Ireland were available from the
archives of OPW, EPA and ESB. Of these, 115 were of Grade A and 67 were Grade B.
Data series at 17 stations had both pre and post-drainage records available. Data from most
of these stations played an active role in the analysis and modelling of QMED reported in
Chapter 2.
Summary statistics (Chapter 4) and probability plots (Chapter 5) were prepared for all these
stations. However, no strong inferences were drawn from those of the Grade B stations.
12.2 Descriptive statistics
Examination of the data for trend and randomness (Chapter 3) found that about 10% of
records displayed some trend or lack of randomness. The usual assumption that AM flows
are statistically independent and identically distributed was nevertheless adopted, supported
by the fact that 90% of records did not show significant trend or lack of randomness.
The descriptive statistics indicate that – in comparison to international data and many UK
data – Irish AM flow data typically have low variability and low skewness, whether judged in
terms of traditional statistics (CV and skewness) or in terms of L-moment ratios (L-CV and
L-skewness). This possibly reflects the shallow longitudinal slopes of many Irish rivers.
Examination of probability plots and moment ratio diagrams suggests that – among
2-parameter distributions – the Extreme Value Type I (EV1) and lognormal (LN2) are
typically the most appropriate for use in Ireland. Some exceptions were found, not least for a
number of stations that display very low skewness. In some cases, this reflects that the three
or four largest values in the record are not appreciably different in magnitude from each
other.
Explanations for this latter phenomenon were sought in Section 5.4 by considering flood
volumes and by referring to the physical catchment descriptor FAI: an index of the
attenuation of flood magnitudes expected from floodplain storage effects. However, no
empirical evidence was found to explain the phenomenon.
12.3 Seasonal analysis
Seasonal analysis shows that two thirds of AM flows occur during the winter months of
October to March. At 11 stations, no AM floods occurred outside this winter season.
July is the least likely month in which AM floods occur while August and September are the
most likely summer months to supply AM floods.
At 20 (10%) of the 202 stations examined, the largest flow on record (i.e. the series
maximum, Qmax) occurred during summer, with eight of the occurrences in August. For 91
(45%) of the 202 stations, the series maximum occurred in the single month of December.
Volume II Flood Frequency Estimation
144
12.4 Estimation of design flood
Estimation of the T-year flood is considered for each of a wide range of circumstances that
can arise in practice: from gauged to ungauged sites and from short to long flow records. The
index flood method (see Sections 6.2 and 7.1) is recommended in which QT is expressed as:
QT = QMED × xT
Here QT is the flood of return period T years, QMED is the median of the AM flows at the
subject site and xT is the growth factor appropriate to the subject site. Unless there is a long
record at the subject site, xT is estimated by pooled analysis of AM flow data from an
appropriate homogeneous pooling group.
QMED is estimated from the gauged AM flow data, where available. If the data series is
long, the QMED value obtained is used directly. If the available flow record is short, or if
QMED is obtained (in the ungauged case) from a PCD-based equation, QMED is adjusted
with the assistance of data from a relevant gauged site judged by the user to be pivotal to
flood estimation at the subject site.
Tests conducted on data of 38 gauging stations located in ten different river systems were
used to assess the efficacy of different data transfer methods for improving the QMED
estimate. Data transfer from a downstream donor site was (on average) found to be slightly
superior to data transfer from an upstream donor site. Both were found to be superior to data
transfer from a catchment judged to be locally relevant or hydrologically similar by other
criteria (i.e. other than being directly upstream or downstream of the subject site).
In practical applications – at all ungauged sites and some gauged sites – the user must justify
their choice of the pivotal catchment. This is the user’s assessment of the most relevant
catchment on which to base a data transfer to the subject site. A worked example in
Section 2.7 illustrates the detailed mechanics of data transfer to refine the estimate of QMED
at an ungauged or short-record site.
Estimation of xT may sometimes be based on single-site analysis, if a sufficiently long data
record exists at the subject site. Otherwise, xT is estimated from the dimensionless L-CV and
L-skewness values obtained by averaging these quantities across a pooling group of gauged
catchments chosen to be hydrologically similar to the subject catchment. There can
sometimes be scope to combine single-site and pooled estimates of xT (see Box 10.2 in
Section 10.5.2).
Suitable members of the pooling group are chosen with the help of a dissimilarity metric dij.
Tests were carried out into the effectiveness of different combinations of PCDs in the
definition of dij. The most effective metric was based on differences in catchment size
(indexed by differences in ℓnAREA), differences in wetness (indexed by differences in
ℓnSAAR) and differences in permeability (indexed by differences in the baseflow index BFI).
Where a gauged value of BFI is unavailable, the physical catchment descriptor BFIsoil is
substituted.
Tests have been carried out into the effect on the estimated value of xT of catchment size
(measured by AREA), peat coverage (indexed by PEAT), “lakiness” (indexed by FARL),
geographical location and the available period of record varying from the 1950s to the 1990s.
Volume II Flood Frequency Estimation
145
None of these effects was judged to be sufficiently influential to make additional provision
for, other than through their influence on QMED.
Although some regard is to be paid to the homogeneity (or not) of pooled flood data,
catchments with discordant values of L-moment ratios should be rejected from a pooling
group only if they are found to be physically dissimilar to the subject catchment. Catchments
should not be rejected on the basis of their flood statistics alone.
The standard errors associated with estimates of xT and QT are investigated in Chapter 9. The
standard error of QT estimated by the index flood method is dominated by SE(QMED).
Consequently, when expressed as a percentage, SE(QT), varies only slightly with T. When
QMED is estimated from gauged data and xT is estimated from a pooling group containing
approximately 500 station-years of data, SE(QT) is of the order of 5 to 10% of QT regardless
of return period. If QMED is estimated from a PCD-based formula alone, and xT is estimated
from a pooling group containing approximately 500 station-years of data, SE(QT) is of the
order of 37% QT. [Editorial note: One route to developing greater certainty about
uncertainty may be to use resampling techniques in which the resampling is arranged by year
across all AM datasets implicated in the pooled and single-site analyses of a particular flood
estimation problem. Such resampling respects important spatial structure in AM flows and is
capable of revealing the degree of sensitivity of the final answers to the particular years of
record available.]
Chapter 10 provides guidelines for the estimation of QT, both from site and pooled data.
Several points that need to be taken into account in practical cases are outlined. It is
emphasised that blind use of a prescribed method can sometimes lead to a QT estimate which
is not always in accordance with the “ground truth” of gauged or historical flood data. This is
especially the case when single-site AM flow data have higher-than-average variability
and/or skewness. A number of examples are discussed, most of which throw up practical
problems of the type met in practice where difficult choices have to be weighed and decisions
made. Discussion is also provided on flood distributions with upper bounds, and on the
choice between 2 and 3-parameter distributions.
In some special cases, the user may need to consider single-site estimate in conjunction with
(and sometimes even in preference to) the pooled method of growth-curve estimation that lies
at the heart of the general FSU recommendations for estimating the T-year flood peak. When
competing estimates of QT are being compared or assessed at a gauged site, it is essential to
view probability plots that also show the AM flow data for the site. It is crucial to be alert to
information about historical floods experienced at the site or elsewhere on the river. It is
better to use such information subjectively than to ignore it entirely.
In conclusion, flood estimation cannot be reduced to a strict formula-based procedure.
Individual analysts must make choices which depend on the problem circumstances and
which take into account their own knowledge and experience. Users will be expected to
present and record a reasoned argument for the choices made. “Because it gives a smaller
number” is never an adequate reason.
Volume II Flood Frequency Estimation
146
Acknowledgements
The work reported was undertaken principally by NUI Galway and NUI Maynooth, with
contributions from OPW and Hydro-Logic Ltd. The help of many organisations and
individuals is gratefully acknowledged.
Particular thanks go to hydrometric staff in gauging authorities and to members of the FSU
Technical Steering Group.
Volume II was edited by Duncan Reed of DWRconsult, who added Sections 2.7 and 11.2.
References
Ahilan, S., O’Sullivan, J.J. and Bruen, M. 2012. Influences on flood frequency distributions
in Irish river catchments. Hydrol. Earth Syst. Sci., 16: 1137–1150.
Bayliss, A.C. and Reed, D.W. 2001. The use of historical data in flood frequency estimation.
Report to MAFF, 87pp. http://nora.nerc.ac.uk/8060/1/BaylissRepN008060CR.pdf
Bhattarai, K. and Baigent, S. 2009. The hydrological analysis for the Fingal East Meath
Flood Risk Assessment and Management Study. Proc. National Hydrology Seminar,
Tullamore, 10 November 2009, 58–67.
Bhattarai, K.P. and O’Connor, K.M. 2004. The effects over time of an arterial drainage
scheme on the rainfall-runoff transformation in the Brosna catchment. Phys. and Chem. of
the Earth, 29: 787–794.
Bruen, M., Gebre, F., Joyce, T. and Doyle, P. 2005. The Flood Studies Report ungauged
catchment method underestimates for catchments around Dublin. Proc. National Hydrology
Seminar, Tullamore, 15 November 2005, 34–45.
http://www.opw.ie/hydrology/data/speeches/F_BRUEN.PDF
Buishand, T.A. 1982. Some methods for testing the homogeneity of rainfall records. J.
Hydrol., 58: 11–27.
Burn, D.H. 1990. Evaluation of regional flood frequency analysis with a region of influence
approach. Wat. Resour. Res., 26: 2257-2265.
Cawley, A.M., Fitzpatrick, J., Cunnane, C. and Sheridan, T. 2005. A selection of extreme
flood events – the Irish experience. Proc. National Hydrology Seminar, 15 November 2005,
14–25. http://www.opw.ie/hydrology/data/speeches/d_cawley.pdf
Cohn, T.A., Lane, W.L. and Baier, W.G. 1997. An algorithm for computing moments-based
flood quantile estimates when historical flood information is available. Wat. Resour. Res., 33:
2089-2096.
Cunnane, C. 1989. Statistical distributions for flood frequency analysis. Oper. Hydrol. Rep.
33, WMO 718, World Meteorological Organization, Geneva, 73pp + appendices.
Dalrymple, T. 1960. Flood frequency methods. In: U.S. Geol. Survey Water Supply Paper
1543A, Washington: 11–51.
Das, S. 2009. Examination of flood estimation techniques in the Irish context. Unpublished
PhD Thesis, NUI Galway, 236pp.
Volume II Flood Frequency Estimation
147
Das, S. and Cunnane, C. 2011. Examination of homogeneity of selected Irish pooling groups.
Hydrol. Earth Syst. Sci., 15: 819–830.
Das, S. and Cunnane, C. 2012. Performance of flood frequency pooling analysis in a low CV
context. Hydrol. Sci. J., 57: 433–444.
Dawson, C.W., Abrahart, R.J., Shamseldin, A.Y. and Wilby, R.L. 2006. Flood estimation at
ungauged sites using artificial neural networks. J. Hydrol., 319: 391–409.
Efron, B. 1987. Better bootstrap confidence intervals. JASA, 82: 171–185.
Fiering, M.B. 1963. Use of correlation to improve estimates of the mean and variance.
Statistical studies in hydrology, Geological Survey professional paper 434-C, US Gov.
Printing Office, Washington, 9pp.
Fisher, R.A. and Tippett, L.H.C. 1928. Limiting forms of the frequency distribution of the
largest or smallest member of a sample. Proc. Cambridge. Phil. Soc., 24:180–191.
Foster, H.A. 1924. Theoretical frequency curves and their application to engineering
problems. Trans. ASCE, 87: 142–173.
Fotheringham, A.S., Brunsdon, C. and Charlton, M. 2002. Geographically weighted
regression: the analysis of spatially varying relationships. Wiley-Blackwell, 284pp. [See
also: http://eprints.ncrm.ac.uk/90/1/MethodsReviewPaperNCRM-006.pdf by same authors.]
Fuller, W.E. 1914. Flood flows. Trans. ASCE, 77: 564–617.
Greis, N.P. and Wood, E.F. 1981. Regional flood frequency estimation and network design.
Water Resour. Res., 17: 1167–1177.
Grover, P.L., Burn, D.H. and Cunderlik, J.M. 2002. A comparison of index flood estimation
procedures for ungauged catchments. Canadian J. of Civ. Engg, 29: 734–741.
Gumbel, E.J. 1941. The return period of flood flows. Annals of Math. Statist., 12: 163-190.
Guo, S.L. 1990. Unbiased plotting position formulae for historical floods. J. Hydrol., 121:
45–61.
Guo, S. and Cunnane, C. 1991. Evaluation of the usefulness of historical and palaeological
floods in quantile estimation. J. Hydrol., 129: 245–262.
Hazen, A. 1930. Flood flows: a study of frequencies and magnitudes. John Wiley, New
York, 199pp.
Hebson, C.S. and Cunnane, C. 1987. Assessment of use of at-site and regional flood data for
flood frequency estimation. In: V.P. Singh (ed.), Hydrologic frequency modelling, Reidel,
Dordrecht, 433–448.
Helsel, D.R. and Hirsch, R.M. 1992. Statistical methods in water resources. Studies in
Environmental Science 49. Elsevier, Amsterdam, The Netherlands, 522pp.
Hirsch, R.M., Helsel, D.R., Cohn, T.A. and Gilroy, E.J. 1993. Statistical analysis of
hydrologic data. In: D.R. Maidment (ed.), Handbook of hydrology, McGraw-Hill, New York,
17.1–17.55.
Hosking, J.R.M. 1990. L-moments: analysis and estimation of distributions using linear
combinations of order statistics. Journal of the Royal Statistical Society, Series B, 52: 105–
124.
Hosking, J.R.M., Wallis, J.R. and Wood, E.F. 1985a. An appraisal of the regional flood
frequency procedure in the UK Flood Studies Report. Hydrol. Sci. J., 30: 85–109.
Volume II Flood Frequency Estimation
148
Hosking, J.R.M., Wallis, J.R. and Wood, E.F. 1985b. Estimation of the generalized extreme-
value distribution by the method of probability-weighted moments. Technometrics, 27: 251–
261.
Hosking, J.R.M. and Wallis, J.R. 1997. Regional frequency analysis: an approach based on
L-Moments. Cambridge Univ. Press, 224pp.
Irish Academy of Engineering 2007. Ireland at Risk, 1: The impact of climate change on the
water environment. Proc. workshop, RDS, Ballsbridge, Dublin, May 2007, available at
www.iae.ie/publications.
Jakob, D., Reed, D.W. and Robson, A.J. 1999. Selecting a pooling-group (B). Chapter 16,
Volume 3, Flood Estimation Handbook, CEH Wallingford, 153–180.
Kaczmarek, Z. 1957. Efficiency of the estimation of floods with a given return period. Proc.
Toronto Symp., IAHS Publ. 45, III: 144–159.
Kendall, M.G. 1975. Rank correlation methods. 4th
ed, Charles Griffin, London, 202pp.
Kimball, B.F. 1949. An approximation to the sampling variance of an estimated maximum
value of given frequency based on fit of doubly exponential distribution of maximum values.
Annals of Math. Statist., 20: 110–113.
Kjeldsen, T., Jones, D., Bayliss, A., Spencer, P., Surendran, S., Laeger, S., Webster, P. and
McDonald, D. 2008. Improving the FEH statistical method. Proc. Flood & Coastal
Management Conference 2008, University of Manchester, 1-3 July 2008, Environment
Agency/Defra. Also available at: http://nora.nerc.ac.uk/3545/, 10pp.
Kundzewicz, Z.W. and Robson, A. (eds.) 2000. Detecting trend and other changes in
hydrological data. World Climate Programme – Water, WCDMP-45, WMO/TD 1013,
World Meteorological Organization, Geneva.
Kundzewicz, Z.W. and Robson, A.J. 2004. Change detection in hydrological records – a
review of the methodology. Hydrol. Sci. J., 49: 7–19.
http://iahs.info/hsj/491/hysj_49_01_0007.pdf
Langbein, W.B. 1949. Annual floods and the partial-duration flood series. Trans. Am.
Geophys. Union, 30: 879–881.
Lehmann, E.L. 1975. Nonparametrics: Statistical methods based on ranks. Holden-Day,
San Francisco, 457pp.
Lettenmaier, D.P., Wallis, J.R. and Wood, E.F. 1987. Effect of regional heterogeneity on
flood frequency estimation. Water Resour. Res., 23: 313–323.
Lowery, M.D. and Nash, J.E. 1970. A comparison of methods of fitting the double
exponential distribution. J. Hydrol., 10: 259–275.
Lu, L-H. and Stedinger, J.R. 1992. Variance of two- and three-parameter GEV/PWM
quantile estimators: Formulae, confidence intervals, and a comparison. J. Hydrol., 138: 247–
267.
Lynn, M. A. 1981. Estimating flood magnitude/return period relationships and the effect of
catchment drainage. Hydrology Unit Report, Office of Public Works.
Mandal, U.K. 2011. Studies in low and flood flow estimation for Irish river catchments.
PhD thesis, College of Engineering and Informatics, NUI Galway, 282pp.
Mann, H.B. 1945. Non-parametric test against trend. Econometrica, 13: 245–259.
Volume II Flood Frequency Estimation
149
Mason, D. W. 1992. Modelling the effect of flood plain storage on the flood frequency
curve. Ph.D. thesis, Univ. of Newcastle upon Tyne, 1992.
Matalas, N.C., Slack, J.R. and Wallis, J.R. 1975. Regional skew in search of a parent. Water
Resour. Res., 11: 815–826.
McCuen, R.H., Leahy, R.B. and Johnson, P.A. 1990. Problems with logarithmic
transformations in regression. ASCE J. Hydraul. Engg, 116: 414–428.
Morris, D.G. 2003. Automation and appraisal of the FEH statistical procedures for flood
frequency estimation. CEH Wallingford report to Defra, Project FD1603, 207pp.
Nash, J.E. and Amorocho, J. 1966. The accuracy of the prediction of floods of high return
period. Water Resour. Res., 2: 191–198.
Nash, J.E. and Shaw, B.L. 1965. Flood frequency as a function of catchment characteristics.
Institn of Civ. Engin
rs, Proc. Symp. on River Flood Hydrology (published 1966), 115–136.
NERC 1975. Flood Studies Report (5 volumes). Natural Environment Research Council,
Wallingford, UK.
OPW 2004. Report of the Flood Policy Review Group. Office of Public Works, Oct 2004,
235pp.
O’Sullivan, J.J., Ahilan, S. and Bruen, M. 2012. A modified Muskingum routing approach
for floodplain flows: theory and practice. J. Hydrol., 470-471: 239–254.
Pandey, G.R. and Nguyen, V.T.V. 1999. A comparative study of regression based methods
in regional flood frequency analysis. J. Hydrol., 225: 92–101.
Payrastre, O., Gaume, E. and Andrieu, H. 2011. Usefulness of historical information for
flood frequency analyses: Developments based on a case study. Water Resour. Res., 47:
W08511, doi:10.1029/2010WR009812.
Pettitt, A. N. 1979. A non-parametric approach to the change point problem. Appl. Statist.,
28: 126–135.
Powell, R.W. 1943. A simple method of estimating flood frequency. Civil Eng., 13: 105–
107.
Reed, D.W. 1999. Deriving the flood frequency curve. Chapter 8, Volume 3, Flood
Estimation Handbook, CEH Wallingford, 46–51.
Reed, D.W. 2011. Letters in applied hydrology. DWRconsult, 86pp.
Reed, D.W., Jakob, D., Robson, A.J., Faulkner, D.S. and Stewart, E.J. 1999. Regional
frequency analysis: a new vocabulary. Proc. IAHS Symp. Hydrological extremes:
understanding, predicting, mitigating (eds Gottschalk, L., Olivry, J-C., Reed, D., Rosbjerg,
D.), Birmingham, July 1999, IAHS Publ. No. 255, 237–243.
Reed, D.W. and Robson, A.J. 1999. Adjusting for urbanisation. Chapter 18, Volume 3,
Flood Estimation Handbook, CEH Wallingford, 191–203.
Robinson, M. 1990. Impact of improved land drainage on river flows. Institute of
Hydrology Report 113, CEH Wallingford, 226pp.
http://www.ceh.ac.uk/products/publications/Impactofimprovedlanddrainageonriverflows.html
Robson, A.J. 1999a. Estimating QMED from flood data. Chapter 12, Volume 3, Flood
Estimation Handbook, CEH Wallingford: 77–99.
Volume II Flood Frequency Estimation
150
Robson, A.J. 1999b. Estimating QMED from catchment descriptors. Chapter 13, Volume 3,
Flood Estimation Handbook, CEH Wallingford: 100–127.
Robson, A.J. 1999c. Adjusting QMED for climatic variation. Chapter 20, Volume 3, Flood
Estimation Handbook, CEH Wallingford: 212–224.
Robson, A.J. and Jakob, D. 1999. L-moments for flood frequency analysis. Chapter 14,
Volume 3, Flood Estimation Handbook, CEH Wallingford: 129–138.
Siegel, S. 1956. Nonparametric statistics for the behavioral sciences. McGraw-Hill, 312pp.
Sneyers, R. 1990. On the statistical analysis of series of observations. Tech. Note 143,
WMO 415, World Meteorological Organization, Geneva, 192pp.
Srikanthan, R., McMahon, T.A. and Irish, J.L. 1983. Time series analysis of annual flows of
Australian rivers. J. Hydrol., 66: 213–226.
Stedinger, J.R. and Tasker, G.D. 1985. Regional hydrologic analysis, 1: Ordinary, weighted
and generalised least squares compared. Water Resour. Res., 21: 1421–1432.
Tasker, G.D. 1980. Hydrologic regression and weighted least squares. Water Resour. Res.,
16: 1107–1113.
Wallis, J.R., Matalas, N.C. and Slack, J.R. 1974. Just a moment! Water Resour. Res., 10:
211–219.
Weisberg, S. 1980. Applied linear regression. 1st ed., Wiley, New York, 283pp.
Volume II Flood Frequency Estimation
151
Appendices
Appendix A Review of stage-discharge relationships
A1 Terminology
The water level at a gauging station is generally known as the stage. The discharge is the
river flow. The stage-discharge relationship is commonly referred to as the rating curve or
simply the rating. All these terms are used interchangeably.
River flow is measured occasionally by field technicians. Historically, this was typically
undertaken using current meters to measure the flow velocity at different points across the
river cross-section. In more recent years, river flow is measured using Acoustic Doppler
Current Profilers (ADCPs). The resultant measured flows are typically known as gaugings.
The rating curve is constructed from the stage and discharge measurements. Because water
level can be recorded continuously, the rating curve allows river flow to be inferred
continuously.
Depths are generally in metres and flows in m3s
-1.
A2 Review of stage-discharge relationships
The review of stage-discharge relationships was necessary in order to provide a robust set of
annual maximum (AM) flows for the FSU research. Hydro-Logic Ltd reviewed all gauging
stations that had the potential for use within the FSU research. The work examined the stage-
discharge relationships held by the relevant hydrometric archiving authority, most notably the
OPW, the EPA and the ESB.
Rating curves are dynamic equations that require adjustment if there is a change in
instrumentation or a change in the channel control (e.g. due to debris/obstructions in the
channel, weed growth, tree growth etc.) Because of this, it was necessary to consider the full
history of the gauging station and to acquire all pre-existing stage-discharge relationships for
the gauging stations under consideration. These ratings were used as the starting point in the
review of stage-discharge relationships.
A2.1 General form of stage-discharge relationships
The stage-discharge relationships held by the OPW take the form:
pDG)(DCQ A.1
where:
Q is the discharge
D is the (gauged) water height or stage
DG is the datum shift (which can be positive or negative)
C is the discharge when the effective depth of flow (D + DG) is equal to 1
p is the slope of the rating curve (on logarithmic paper)
Volume II Flood Frequency Estimation
152
The C and p values sometimes derive from theory if the gauging station comprises a formal
structure. More typically, they derive from regression analysis of flow and depth gaugings.
Some hydrometric agencies use the alternate notation:
βahcQ A.2
where:
Q is the discharge
h is the (gauged) water height or stage
a is the gauge datum nominally corresponding to zero flow
c (constant) is the discharge when the effective depth of flow (i.e. h – a) is 1.0
(constant) is the slope of the rating curve (on logarithmic paper)
The stage-discharge relationship can be written in its equivalent logarithmic form. For
example, Equation A.2 can be transformed to:
ahlnβclnQln A.3
In the simplest case where the datum a is known – and where all gaugings are treated in a
single group – the parameters ℓn c and β can be obtained by linear regression analysis.
Appendix C provides an overview of regression methods.
A2.2 Analysis tools and background information
Analysis and evaluation tools in standard hydrometric software packages were used to review
any existing or proposed rating. Rating curves and gaugings were plotted using the Hydro-
Logic “Gauger Analysis” system (a system for rating development) or using the rating curve
editor utility “SKED” within the WISKI water management information system.
A key ingredient was to study the deviations between current-meter gauged flows (Qg) and
rating-curve calculated flows (Qc). These deviations were plotted both against stage and
against date. Such “deviation plots” assist in the identification and diagnosis of
discontinuities in the ratings.
Detailed reference was made to files of information about the station and its gauging history.
Metadata extracted included information on datum shifts, changes in flow control and other
technical matters. Contact was made with the hydrometric authority and, where appropriate,
with field technicians. This process helped to assess the validity of ratings and to judge
whether new work was needed to develop a better flood-flow rating.
A2.3 Review procedure
The following steps were undertaken in the review:
i A database was established and populated with readily available information. This
was refined and updated as the project progressed.
ii A prior estimate of the mean annual flood (Qmean) was generally available for each
gauging station. This was taken with the highest gauged flow (HGF) and the ratio
HGF/Qmean calculated. This was used in the initial classification of the gauging
station in terms of the likely quality of its high-flow measurements (see Section A3).
Volume II Flood Frequency Estimation
153
iii An exploratory analysis was undertaken by trawling through ratings classed as A, B
or C grade to gain a first impression of their quality and characteristics. Some sites
were identified for which some basic survey information could assist with the rating
classification. For example, simple surveys were arranged for a number of key sites
to establish the level of the river bank relative to the datum of the water level
recorder. Some further details are given in Section A2.4.
iv Periods of record for which a particular rating curve could be used were confirmed or
identified. Where arterial drainage works have affected the gauged site or its
catchment, the ratings and flow series were split into pre-drainage and post-drainage
elements.
v On completion of the exploratory analysis, a more detailed assessment of the ratings
began for stations initially graded A, B or C. During these more detailed
investigations it was ascertained whether the existing ratings could be used or whether
a refined or totally new rating was required. On the basis of this assessment, ratings
were selected for use in an uncertainty analysis (see Section A3.3). The final
recommended ratings were entered on the database.
vi Annual maximum (AM) water levels were provided by the OPW in Excel
spreadsheets. For the EPA sites, the required AM water levels were extracted from
the WISKI information management system, after filling or checking any gaps by
scrutiny of chart data.
vii AM flows were determined using the existing OPW/EPA rating. For sites where
there was scope to produce a better rating, AM flow series were also extracted using
the proposed new rating. If the differences between the two flow series were minor,
the existing OPW/EPA rating was generally adopted for use in the FSU. Differences
were normally considered acceptable if the QMED values calculated for the existing
and for the revised series agreed to within 10%. In other cases, a further review of the
ratings and station history files was undertaken to explore the differences.
viii Those ratings that were required to be changed for the purposes of the FSU research
were discussed with the OPW. If the proposed changes were deemed reasonable, the
revised ratings were adopted and used to produce the AM flow series for the FSU
research.
A2.4 Gauging station surveys
By examining the rating periods (see Step iv above), their respective gaugings and the
bankfull level it was possible in some cases to extrapolate the flood ratings above the HGF.
Initially, if the channel had a relatively regular cross-section and bankfull levels were known,
the bankfull levels could be used to set the limits of extrapolation.
For sites with a more complex stage-discharge relationship, a more detailed and thorough
topographic survey is required. However, it was not possible for the hydrometric authority to
undertake such detailed surveys within the timescale of the review.
Based on the exploratory analysis of ratings, a list of priority sites for bankfull-level surveys
was forwarded to the relevant OPW field technicians. This survey work was incorporated
into the routine visits to these sites. Within the timescale of the review, this was not possible
for EPA sites. Such survey information could be useful in further improvement and
refinement of the ratings.
Volume II Flood Frequency Estimation
154
For the sites identified by ESB for inclusion in the FSU, the bankfull levels were available
from previous reports undertaken for these sites.
A3 Gauging station classification
A3.1 Initial site classification
Prior to the review, the OPW had produced an initial classification of gauging stations. This
can be summarised as follows:
A sites Sites having stage-discharge ratings considered good for determining high and
flood flows.
B sites Sites with good high-flow ratings but for which there were some concerns as to the
quality of the flood-flow rating.
C sites Sites with reasonable medium to high-flow ratings but for which it was not
possible to determine flood flows with any confidence, due to the fact that at high
flows the site was either not rateable (e.g. due to hydraulic effects or bypassing) or
because there were insufficient gaugings to produce a rating.
P sites Sites classified as poor and not considered suitable for high and flood-flow
determinations. It is possible that some of these sites could be used in future if
sufficient gaugings and other information were available.
U sites Sites for which the data were unusable for determining high flows. These could
for example be “level-only” stations at which it is not possible to measure
discharges.
P and U sites were not considered in any detail during the project.
For sites archived by the EPA, the classification of ratings had been produced some years
earlier and was not complete for all sites. Values of the mean annual flood (Qmean) were
derived from the EPA's archive database for all sites considered for inclusion in the FSU.
The ESB also has a network of hydrometric gauging stations with an associated archive
dataset. Even though their network is small, the sites have extensive records going back for
many years and in some cases for up to 60 years. After exploration, six of the ESB stations
were identified for potential use in the FSU research.
From the list of sites produced, a simple ranking index was created, based on the ratio
HGF/Qmean where HGF denotes the highest gauged flow.
A3.2 FSU station classification
The basis for the first level of site categorisation for the FSU was as follows:
Grade A station – Suitable for flood frequency analysis: Site for which HGF is at
least 1.3 QMED and which has a stage-discharge relationship that the OPW judges
represents extreme floods reasonably.
Volume II Flood Frequency Estimation
155
Grade B station – Suitable for flows up to QMED: Site for which HGF is at least
0.95 QMED and for which the cross-section is thought to be without significant
change in channel geometry up to the stage corresponding to QMED.
Grade C station – Possibly suitable for extrapolation up to QMED: Site for
which the rating is considered well-defined up to about 0.8 QMED.
The categorisation was further refined to distinguish two classes of Grade A station:
Grade A1 station – Very good – Providing the very best flood data: Grade A
station with confirmed ratings that are considered good for flood flows well above
QMED and with some confidence of valid extrapolation up to 2.0 QMED using
suitable survey data and allowing for any flows across the floodplain.
Grade A2 station – Good – Providing the next best flood data: Other Grade A
stations.
A3.3 Uncertainty analysis
The FSU categorisation makes particular use of the ratio of HGF to QMED. Being defined
as the median of the AM flows, QMED is a relatively robust measure that is unaffected by
uncertainty in the measurement of the very highest flood flows. The HGF is taken as the
largest flow that is measured at the station with a fair degree of confidence. The ratio
HGF/QMED therefore provides a basic index of the likely quality of the flood-flow rating.
If, for example, HGF is 85 m3s
-1 and QMED is 53 m
3s
-1, HGF/QMED is 1.6. This suggests
that the rating relationship is potentially valid for relatively rare floods well in excess of
QMED. Unless there were specific doubts about the quality of the high-flow gaugings, the
station would be assigned to Grade A1.
The above classification system inevitably has a certain amount of subjectivity. Following
discussion with the OPW, a more analytical approach was explored in which the likely
uncertainty in the high-flow ratings was assessed from the scatter (in gaugings about the
stage-discharge relationship) at high flows, taking due account of the number of high-flow
gaugings used in its construction.
The uncertainty analysis constructed 95% confidence intervals about the linear regression
(e.g. Equation A.3) representing the stage-discharge relationship. This was done using the
Hydro-Logic Gauger Analysis software, in which it is possible to enter the stage value
corresponding to QMED, or any other stage value, and obtain 95% confidence limits on the
estimation of flow using the stage-discharge relationship.
During the review and analysis of the ratings, stage values corresponding to QMED for all
Grade A and B stations were entered and confidence limits obtained. This provided a
quantitative expression of the uncertainty in flow measurement using the stage-discharge
relationship. Stations were considered good (i.e. worthy of Grade A2) if 95% confidence
intervals for flows at the level of QMED were smaller than 30% and very good (i.e. worthy of
Grade A1) if they were smaller than 10%. This uncertainty analysis helped to confirm the
final classification of gauging stations.
Volume II Flood Frequency Estimation
156
A4 Production of annual maximum flood series
With the gauging station review complete, and in association with the EPA and Hydro-Logic
Ltd, OPW hydrometric staff prepared the AM flow datasets subsequently used in the research
reported in this volume. In the majority of cases, flow data supplied for use in the Volume III
research on hydrographs were also based on the rating curves confirmed or revised in the
above review of stage-discharge relationships.
The FSU research chiefly used flow data from Grade A1, Grade A2 and Grade B stations.
Data from other gauging stations should be considered in site-specific flood studies.
Volume II Flood Frequency Estimation
157
Appendix B Flood data exclusions
B1 Stations omitted from the QMED modelling research
Sixteen out of 206 stations were omitted from the QMED modelling research of Chapter 2 for
the reasons highlighted below:
15003
Station 15003 Dinin at Dinin Bridge was one of three stations questioned during model-
building because of their exaggerated influence on model coefficients. Further inspection
revealed a number of low outliers, i.e. exceptionally small annual maxima. This station is
known to be an extremely flashy catchment in a karst area (Castlecomber Plateau). QMED
may not provide an adequate index flood in these circumstances. [Editorial note: Quite
extensive use was made of this Grade A2 station in other chapters, most notably in the study
of flood volumes reported in Section 5.4 and Appendix J.]
19014, 19015, 19016 and 19031
Annual maximum flow series for Stations 19014, 19015, 19016 and 19031 in the Lee basin
arrived too late for inclusion in the Chapter 2 research. [Editorial note: Limited use was
made of Stations 19014, 19016 and 19031 in the research reported in other parts of Volume
II. These ESB stations are categorised Grade B.]
20006
Station 20006 Argideen at Clonakilty WTW was one of three stations questioned during
model-building because of their exaggerated influence on model coefficients. Further
inspection revealed criteria for omission. The station has an unusually small QMED.
Investigation of the flow series revealed a large number of years with missing months. This
raises the suspicion that the true AM flow may have been missed in a number of years.
[Editorial note: Some limited use of this Grade B station was made in other parts of Volume
II.]
25001, 25002, 25003 and 25005
Four stations in the Mulkear basin were omitted from analysis in line with recommendations
made by Joyce (2006, pers. comm.):
This river was subjected to a District Drainage Scheme in the late 1920s and early
1930s that protects large areas of land by extensive lengths of embankments that are
overtopped about once in five years. This means that the catchment responds almost
without storage attenuation for the smaller annual maxima, including QMED, and
with massive storage attenuation for the larger events.
[Editorial note: These stations were omitted from much of the research reported in this
volume. However, they are included in the Section 5.5 study of flood seasonality and in the
trend analyses reported in Chapter 3 and Appendix E.]
Volume II Flood Frequency Estimation
158
26010
The AM series for Station 26010 Cloone at Riverstown revealed suspicious outliers. On
investigation, it was found that a partially developed flood rating had been erroneously
applied. Neither the AM flow series based on the OPW rating, nor that based on the
Hydro-Logic Ltd rating, appeared to reflect changes in the stage-discharge relationship
adequately. [Editorial note: Some limited use was made of this Grade B station in other
parts of Volume II.]
30037
Station 30037 Robe at Clooncormick was one of three stations questioned during model-
building because of their exaggerated influence on model coefficients. The gauged QMED of
1.79 m3s
-1 is exceptionally small for a catchment area of 210 km
2. This is well below any
other catchment of a similar size. [Editorial note: Some limited use was made of this
Grade B station in other parts of Volume II.]
31075, 34005 and 36020
Discrepancies were found in the metadata descriptions of Stations 31075, 34005 and 36020.
QMED values were provided but without the AM values themselves. [Editorial note: None
of these stations was used in Volume II.]
36027
The AM series for Station 36027 Ballyconnel Canal East at Bellaheady revealed some
extremely low outliers. Given the station’s name, and the fact that it typically records an AM
flood of only 25 m3s
-1 from a nominal drainage area of 1501 km
2, it seems unlikely that this
represents a meaningful flood series for a natural river. [Editorial note: The only use of this
Grade A2 station in Volume II was in the Section 5.5 study of flood seasonality.]
B2 Other notes on stations omitted from research
The list in Section B1 is not comprehensive. Different researchers make different choices.
Part of the Volume II research considered data for Station 07004 (Kells) Blackwater at
Stramatt. However, for some unrecorded reason, the Chapter 2 research on QMED
modelling did not consider this Grade A2 station.
27070
The annual maximum series for Station 27070 Lough Inchiquinn at Baunkyle was excluded
from most parts of Volume II because of suspected extreme lake effects. Its retention in the
QMED modelling of Chapter 2 does not appear to have led to undue difficulty or influence.
B3 General recommendation
Practitioners are encouraged to study all relevant flood data carefully. A flood series rejected
for national study may include information that is valuable in a site-specific study.
Investigation or modelling may rehabilitate flood data considered suspect or unreliable.
Volume II Flood Frequency Estimation
159
Appendix C Notes on the regression methods used in Chapter 2
C1 Ordinary least-squares (OLS) regression
The catchment descriptor equation was fitted using multiple linear least-squares regression.
Under this approach, the relationship can be written in vector notation as:
= + y X β e C.1
where is the vector of dependent variables, is the matrix of independent variables, is the y X β
vector of regression coefficients and is the vector of random errors. The errors are assumed e
to be uncorrelated and N(0, γ2), meaning that they are Normally distributed with mean of zero
and a variance of γ2. This is referred to as the model error variance.
Grover et al. (2002) highlight that in hydrology the true value of (e.g. the vector holding y
QMED values observed at a set of sites) is typically unknown, and there is therefore an error
associated with its estimation. Adopting widely used notation (e.g. Stedinger and Tasker,
1985), if y is an unbiased estimate of the variable of interest then:
E[ ] = ŷ y C.2
and
Var[ ] = ŷ Σ C.3
where is the sampling covariance matrix associated with the estimate of . Therefore Σ ŷ
Equation C.1 is written:
= + y X β u C.4
where is a random vector of errors that are a combination of model and sampling errors u
defined as:
Var[ ] = = u Λ γ2 I + Σ C.5
where is defined as the full covariance residual matrix, is a vector of modelling errors, Λ γ2
I
is the identity matrix and is a matrix of sampling errors. The least-squares estimate for in Σ β
Equation C.4 is given by:
β = (T -1
)-1
T -1
X Λ X X Λ ŷ C.6
This is also known as the generalised least-squares estimator.
In this study, three least-squares methods – ordinary, weighted and generalised – were
applied to solve Equation C.6, thereby estimating the model parameters . Ordinary least-β
squares (OLS) is the simplest method and is suitable when the sampling error in the data is
small ( ≈ 0) and the error terms have equal variances and are uncorrelated. Σ
C2 Weighted least-squares (WLS) regression
The weighted least-squares (WLS) procedure for hydrologic regression introduced by Tasker
(1980) attempts to limit the uncertainty introduced by unequal record lengths. [Unequal
record lengths tend to lead to heteroscedacity, i.e. to observations that have unequal sampling
errors. For example, one will generally have a much better estimate of QMED (the median
annual maximum flood) where the median is taken from 30 rather than (say) eight annual
Volume II Flood Frequency Estimation
160
maxima. In consequence, the assumption of constant variance in OLS is no longer valid.] In
the WLS approach, a weighting term – proportional to the square root of the record length –
is used to represent the sampling error. The procedure used here follows Weisberg (1980).
C3 Generalised least-squares (GLS) regression
The generalised least-squares (GLS) procedure – introduced to hydrological application by
Stedinger and Tasker (1985) – is an extension of WLS which seeks to account for inter-site
correlation in flood data.
In applying the GLS procedure, it was assumed here that the inter-site correlations in AM
flow data provide a reasonable approximation to the correlations in the regression errors.
Inter-site correlation was assessed and represented by an exponential decay with distance,
with inter-site correlation falling to approximately 0.5 at a distance of 50 km (see Figure C.1).
An exponential spatial correlation was incorporated into the GLS approach. While it is
evident from Figure C.1 that this is an approximation only, its incorporation into the
estimation procedure provides some recognition of the effect of inter-site correlation.
Figure C.1: Fitted model for inter-site correlation
[Editorial note: The Figure C.1 plot of inter-station correlation in AM flows against inter-
catchment distance is a little puzzling. The perfect correlation of a small number of station-
pairs is explained by records that overlap in one year only. However, one would expect some
station-pairs – especially those having only a few years of record in common – to exhibit
negative correlation. It seems likely that this is a mistake in the plot rather than in the
analysis. The decay rate of the fitted exponential is a little slower than that reported by
Robson (1999b) for UK flood data.]
From the modelling conducted, it was found that the assumptions of the OLS approach – i.e.
Normally distributed residuals, equal variance and uncorrelated sampling errors in the data –
were broadly satisfied. The simpler approach was therefore adopted. The extension of the
methodology to WLS and GLS returned only very slight changes in model performance and
parameter values. The results are therefore not reported further.
2001751501251007550250
Distance (km)
1.00
0.90
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
Co
rrela
tio
n
0 50 100 150 200
Inter-centroid distance, d (km)
Corr
elat
ion,
r
1.0
0.8
0.6
0.4
0.2
0.0
r = exp(-0.0137 d)
Volume II Flood Frequency Estimation
161
C4 Geographically-weighted regression (GWR)
A few details of geographically weighted regression (GWR) are given. A global regression
model can be presented as:
εxvμ,β....xvμ,βvμ,βy nn110 C.7
where (μ, ν) denotes the coordinates of the samples in space. In GWR, the parameter
estimates are made using an approach in which the contribution of a sample to the analysis is
weighted based on its spatial proximity to the specific location under consideration. Thus,
the weighting of an observation is no longer constant in the calibration but varies with
location. Data from observations close to the location under consideration are weighted more
than data from observations far away.
The parameters are estimated from a variation of Equation C.6 (in Section C1) in which the
covariance matrix -1
is replaced by a special matrix of weights: Λ
β (μ, ν) = (T
(μ, ν) )-1
T
(μ, ν) X W X X W ŷ C.8
Here, β (μ, ν) is an estimator of , is the matrix of independent variables, and (μ, ν) is the β X W
weighting matrix. β is the vector of regression coefficients.
Several methods have been proposed to determine the weighting matrix. Let Wij denote the
weight of the specific point j in the space at which data are observed to any point i in the
space from which parameters are estimated. Under certain assumptions it is reasonable to
represent Wij – known as the kernel – by the continuous function:
2
/bdexpW
2
ij
ij C.9
where dij denotes the distance between i and j, and b is referred to as the bandwidth. An
alternative kernel utilises the bi-square or quartic function:
22
ijij /bd1W if dij < b C.10
0Wij otherwise
Fixed kernels in regions where data are dense may suffer from bias when the kernels are
larger than needed. When the kernels are smaller than needed, they may not estimate the
parameters reliably where data are scarce. Thus, spatially varying kernels have also been
proposed.
Parameter estimation in GWR is highly dependent on the weighting function and the
bandwidth of the kernel used. As the bandwidth increases, the parameter estimates will tend
to the estimate from a global model. The selection of the weighting function and bandwidth
can be determined using a cross-validation approach. In this work, GWR was deployed using
an adaptive bi-square kernel, and the selection of weighting functions and bandwidth was
based on cross-validation.
Volume II Flood Frequency Estimation
162
Table C.1 shows the results of the tests for spatial stability in model parameter coefficients
for each independent variable in the rural ℓnQMED model of Equation 2.7. It is evident that
the majority of parameter coefficients in the model are indeed spatially constant. The
exception is the coefficient of ℓnFARL. [Editorial note: The use of p-values is discussed in
Section E3.4.]
Table C.1: Test of significance of spatial variability in ℓnQMED model parameters
Parameter p-value Significance Parameter p-value Significance
Intercept 0.14
None
ℓnSAAR 0.13 None
ℓnAREA 0.84 ℓnFARL 0.05 Just significant
ℓnDRAIND 0.16 ℓnBFIsoil 0.36 None
ℓnS1085 0.24 ℓn(1+ARTDRAIN2) 0.86
Map C.1 suggests that the coefficient of ℓnFARL takes generally higher values in the East
and North West and lower values around the Upper Shannon basin and in the West.
Comparisons with Map 2.2 andMap 2.3 suggest that the mapped minimum just west of
Wexford is driven by results from a single Grade B site: Station 13002 Corock at Foulk’s
Mill. [Editorial note: The FARL value for this catchment is 1.0, implying no major
reservoir or lake on the river network. Thus, it appears that the isolated overestimation of
QMED (evident in Map 2.2) is leading to the minimum west of Wexford, rather than any
feature related to FARL. This illustrates that automated methods can lead to over-fitting of
observations and can bypass tests of physical reasoning. What is really needed is
intensification of the gauging network. It is understandable why the GWR approach was not
selected for implementation.]
Volume II Flood Frequency Estimation
163
Map C.1: Spatial variation in the FARL coefficient as interpolated from GWR
Coefficient of
ℓnFARL
-1.514 – -0.552
-0.552 – 0.411
0.411 – 1.374
1.374 – 2.337
2.337 – 3.299
3.299 – 2.337
1.374 – 4.262
4.262 – 5.225
5.225 – 6.188
6.188 – 7.151
Wexford
Volume II Flood Frequency Estimation
164
Appendix D QMED models for drained/undrained catchments
D1 Partitioned approach
A more detailed approach to modelling QMED was explored by deriving separate models for
drained and undrained catchments. The calibration set of 145 essentially rural catchments
comprises 95 undrained (i.e. without arterial drainage) and 50 drained (i.e. with arterial
drainage). Some of the undrained cases are pre-drainage records from catchments
subsequently drained.
D2 Undrained catchments
Following the methodology used to derive the rural model of Section 2.2, QMED values for
the 95 undrained stations were subject to an exhaustive search to derive the best combination
of catchment descriptors for model-building. Details of the 6-variable model resulting are
given in Table D.1. This is similar in make-up to the general rural model of Equation 2.7
except that the arterial drainage term is redundant. The similarity of the results (to those of
Table 2.7) increases confidence in the manner in which ℓn(1+ARTDRAIN2) indexes
drainage and in the meaningfulness of the other six variables.
Table D.1: Coefficient and collinearity statistics for undrained ℓnQMED model
Term/regressor Coefficient Standard
error
β
value
t
statistic
95% confidence
interval Variance
inflation
factor (VIF) Lower Upper
Constant -11.145 1.44 -7.74 -14.01 -8.28
ℓnAREA 0.910 0.04 1.05 21.34 0.82 1.00 2.00
ℓnSAAR 1.328 0.22 0.35 6.03 0.89 1.77 2.79
ℓnFARL 2.762 0.41 0.31 6.78 1.95 3.57 1.71
ℓnS1085 0.214 0.05 0.24 4.07 0.11 0.32 2.76
ℓnDRAIND 0.477 0.12 0.18 4.03 0.24 0.71 1.69
ℓnBFIsoil -0.590 0.21 -0.13 -2.76 -1.02 -0.16 1.87
Nevertheless, there are substantial differences in some of the model coefficients. Of
particular note is the reduced importance of ℓnBFIsoil, indicated by the more modest
coefficient and β value in Table D.1. The coefficients of ℓnDRAIND (especially), ℓnFARL
and ℓnS1085 all increase relative to the general rural model (i.e. compared to the
corresponding values in Table 2.7).
Overall, the model provides good results with an r2
of 0.892 and a factorial standard error
(FSE) of e0.315
= 1.370. From the model diagnostics, the OLS approach is again judged
acceptable, with the logged residuals Normally distributed and showing little evidence of
heteroscedacity (i.e. uneven error variance).
Volume II Flood Frequency Estimation
165
Figure D.1 summarises the performance achieved for the undrained stations. Using the
subscript ud to denote undrained, the OLS model is:
ℓnQMEDud = –11.145 + 0.910 ℓnAREA + 1.328 ℓnSAAR + 2.762 ℓnFARL +
0.214 ℓnS1085 + 0.477 ℓnDRAIND – 0.590 ℓnBFIsoil D.1
50020010050201052
500
200
100
50
20
10
5
2
Observed QMED (m s )
Pre
dic
ted Q
ME
D
1:1 line
3 -1
Station 06030 River Big at Ballygoly
Figure D.1: Performance achieved with undrained model (on 95 undrained catchments)
D3 Drained catchments
In order to fit a regression model to the drained catchments (i.e. catchments in their post-
drainage state), the exhaustive regression approach was again employed. Using the subscript
d to denote drained, the linear form of the selected 4-variable model is:
ℓnQMEDd = –11.214 + 0.976 ℓnAREA – 1.780 ℓnBFIsoil + 1.230 ℓnSAAR +
1.328 ℓnFARL D.2
Further details of the model and its performance are given in Table D.2 and Figure D.2.
Overall, the model performs well with an FSE of e0.318
= 1.374.
It is possible that the excellent degree of fit achieved (r2=0.936) reflects that catchments that
have undergone drainage tend to be of a similar ilk, so there is less variance for a model to
have to capture. The variance inflation factors in Table D.2 confirm the relative robustness
of the model.
Nevertheless, there are some interesting features. The coefficient of ℓnBFIsoil is about three
times stronger than in the undrained model, and the variable is second only to ℓnAREA in
Volume II Flood Frequency Estimation
166
helping to explain the variation in ℓnQMED. This suggests the particular usefulness of
BFIsoil in characterising the behaviour of drained catchments.
Table D.2: Coefficient and collinearity statistics for drained ℓnQMED model
Term/regressor Coefficient Standard
error
β
value
t
statistic
95% confidence
interval Variance
inflation
factor (VIF) Lower Upper
Constant -11.213 2.09 -5.37 -15.42 -7.01
ℓnAREA 0.976 0.04 0.96 22.69 0.89 1.06 1.27
ℓnBFIsoil -1.780 0.28 -0.28 -6.30 -2.35 -1.21 1.41
ℓnSAAR 1.230 0.31 0.16 3.98 0.61 1.85 1.17
ℓnFARL 1.328 0.61 0.11 2.18 0.10 2.56 1.67
500200100502010521
500
200
100
50
20
10
5
2
1
Observed QMED (m s )
QM
ED
pre
dic
ted b
y d
rain
ed m
odel 1:1 line
Station 24022
25017
3 -1
Figure D.2: Performance achieved with drained model (on 50 drained catchments)
D4 Choosing a general purpose model
The partitioned approach has derived separate models for predicting ℓnQMED on drained
and undrained catchments. To test the merit of the approach, its performance was assessed
on the 25 stations in the validation dataset, and against the performance achieved by the
general model of Section 2.2.
The validation set comprises 17 undrained stations and eight drained stations. Table D.3
compares the success of the partitioned and general models in explaining variation in
ℓnQMED across the validation stations.
Volume II Flood Frequency Estimation
167
The drained-catchment model of Equation D.2 is seen to provide only a modest improvement
over the general model when tested on the eight drained catchments in the validation set.
More notably, the undrained-catchment model of Equation D.1 is outperformed by the
general model when tested on the 17 undrained catchments in the validation set. This is
likely due to the fact that all six PCDs appearing in the partitioned model are no different to
those already in the general 7-variable model for ℓnQMED. The general model includes
ARTDRAIN2.
Table D.3: Validation of partitioned and general models for ℓnQMED
Validation stations r
2 by partitioned model r
2 by general (Section 2.2)
model Undrained Drained
17 undrained 0.898 0.915
8 drained 0.891 0.848
All 25 0.906
Given the mixed success of the partitioned approach in validation – and the fact that it has
fitted 12 parameters (seven for the undrained and five for the drained) as against eight
parameters for the general model – it was judged prudent to maintain the recommendation in
Section 2.2.5 to adopt the 7-variable general model to estimate the rural component of
QMED on all catchments: drained and undrained.
Nevertheless, the partitioned analysis has shown BFIsoil and ARTDRAIN2 to be particularly
important in characterising the post-drainage flood response of a catchment, and the
descriptors DRAIND and S1085 to be relatively important when estimating QMED on
undrained catchments. These findings may help the experienced user to judge the most
appropriate data transfer when estimating QMED at an ungauged subject site (see
Section 11.2).
Volume II Flood Frequency Estimation
168
Appendix E Trend analysis of AM flows in Irish rivers
E1 Introduction
A major assumption in flood frequency analysis is that all observations in the dataset are
independent (random) and identically distributed. The study reported here is by Mandal and
examines the validity of the independent and identically distributed (iid) assumption for
annual maximum (AM) flows in Irish rivers. See also Mandal (2011).
Randomness cannot be proved but it can be disproved by the presence of a non-random
feature such as trend. Statistical methods lie at the heart of testing for change in hydrological
data series. Before the tests are described, the purpose, importance and language of trend
analysis are first introduced.
E2 Purpose
The purpose of trend testing is to determine if the values of a variable generally increase (or
decrease) over some period of time (Helsel and Hirsch, 1992; Hirsch et al. 1993).
E2.1 Importance of testing whether hydrological processes are stationary
Detection of abrupt or gradual changes in hydrological records is of scientific and practical
importance, and fundamental to planning water resources and managing flood risks
effectively. Traditional “design rules” (e.g. for water resource or flood alleviation systems)
are based on the assumption that hydrological processes are stationary (see Box E.1) and the
principle that the past is the key to the future.
Box E.1: Stationarity
The assumption of stationarity is clearly questionable in the era of global change. If it is
incorrect, procedures for water-related structures such as dams and river embankments will
need to be reassessed to ensure that risks are properly appreciated. Systems might be under-
designed and therefore fail to perform adequately. On the other hand, if systems are over-
designed, they may not be cost-effective.
E2.2 Types of change
Changes in river flows can be caused directly by human activities such as urbanisation,
reservoir construction, drainage works, water abstraction and agricultural changes, or by
incidental factors such as changes in channel morphology. An apparent change in flood
behaviour may arise from measurement practices, e.g. a change in instrumentation, a change
in data processing or the failure to apply the flood rating curve appropriate to the particular
period of record.
A stochastic process is said to be strictly stationary if its probability distribution is
completely independent of time. It is said to be stationary in the wider sense if the mean
and variance are independent of time. The statistics of samples drawn from a stationary
process may vary due to sampling variability but not due to their size or their position in
the population.
Volume II Flood Frequency Estimation
169
Climate is the principal driver of the hydrological cycle. Since the climate system and water
cycle are intimately linked, any change in one induces change in the other. Widespread
effects therefore arise from climate variability and from human-induced climate change.
Change in a series can manifest itself in a number of ways. Attention typically focuses on
gradual progressive change – referred to as trend – and abrupt change – often referred to as
step-change. Cyclical behaviour makes for a third kind of non-stationarity. For example,
several major floods may occur in a so-called flood-rich period only to be succeeded by a
flood-poor period in which no major flood occurs for many years.
Attention usually focuses on testing for trend or step-change in the typical value of a data
series: for example, a change in the mean or median. Non-stationarity may also be revealed
in a change in variability or in autocorrelation, or in almost any aspect of data.
With a range of effects to be considered, exploratory data analysis and graphical display of
data series are essential.
E3 Procedure
E3.1 Steps in the analysis
Kundzewicz and Robson (2004) list the main stages in a statistical analysis of change:
Decide what type of data series (e.g. monthly averages, annual maxima, peaks-over-
threshold series, …) to test to meet the issue of interest;
Decide what types of change are of interest (e.g. gradual or step-change);
Check the assumptions in an exploratory data analysis or by applying a formal test;
Select the statistical methods:
A particular test (noting that it is good practice to use more than one);
A particular test statistic;
A method for evaluating significance levels;
Evaluate the significance levels;
Investigate and interpret the results.
E3.2 The idea of exploratory data analysis
Exploratory data analysis (EDA) is an advanced visual examination of the data and forms an
integral part of any study of change. Graphical presentation is pivotal. An advantage of a
large study such as that reported here is the ability to assess behaviour at many sites. Any
effect arising from climate is likely to influence more than one flood series.
The first stage of EDA is to examine the raw data to identify notable features such as:
Peculiarities or problems (e.g. unusually large/small values or gaps in the record);
Temporal patterns (e.g. trend, step-change or marked seasonality);
Regional/spatial patterns.
Exploratory data analysis also plays an important role in checking assumptions such as
independence or the presumed statistical distribution of data values. Types of graph that can
Volume II Flood Frequency Estimation
170
be useful for checking hydrological data series include histograms, probability plots,
autocorrelation plots, scatter-plots and smoothing curves (see Box E.2).
Box E.2: Smoothing
A well-conducted EDA is sometimes such a powerful tool that it eliminates the need for a
formal statistical analysis (Kundzewicz and Robson, 2004).
E3.3 Hypothesis testing
A statistical test requires a formal procedure that includes:
Stating a null hypothesis (the convention is to label this H0);
Declaring a test statistic and its distribution under the null hypothesis;
Stating a critical region for the test statistic in which – under the null hypothesis – the
value of the test statistic falls with probability α;
Computing the test statistic for the data sample;
Accepting/rejecting the null hypothesis according to whether the observed test
statistic value lies inside/outside the critical region.
When interpreting results it is necessary to remember that no statistical test is perfect, even if
all test assumptions are met. Adoption of a 5% significance level means that an error will be
made, on average, 5% of the time. [If the 5% significance level is adopted and the null
hypothesis is actually true, about one in 20 test results will (as a matter of chance) yield a test
statistic greater than the critical value. In consequence, the null hypothesis will be (wrongly)
rejected.]
It is relatively commonplace to term a result significant if the null hypothesis is rejected at the
5% level and to denote it highly significant if it is rejected at the 1% level. However, there is
no universal acceptance of terminology and it is widely recognised that the choice of significance level, α, is arbitrary.
[Editorial note: A statistical test determines the statistical significance of a result. This is
not to be confused with the practical significance of a feature. For example, data may reveal
a trend that is very highly significant statistically. Yet the scale of the trend may be too
minor to be of practical importance.] Where many tests are applied to a large number of data samples, there is considerable scope
for significant results to arise by chance. In consequence, the interpretation of results can
become complex. Even the presence of a highly significant result may provide only weak
evidence of change. However, if a batch or “basket” of significant results is obtained – e.g. if
Smoothing techniques are used to reduce irregularities (random fluctuations) in time series
data. They can provide a clearer view of the underlying behaviour of a series.
In time series where it is naturally strong, seasonal variation can impede the detection of
trends or cycles. Smoothing can remove seasonality and make long-term fluctuations in
the series stand out more clearly. A typical smoother is a moving average filter in which
each observation is replaced by the average of observations in a time window centred on
the time of the particular observation.
Volume II Flood Frequency Estimation
171
the null hypothesis is rejected by several tests or if a particular test rejects the null hypothesis
at a number of sites on a particular river or in a particular district – there can be greater
confidence that an important effect has been detected.
E3.4 Use of p-values
An alternative way to conclude a test of hypotheses is to compare the p-value of the sample
test statistic against the significance level (α). The p-value of the sample test statistic is the
smallest level of significance for which we can reject H0. In other words, the p-value of a
statistical hypothesis test is the probability of getting a value of the test statistic as extreme as
or more extreme than that observed by chance alone, if the null hypothesis H0 is true.
The p-value is compared with the actual significance level of our test and, if it is smaller, the
result is deemed significant. That is, if the null hypothesis were to be rejected at the 5%
significance level, this would be reported as “p < 0.05”. The smaller the p-value, the more
convincing is the rejection of the null hypothesis. It indicates the strength of evidence for
rejecting the null hypothesis H0, rather than merely concluding “Reject H0” or “Do not reject
H0”.
The p-value serves a valuable purpose in the evaluation and interpretation of research
findings. It enables the researcher to set their own level of significance and to reject or accept
the null hypothesis in accordance with their own criterion rather than that at a fixed level of
significance.
E4 Statistical tests
E4.1 Parametric and non-parametric tests
Various test procedures have been developed for detecting trend in hydrological time series.
An important classification is whether the test is parametric or non-parametric. A parametric
test is one that depends on the form of parent distribution from which the sample is assumed
to be drawn. Such tests are able to make full use of the available information but carry the
penalty that the distributional assumption may be incorrect. This may bias the results
obtained.
In contrast, non-parametric tests require few if any assumptions about the shape (e.g.
skewness) of the underlying population distributions. Consequently, they are also known as
distribution-free methods. Typically, non-parametric tests use ordinal information (often, the
ranks of the data values) rather than the values themselves.
Inevitably, non-parametric tests are less powerful than parametric tests. Thus, when the
relevant distributional assumption is met, it is preferable to use the parametric test.
Hydrological data are often strongly non-Normal. This means that tests which assume an
underlying Normal distribution are inadequate. In many real-life situations, distribution-free
methods are to be recommended because they require only minimal assumptions to be made
about the data series.
The Flood Studies Report (NERC, 1975) found that the EV1 distribution could adequately
describe Irish AM flow series. The use of tests based on the Normal distribution was
therefore not considered to be appropriate. Non-parametric tests are used in order to test the
Volume II Flood Frequency Estimation
172
underlying assumptions of independent and identically distributed (iid) time series.
However, in order to compare the power of the tests, two parametric tests are also used.
E4.2 Null hypotheses
The null hypothesis H0 for the tests for trend is that there is no trend in the data. The null
hypothesis H0 for the tests for changes or difference in means/medians is that there are no
changes or difference in the means/medians between two data periods. The null hypothesis
H0 for the tests for randomness (independence) is that the data come from a random process.
It is considered that the alternative hypotheses for all tests are non-directional i.e. all tests are
two-tailed tests.
E4.3 Tests adopted
Six tests were used in detecting trends, shift and serial dependency in the AM flood series:
Tests for trend
Mann-Kendall (non-parametric test for trend)
Spearman’s Rho (non-parametric test for trend)
Mean-weighted linear regression (parametric test for trend)
Test for step-change
Mann-Whitney U (non-parametric test for step-change in mean/median)
Tests for serial dependency of time series
Turning point (non-parametric test for randomness)
Rank difference (non-parametric test for randomness)
E4.4 Mann-Kendall test
The Mann-Kendall test is a non-parametric test for identifying trends in time-series data. The
data need not conform to any particular distribution.
The test compares the rankings of the sample data rather than the data values themselves.
The n time-series values {X1, X2, X3, …, Xn} are replaced by their relative ranks {R1, R2, R3,
…, Rn} where R1 marks the lowest value and Rn the highest. Where equal-ranking (i.e. tied)
values are found, each is assigned a mean rank. Thus, if the 4th
, 5th
and 6th
ranked values are
equal, each is assigned a rank of 5.0. See Box E.3 for a discussion of tied values.
Box E.3: Tied values
The Mann-Kendall test statistic S (known as Kendall’s sum) is given by:
1
1 1
sgnn
i
n
ij
ji RRS E.1
where sgn is the sign function defined by:
Editorial note: Exactly equal values are known as tied values or ties. For various
reasons – notably the rounding of water-level observations to a fixed number of decimal
places – ties are quite common in flood peak data. It can be important for methods to
explicitly recognise this feature, if best estimates are to be obtained.
Volume II Flood Frequency Estimation
173
sgn(x) = 1 for x > 0
sgn(x) = 0 for x = 0 E.2
sgn(x) = -1 for x < 0
A positive value of S indicates a possible upward trend; a negative value indicates a possible
downward trend.
Mann (1945) and Kendall (1975) document that if the null hypothesis H0 is true (i.e. if there
is no trend) then for n ≥ 8, the statistic S is approximately Normally distributed with mean
zero and variance:
18
5)1)(2ii(it5)1)(2nn(n
(S)Var
n
1i
i
E.3
where ti is the number of ties of extent i.
The standardised Mann-Kendall (MK) statistic, ZMK, is computed by:
SVar
1SZMK
for S > 0
0ZMK for S = 0 E.4
SVar
1SZMK
for S < 0
ZMK follows the N(0, 1), i.e. the standard Normal distribution with a mean of zero and a
variance of one.
The p-value of the MK statistic is found from the cumulative distribution function (CDF) of
the Normal distribution using:
ZΦ0.5p where dte2π
1ZΦ
Z
0
2t2
E.5
When the p-value is small enough, the trend is considered to be statistically significant. For
example, if p < 0.05 the trend is considered significant (if a 5% significance level is being
applied).
E4.5 Spearman’s ρ test
Spearman’s ρ (rho) is a rank-based test that determines whether the correlation between two
variables is significant. When applied to trend analysis of annual maxima, one variable is
taken as the year number and the other as the annual maximum value. Both variables are
replaced by their ranks. [Editorial note: This substitution is less suitable if there is a period
of record missing in the AM series.]
If the time series consists of n distinct values, the ranks will be the numbers 1 to n, with 1
corresponding to the lowest value in the series and n to the highest. If there are ties (see Box
E.3) in the series, each value in the tie group is assigned the same (mean) rank. Thus the
ranks can sometimes be non-integer.
Volume II Flood Frequency Estimation
174
The null hypothesis H0 of Spearman’s ρ test is that there is no relationship between the
variables. Given a time series of values {Xi, i=1, 2, … n}, rejection of the null hypothesis
implies that Xi increases or decreases with i (i.e. that a trend exists with time). The relevant
test statistic given by Sneyers (1990) is:
1nn
i-R6
1D2
n
1i
2
i
E.6
where Ri is the rank of the ith
observation, Xi.
Under the null hypothesis – following Lehmann (1975) and Sneyers (1990), and for large
sample size (n > 30) – the distribution of D can be taken to be Normal with mean zero and
variance 1/(n-1).
Dividing by the standard deviation, the standardised test statistic:
D1n1n1
D
DVar
DZSR
E.7
follows the standard Normal distribution Z ~ N(0, 1). The p-value of the statistic can then be
evaluated in the usual way. This test is particularly useful for the detection of gradual change
in time series.
Box E.4: Resampling methods
E4.6 Mean-weighted linear regression test
This is a parametric test that assumes that the data come from an EV1 distribution. It tests
whether there is a linear trend by examining the relationship between time (x) and the
variable of interest (y). The regression gradient is estimated by:
Editorial note: An alternate approach to testing whether an observed value of
Spearman’s rank correlation is significantly different from zero is by permutation
resampling. The time series of AM values is randomly re-ordered and the rank correlation
of the resample rs calculated. The procedure is repeated many times. If the actual rank
correlation lies in the top 2.5% or bottom 2.5% of the distribution of rs values obtained by
permutation resampling, the trend can be declared significant at the 5% level.
Critical values of test statistics in rank-based methods are typically complicated by the
presence of tied values (see Box E.3). A well-executed approach based on permutation
resampling will deal with this automatically, whereas a theoretical approach may require
specially designed adjustments such as those in Equation E.3.
Resampling methods can be used to evaluate test statistics or to construct confidence
intervals in a wide range of hydrological applications. The resampling method can be
designed to preserve important structure in the dataset, e.g. resampling AM series by year
number when estimating confidence intervals for a pooled flood growth curve.
Volume II Flood Frequency Estimation
175
n
1i
2
i
n
1i
ii
xx
yyxx
b E.8
The test statistic bs is:
y
bbs where
n
1i
iyn
1y E.9
Under the null hypothesis of bs = 0 (i.e. the hypothesis of independent and identically
distributed samples drawn from an EV1 distribution), the critical values of bs for various
significance levels were calculated by sampling experiments.
Ten thousand time series each of size n = 50 were generated from the EV1 population with a
mean μ = 100 and coefficient of variation CV = 0.3. The mean value of 100 was chosen
arbitrarily. The CV of 0.3 is chosen to be typical of Irish AM flood series. Critical values at
various significance levels for sample sizes from three to 60 were derived. They are
presented later in Table E.6 of Section E8.
E4.7 Mann-Whitney U test
The Mann-Whitney U test is a non-parametric test of the null hypothesis that two populations
are the same against the alternative hypothesis that one population tends to have larger values
than the other. When applied to test for step-change in a series, the full sample is broken
down into early and late samples of size n1 and n2.
For some flood series, the separation into subsamples is known a priori: the date of change
corresponding to drainage works or to a known change in flow measurement. In other cases,
the date of any step-change is speculative or unknown. [Editorial note: Kundzewicz and
Robson (2004) recommend use of the median change-point test (Pettitt, 1979) in these cases.]
The Mann-Whitney U test labels observations from the two subsamples before ranking them
jointly. If the subsamples differ substantially, their elements will be poorly mixed in the joint
ranking, with the elements of one subsample displaying typically low rank numbers and those
of the other displaying typically high rank numbers.
The test statistic U is calculated as the smaller of U1 and U2 where:
111
1 R2
1nnU
and 1212 UnnU E.10
Here, n1 and n2 are the sizes of the first and second subsamples and R1 is the sum of the ranks
attributed to members of the first subsample in the ranked total sample. An alternate name is
hence the rank-sum test.
The test statistic U is designed to take a low value when the subsamples are not well mixed.
If the observed U value is less than a certain critical value Ucr, the hypothesis that there is no
difference (in typical values) between subsamples is rejected, and a step-change declared
significant. The quantity Ucr depends on the significance level adopted and the subsample
Volume II Flood Frequency Estimation
176
sizes. It is tabulated by Siegel (1956). When both subsamples are larger than 10, U is
approximately Normally distributed with mean n1n2/2 and variance n1n2(n1+n2+1)/12.
E4.8 Turning points test (Kendall’s test)
This non-parametric test is based on counting turning points in the series. Turning points are
triples of consecutive values xi-1, xi, xi+1 such that xi-1 < xi > xi+1 or xi-1 > xi < xi+1. If N is the
number of turning points, then the test statistic is:
2916n
1042n3NS
E.11
Under the null hypothesis of iid values, S is N(0, 1) distributed (Srikanthan et al., 1983).
E4.9 Rank difference test (Meacham test)
This is a nonparametric test based on computing differences in ranks of consecutive values in
the time series. The test statistic is:
74n1n2n
101n3US
2
E.12
where U denotes the sum of absolute values of the rank differences. Under the null
hypothesis of iid values, S is N(0, 1) distributed (Srikanthan et al., 1983).
E5 Exploratory data analysis
E5.1 Selection of data
It is important to consider carefully the form and frequency of the data that should be
analysed. This usually depends on the focus of the study. For floods, the biggest flow is
often of interested; for droughts, it may be the duration of low flows. Selection of which
stations to use in a study is also important (Kundzewicz and Robson, 2004). For example, in
order to study the climate-change signature in river flow, data should ideally be taken from
baseline rivers and should be of high quality and extend over a long period. Data should be
quality-controlled before commencing an analysis of change.
For the present study, annual maximum flood series were selected for 117 stations graded A1
or A2. Some 79 of these have continuous records ranging from 17 to 58 years in length. The
remaining 38 stations have intermittent records with the number of consecutive missing
values in the series ranging from 1 to 16.
As part of the EDA, the AM flows for each station were plotted and a linear trend fitted using
Excel’s data analysis tool. Most of the AM series were found to be well behaved with only a
few series showing outliers.
E5.2 Stations showing trend
Twelve of the 117 series showed a notable upward or downward trend (see Table E.1). The
two stations marked in red are amongst those rejected from general study for reasons
discussed in Appendix B.
Volume II Flood Frequency Estimation
177
Table E.1: Stations with significant trends
Station
number Water-body and location
#
AM
Station
number Water-body and location
#
AM
07004 Blackwater (Kells) at Stramatt 48 25014 Silver at Millbrook 54
15004 Nore at Mcmahons Bridge 51 25023 Little Brosna at Milltown 52
16003 Clodiagh at Rathkennan 51 25029 Nenagh at Clarianna 33
24001 Maigue at Croom 51 26005 Suck at Derrycahill 51
25002 Newport at Barrington’s Bridge 51 34018 Castlebar at Turlough 27
25003 Mulkear at Abington 51 36011 Erne at Bellahillan 49
The data series and trend-lines for these 12 series are shown in Figure E.1. Note that each
diagram shows two stations. In addition to the rejected Mulkear at Abington (Station 25003),
the Nore at MacMahon’s Bridge and Nenagh at Clarianna also show significant downward
trend. The remaining nine series show upward trend. [Editorial note: Two missing annual
maxima – for 1981 at Station 07004 and for 1984 at Station 36011 – have been inadvertently
ascribed a value of 0.0 in the trend analysis. The gap at Station 07004 corresponds to when
arterial drainage was undertaken.]
E5.3 Stations with pre and post-drainage records
The likely impact of arterial drainage works on the AM flows in some of these catchments is
evident. Fourteen stations with pre and post-drainage records are listed in Table E.2 and
illustrated in Figure E.2. [Editorial note: Station 07004 shown in Figure E.1e is a 15th
such
station. It has been added to Table E.2.]
Table E.2: Stations with pre and post-drainage records
Station
number Water-body and location
#
AM
Station
number Water-body and location
#
AM
07002 Deel at Killyon 46 23002 Feale at Listowel 59
07003 Blackwater (Enfield) at Castlerickard 46 24001 Maigue at Croom 51
07004 Blackwater (Kells) at Stramatt 48 25001 Mulkear at Annacotty 49
07005 Boyne at Trim 47 26012 Boyle at Tinacarra 48
07007 Boyne at Boyne Aqueduct 45 30001 Aille at Cartronbower 48
07010 Blackwater (Kells) at Liscartan 46 30004 Clare at Corrofin 35
07011 Blackwater (Kells) at O’Daly’s Bridge 44 30005 Robe at Foxhill 49
07012 Boyne at Slane Castle 65
In most cases an upward step-change is evident in the AM flows following arterial drainage.
Station 25001 Mulkear at Annacotty is amongst those rejected from general study for reasons
discussed in Appendix B. It is seen from Figure E.2ℓ that Station 30001 Aille at
Cartronbower shows a downward step-change.
Volume II Flood Frequency Estimation
178
Figure E.1: Stations for which AM flow series shows significant trend
Annual Maximum Flow
y = 0.0979x - 181.51
y = -0.119x + 253.06
0
5
10
15
20
25
30
1951 1961 1971 1981 1991 2001
Year
Flo
w (
m3/
s)
25014 2523 Linear (2523) Linear (25014)
Annual Maximum Flow
y = 0.6966x - 1330.7
y = -0.2524x + 560.58
0
10
20
30
40
50
60
70
80
1972 1977 1982 1987 1992 1997 2002
Year
Flo
w (
m3/s
)
25029 25002 Linear (25029) Linear (25002)
Annual Maximum Flow
y = 0.4125x - 723.53
y = 0.8654x - 1605.7
0
50
100
150
200
250
1953 1963 1973 1983 1993 2003
Year
Flo
w (
m3/s
)
24001 26005 Linear (26005) Linear (24001)
Annual Maximum Flow
y = 0.2256x - 415.25
y = 0.1887x - 354.92
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
40.00
45.00
50.00
1953 1963 1973 1983 1993 2003
Year
Flo
w (
m3/
s)
7004 16003 Linear (16003) Linear (7004)
Annual Maximum Flow
y = 0.0763x - 133.58
y = 0.1354x - 257.85
0
5
10
15
20
25
1953 1963 1973 1983 1993 2003
Year
Flo
w (
m3/s
)
34018 36011 Linear (36011) Linear (34018)
(a)
(b)
(c)
(d)
(e)
(f)
15004
25014
25023
25029
25002
26005
24001
16003
36011
34018
07004
25003
Volume II Flood Frequency Estimation
179
Figure E.2: AM flow series for stations with arterial drainage works during record
Annual Maximum Flows - Deel at Killyon (7002)
0
5
10
15
20
25
30
35
40
1953 1963 1973 1983 1993 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Blackwater at Castlerickard (7003)
0
5
10
15
20
25
30
35
40
1953 1963 1973 1983 1993 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Boyne at Trim (7005)
0
50
100
150
200
250
1953 1963 1973 1983 1993 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Boyne at Boyne Aqueduct (7007)
0
10
20
30
40
50
60
70
80
1953 1963 1973 1983 1993 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Blackwater at Liscarton (7010)
0
10
20
30
40
50
60
70
80
90
100
1953 1963 1973 1983 1993 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Blackwater at O'Daly's Bridge (7011)
0
10
20
30
40
50
60
70
1958 1963 1968 1973 1978 1983 1988 1993 1998 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Boyne at Slane Castle (7012)
0
50
100
150
200
250
300
350
400
450
500
1940 1950 1960 1970 1980 1990 2000
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Feale at Listowel (23002)
0
100
200
300
400
500
600
700
800
900
1946 1956 1966 1976 1986 1996
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Maigue at Croom (24001)
0
50
100
150
200
250
1953 1958 1963 1968 1973 1978 1983 1988 1993 1998 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Mulkear at Annacotty (25001)
0
20
40
60
80
100
120
140
160
180
200
1953 1963 1973 1983 1993 2003
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Boyle at Tinacarra (26012)
0
10
20
30
40
50
60
70
1957 1962 1967 1972 1977 1982 1987 1992 1997 2002
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Aille atCartonbower (30001)
0
5
10
15
20
25
30
35
1952 1957 1962 1967 1972 1977 1982 1987 1992 1997
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Clare at Corrofin (30004)
0
20
40
60
80
100
120
140
1951 1961 1971 1981 1991 2001
Time
Flo
w (
m3/s
)
Annual Maximum Flows - Robe at Foxhill (30005)
0
10
20
30
40
50
60
1955 1960 1965 1970 1975 1980 1985 1990 1995 2000
Time
Flo
w (
m3/s
)
07010 07011
07012
07005
23002
26012
07002 07003
07007
24001 25001
30005
30004
30001
(a)
(c)
(e)
(g)
(i)
(k)
(m)
(b)
(d)
(f)
(h)
(j)
(l)
(n)
Volume II Flood Frequency Estimation
180
E6 Trend analysis and results
E6.1 Formats of flood data tested for trend
Based on the availability of records and the evidence of drainage improvement works,
statistical tests were applied to the flood data in two formats:
Full AM flood series;
Median of 5-yearly blocks of the AM flood series.
The use of 5-year medians can be viewed as a method of smoothing. Moreover, it is tolerant
of data imperfection. Here, the median of four recorded AM values was accepted as a
reasonable estimate of the 5-year median in cases where a gap in measurement or processing
led to the AM value being unknown in isolated years.
Inferences from the two formats of flood data did not differ greatly. Consequently, only the
results for the conventional analysis of AM series are reported here. Results of applying tests
to AM series for 94 of the 117 A1 + A2 stations are now presented. The other 23 stations had
more than one consecutive AM value missing.
Fifteen of these 94 stations had isolated years where the AM value was missing. At these
sites, the missing years were disregarded and the set of available values assumed to have
occurred consecutively in the series.
E6.2 Main test results
Trends in both data formats were analysed using the tests described in Section E4. A Fortran
program was developed and used in all relevant calculations. Values of the various test
statistics are presented in Table E.3 for the annual maximum series from each of 94 stations.
Results significant at the 5% and 1% level are highlighted.
Table E.3: Test statistics for trend and change in AM flow series
Emboldened entries are significant at 5%; values in red are highly significant at 1% level
Station
#
#
AM
Mann-
Kendall
Spearman’s
ρ
Linear
regression
Mann-
Whitney
U
Turning
Point
Rank
difference
06011 48 0.248 0.281 0.558 0.117 0.415 0.327
06013 30 0.803 0.963 0.944 0.254 0.551 0.770
06014 30 0.301 0.365 0.234 0.852 0.234 0.267
06025 30 0.027 0.035 0.253 0.008 0.551 0.968
06026 46 0.130 0.125 0.038 0.860 0.812 0.319
07004 48 0.000 0.000 0.001 0.000 0.000 0.000
07006 19 0.972 0.913 0.983 0.221 0.127 0.853
07009 29 0.409 0.336 0.236 0.012 0.172 0.453
07033 25 0.176 0.184 0.133 0.064 0.870 0.171
08002 21 0.740 0.814 0.637 0.097 0.367 0.819
09001 48 0.644 0.601 0.699 0.877 0.642 0.498
Volume II Flood Frequency Estimation
181
Emboldened entries are significant at 5%; values in red are highly significant at 1% level
Station
#
#
AM
Mann-
Kendall
Spearman’s
ρ
Linear
regression
Mann-
Whitney
U
Turning
Point
Rank
difference
09002 24 0.286 0.363 0.001 0.707 0.401 0.502
10021 24 0.172 0.105 0.186 0.006 0.737 0.587
11001 33 0.745 0.728 0.631 0.759 0.571 0.910
14006 51 0.153 0.144 0.144 0.777 0.143 0.643
14009 25 0.072 0.064 0.214 0.663 0.412 0.560
14013 50 0.300 0.255 0.300 0.509 0.733 0.934
14018 51 0.065 0.062 0.111 0.019 0.910 0.777
14019 51 0.929 0.852 0.984 0.492 0.215 0.595
14034 17 0.564 0.700 0.748 0.229 1.000 0.416
15003 51 0.968 0.977 0.763 0.095 0.910 0.477
15004 51 0.002 0.002 0.001 0.000 0.024 0.024
16002 51 0.121 0.131 0.091 0.865 0.215 0.874
16003 51 0.000 0.000 0.008 0.000 0.910 0.000
16004 48 0.319 0.348 0.429 0.734 0.816 0.310
16005 30 0.972 0.930 0.724 0.309 0.882 0.374
16008 51 0.024 0.018 0.317 0.002 0.215 0.440
16009 52 0.068 0.076 0.321 0.015 0.655 0.959
18005 50 0.118 0.129 0.151 0.014 0.733 0.352
19001 48 0.756 0.581 0.754 0.503 0.561 0.461
19020 28 0.086 0.069 0.030 0.223 0.757 0.787
23001 45 0.256 0.338 0.310 0.021 0.548 0.669
23002 59 0.099 0.087 0.047 0.010 0.060 0.780
23012 18 0.058 0.046 0.147 0.001 0.001 0.002
24001 51 0.008 0.003 0.003 0.000 0.573 0.000
24008 30 0.432 0.424 0.356 0.309 0.882 0.582
24013 36 0.099 0.081 0.474 0.028 0.787 0.897
24022 20 0.015 0.015 0.001 0.000 1.000 0.110
24082 28 0.040 0.029 0.042 0.085 0.757 0.162
25002 51 0.004 0.004 0.123 0.025 0.573 0.031
25003 51 0.008 0.005 0.176 0.076 0.024 0.149
25006 52 0.336 0.227 0.149 0.552 0.655 0.142
25014 54 0.001 0.001 0.007 0.017 0.584 0.009
25017 55 0.133 0.163 0.329 0.031 0.665 0.738
25021 44 0.015 0.032 0.354 0.001 0.144 0.220
25023 52 0.005 0.004 0.002 0.001 0.577 0.026
25027 43 0.121 0.116 0.219 0.002 0.805 0.808
25029 33 0.007 0.004 0.014 0.000 0.571 0.058
25030 48 0.290 0.282 0.655 0.228 0.561 0.996
Volume II Flood Frequency Estimation
182
Emboldened entries are significant at 5%; values in red are highly significant at 1% level
Station
#
#
AM
Mann-
Kendall
Spearman’s
ρ
Linear
regression
Mann-
Whitney
U
Turning
Point
Rank
difference
25158 18 0.096 0.209 0.342 0.047 0.116 0.559
26002 53 0.759 0.810 0.579 0.346 0.185 0.925
26005 51 0.003 0.004 0.105 0.001 0.910 0.681
26006 53 0.203 0.277 0.279 0.057 0.008 0.272
26007 53 0.730 0.739 0.990 0.233 1.000 0.167
26008 50 0.044 0.052 0.391 0.012 0.088 0.071
26009 35 0.191 0.229 0.612 0.987 0.681 0.496
26017 49 0.241 0.417 0.817 0.059 0.066 0.167
26018 49 0.163 0.198 0.440 0.052 0.420 0.116
26019 51 0.038 0.051 0.180 0.246 0.367 0.550
26020 33 0.133 0.221 0.307 0.000 0.777 0.276
26021 30 0.830 0.960 0.676 0.361 0.766 0.419
26022 33 0.448 0.489 0.588 0.589 0.571 0.589
26059 23 0.023 0.028 0.227 0.829 0.303 0.359
27001 30 0.015 0.016 0.280 0.068 0.766 0.402
27002 51 0.266 0.292 0.418 0.078 0.215 0.770
27003 48 0.488 0.464 0.713 0.773 0.201 0.864
27070 29 0.378 0.371 0.033 0.879 0.006 0.001
29004 32 0.517 0.366 0.590 0.070 0.031 0.902
29011 22 0.085 0.065 0.030 0.045 0.159 0.606
29071 29 0.051 0.047 0.113 0.000 1.000 0.473
30007 31 0.586 0.453 0.968 0.594 0.770 0.314
31002 26 0.567 0.476 0.529 0.228 0.054 0.075
32012 24 0.655 0.755 0.923 0.977 0.737 0.977
33070 25 0.118 0.128 0.450 0.007 0.511 0.091
34001 36 0.017 0.014 0.164 0.004 0.787 0.166
34003 29 0.764 0.775 0.738 0.034 0.363 0.565
34009 33 0.938 0.726 0.919 0.368 0.024 0.077
34011 30 0.284 0.224 0.433 0.245 0.882 0.003
34018 27 0.007 0.006 0.097 0.001 0.753 0.128
34024 29 0.722 0.730 0.691 0.285 0.363 0.632
35001 29 0.311 0.281 0.373 0.432 0.006 0.038
35002 34 0.328 0.385 0.470 0.196 0.329 0.369
35071 30 0.261 0.209 0.277 0.018 0.457 0.361
35073 30 0.695 0.770 0.917 0.263 0.297 0.968
36010 50 0.122 0.175 0.330 0.029 0.494 0.013
36011 49 0.004 0.005 0.097 0.001 0.420 0.038
36012 47 0.011 0.009 0.029 0.009 0.078 0.003
Volume II Flood Frequency Estimation
183
Emboldened entries are significant at 5%; values in red are highly significant at 1% level
Station
#
#
AM
Mann-
Kendall
Spearman’s
ρ
Linear
regression
Mann-
Whitney
U
Turning
Point
Rank
difference
36015 33 0.020 0.015 0.066 0.001 0.479 0.101
36018 50 0.019 0.020 0.132 0.014 1.000 0.798
36019 47 0.011 0.017 0.138 0.001 0.158 0.010
36021 27 0.662 0.596 0.521 0.789 0.270 0.287
36031 30 0.169 0.213 0.567 0.534 0.766 0.793
39008 33 0.780 0.763 0.758 0.113 0.571 0.651
39009 33 0.828 0.825 0.943 0.126 0.777 0.972
The numbers of times the null hypothesis is rejected in the records taking each test are
summarised in Table E.4.
Table E.4: Number of cases (out of 94) in which the null hypothesis is rejected
Significance
level
Trend tests
Step-
change
test
Randomness tests Number
expected
(from 94
cases) by
chance
alone
Mann-
Kendall
Spearman
ρ
Mean-
weighted
linear
regression
Mann-
Whitney
U
Turning
point
Rank
difference
1% 12 13 8 24 5 9 ≈1
5%
but not 1% 14 13 8 15 4 6 ≈4
5%
(all cases) 26 26 16 39 9 15 ≈5
E6.3 Non-randomness
The tests of randomness identified significant non-randomness in some 15 (or 16%) of the 94
AM series. This is considerably more than the five expected by chance alone. There is
therefore evidence of some non-randomness in Irish AM flow data.
E6.4 Trend
In the trend tests, 26 out of 94 cases show statistically significant trends at the 5% level
according to the non-parametric (Mann-Kendall and Spearman ρ) tests. There is therefore
strong evidence of trend in Irish AM flow data.
The AM series that show the most marked departures from the null hypothesis (of zero trend)
are listed in Table E.5, where the direction of trend is also noted. These are taken from the
results presented in Table E.3. [Editorial note: The stations have been informally ranked
according to the sum of the p-values indicated by the three tests of trend.]
Volume II Flood Frequency Estimation
184
It is the nature of applying such tests widely that some of the stations exhibiting trend have
their own peculiarity. Possible explanations for trend are indicated in the comment column.
Arterial drainage is seen to feature in four of the seven series most clearly showing trend.
Table E.5: Stations for which the trend is judged most highly significant
Σ p-
values
Station
number Station name
Trend
noted Comment AD
* AD2
*
0.001 07004 Blackwater (Kells)
at Stramatt Increase
Arterial drainage completed 1981;
more step-change than trend 0.380 0.138
0.005 15004 Nore at Mcmahons
Bridge Increase
Arterial drainage completed 1981;
more step-change than trend 0.000 0.000
0.008 16003 Clodiagh at
Rathkennan Increase
Different rating before and after
1978 0.000 0.000
0.009 25014 Silver at Millbrook Decrease
Different rating before and after
1971; temporarily different rating
used 1984-1988
0.258 0.371
0.011 25023 Little Brosna at
Milltown Increase
Different rating before and after
1971 0.000 0.000
0.014 24001 Maigue at Croom Increase Maigue drainage works spanned
1973 to 1986; rating change in 1989 0.124 0.513
0.025 25029 Nenagh at
Clarianna Increase
Unexplained; drainage works pre-
dates record 0.076 0.411
*AD and AD2 denote the PCDs of arterial drainage: ARTDRAIN and ARTDRAIN2
The AM flow series for these stations were shown in Figure E.1. Station 24001 also appears
in Figure E.2i. An example of the likely influence of arterial drainage is shown in Figure E.3.
2010200019901980197019601950
30
20
10
0
AM
peak
flo
w
Pre-drainage
During drainage
Post-drainage
07004 Blackwater (Kells) at Stramatt
Figure E.3: AM flow series showing likely influence of arterial drainage
Volume II Flood Frequency Estimation
185
E6.5 Discussion
The possibility of trend in the Irish AM flow series cannot be dismissed. Besides
hydrometric reasons such as rating curve uncertainties and weir changes, there are of course
possible hydrological reasons for trend based on land-use changes such as hydroelectric
development, channel improvements and land drainage – and, of course, long-term variation
or change in climate.
It is important to examine the test results alongside graphs of the data, and with as much
historical knowledge about the data as possible (Kundzewicz and Robson, 2004). The impact
of drainage works on the AM flows at 12 stations was identified through EDA, as discussed
previously and illustrated in Figure E.2. An abrupt change in AM flows after drainage is
noted at most of these stations. In cases where drainage works are absent or pre-date the
record, further examination is recommended of series exhibiting trends in the AM flow.
Given effects from arterial drainage, changed hydrometry and other factors, it is barely
practical to explore the possible effect of climate change on the AM flows by empirical
analysis alone. Exploration of a small number of records proved inconclusive.
E7 Summary
In order to test for trend and other non-randomness in the flood series of Irish rivers,
six statistical tests were applied to AM flow series of 94 stations.
Non-parametric (i.e. distribution-free) statistical tests are considered appropriate to
detecting trend and other non-randomness.
However, a mean-weighted linear regression trend test method has been developed
based on a distribution known to be relatively suitable for AM flow series on Irish
rivers, namely: an EV1 distribution with CV = 0.3. Critical values of the test statistic
were obtained from sampling experiments and are presented in Section E8.
Tests for randomness identified non-randomness in 15 (or 16%) of the 94 AM series
examined. This is appreciably more than the five expected by chance alone. There is
therefore evidence of some non-randomness in Irish AM flow data.
Trend tests revealed some notable effects. Out of 94 records, 26 (28%) show
statistically significant trends (≈13 at the 1% level and a further 13 at the 5% level).
This is very much greater than the number expected by chance alone. Most trends
were upward rather than downward.
Tests for step-change revealed notable jumps in many AM flow series. Out of 94
records, 39 (41%) show statistically significant jumps (24 at the 1% level and a
further 15 at the 5% level). This is very much greater than the number expected by
chance alone. Most jumps were upward rather than downward.
The tests applied do not discriminate well between trend and step-change.
Arterial drainage is implicated in a considerable number of the significant step-
changes and trends reported. Hydrometric changes may account for some effects.
In cases where drainage works are absent or pre-date the record, further examination
of records is recommended for stations exhibiting trends in the AM flow.
Exploration of the possible effect of climate change on the AM flood flows proved
inconclusive.
Volume II Flood Frequency Estimation
186
E8 Critical values for mean-weighted linear-regression test
Critical values of the test statistic bs for the mean-weighted linear regression are given in
Table E.6. These are based on 10,000 simulated random samples from an EV1 distribution
with CV = 0.3. This is a typical value of CV for Irish AM flow data.
Table E.6: Critical values of the test statistic bs
Sample size
n
Lower tail Upper tail
0.005 0.01 0.025 0.05 0.05 0.025 0.01 0.005
3 -0.5247 -0.4742 -0.3993 -0.3388 0.3250 0.3949 0.4724 0.5221
4 -0.3360 -0.3018 -0.2585 -0.2104 0.2083 0.2489 0.2995 0.3364
5 -0.2323 -0.2109 -0.1794 -0.1482 0.1497 0.1793 0.2171 0.2422
6 -0.1822 -0.1662 -0.1387 -0.1182 0.1142 0.1367 0.1631 0.1815
7 -0.1434 -0.1321 -0.1106 -0.0921 0.0921 0.1094 0.1298 0.1417
8 -0.1184 -0.1053 -0.0885 -0.0744 0.0752 0.0884 0.1047 0.1186
9 -0.1004 -0.0910 -0.0756 -0.0633 0.0612 0.0738 0.0886 0.0986
10 -0.0832 -0.0765 -0.0655 -0.0540 0.0529 0.0632 0.0761 0.0872
11 -0.0737 -0.0666 -0.0558 -0.0460 0.0471 0.0565 0.0672 0.0741
12 -0.0639 -0.0584 -0.0488 -0.0410 0.0398 0.0474 0.0561 0.0623
13 -0.0578 -0.0513 -0.0427 -0.0352 0.0352 0.0433 0.0504 0.0571
14 -0.0514 -0.0454 -0.0377 -0.0319 0.0325 0.0384 0.0470 0.0516
15 -0.0456 -0.0403 -0.0344 -0.0290 0.0292 0.0342 0.0404 0.0459
16 -0.0404 -0.0365 -0.0308 -0.0255 0.0256 0.0304 0.0355 0.0388
17 -0.0378 -0.0344 -0.0281 -0.0237 0.0230 0.0274 0.0328 0.0362
18 -0.0328 -0.0308 -0.0257 -0.0216 0.0219 0.0262 0.0311 0.0348
19 -0.0310 -0.0280 -0.0236 -0.0197 0.0200 0.0237 0.0279 0.0311
20 -0.0302 -0.0269 -0.0219 -0.0182 0.0186 0.0217 0.0257 0.0287
21 -0.0270 -0.0244 -0.0203 -0.0174 0.0173 0.0204 0.0239 0.0267
22 -0.0255 -0.0235 -0.0189 -0.0158 0.0162 0.0196 0.0232 0.0256
23 -0.0235 -0.0213 -0.0180 -0.0151 0.0149 0.0176 0.0209 0.0232
24 -0.0226 -0.0203 -0.0170 -0.0144 0.0140 0.0167 0.0199 0.0220
25 -0.0214 -0.0193 -0.0158 -0.0132 0.0133 0.0160 0.0190 0.0215
26 -0.0192 -0.0171 -0.0144 -0.0122 0.0123 0.0149 0.0179 0.0201
27 -0.0184 -0.0167 -0.0142 -0.0119 0.0117 0.0138 0.0165 0.0186
28 -0.0177 -0.0157 -0.0134 -0.0112 0.0111 0.0135 0.0162 0.0175
29 -0.0163 -0.0146 -0.0127 -0.0105 0.0106 0.0126 0.0149 0.0166
30 -0.0161 -0.0146 -0.0121 -0.0101 0.0096 0.0116 0.0136 0.0154
Volume II Flood Frequency Estimation
187
Sample size
n
Lower tail Upper tail
0.005 0.01 0.025 0.05 0.05 0.025 0.01 0.005
31 -0.0147 -0.0134 -0.0113 -0.0094 0.0094 0.0112 0.0133 0.0150
32 -0.0147 -0.0135 -0.0110 -0.0092 0.0091 0.0108 0.0132 0.0146
33 -0.0137 -0.0123 -0.0105 -0.0089 0.0085 0.0103 0.0124 0.0138
34 -0.0133 -0.0122 -0.0100 -0.0084 0.0083 0.0100 0.0118 0.0130
35 -0.0124 -0.0113 -0.0095 -0.0081 0.0082 0.0098 0.0114 0.0124
36 -0.0124 -0.0108 -0.0091 -0.0077 0.0078 0.0092 0.0110 0.0125
37 -0.0124 -0.0103 -0.0086 -0.0073 0.0073 0.0087 0.0104 0.0114
38 -0.0124 -0.0101 -0.0086 -0.0071 0.0069 0.0083 0.0098 0.0110
39 -0.0124 -0.0095 -0.0081 -0.0070 0.0068 0.0080 0.0097 0.0110
40 -0.0124 -0.0090 -0.0078 -0.0065 0.0065 0.0077 0.0092 0.0102
41 -0.0124 -0.0090 -0.0075 -0.0063 0.0064 0.0076 0.0089 0.0099
42 -0.0124 -0.0086 -0.0072 -0.0060 0.0061 0.0073 0.0088 0.0097
43 -0.0124 -0.0085 -0.0072 -0.0059 0.0059 0.0070 0.0083 0.0092
44 -0.0124 -0.0078 -0.0067 -0.0056 0.0056 0.0067 0.0080 0.0092
45 -0.0124 -0.0078 -0.0067 -0.0055 0.0055 0.0066 0.0080 0.0089
46 -0.0124 -0.0075 -0.0063 -0.0054 0.0053 0.0063 0.0074 0.0081
47 -0.0124 -0.0072 -0.0062 -0.0053 0.0051 0.0061 0.0073 0.0081
48 -0.0124 -0.0070 -0.0059 -0.0049 0.0050 0.0060 0.0071 0.0077
49 -0.0124 -0.0069 -0.0057 -0.0048 0.0049 0.0058 0.0071 0.0077
50 -0.0124 -0.0065 -0.0055 -0.0046 0.0047 0.0056 0.0068 0.0075
51 -0.0124 -0.0063 -0.0054 -0.0045 0.0046 0.0054 0.0064 0.0071
52 -0.0124 -0.0061 -0.0052 -0.0044 0.0044 0.0053 0.0062 0.0070
53 -0.0124 -0.0060 -0.0051 -0.0043 0.0041 0.0049 0.0059 0.0065
54 -0.0124 -0.0059 -0.0050 -0.0042 0.0041 0.0049 0.0058 0.0064
55 -0.0124 -0.0059 -0.0049 -0.0041 0.0040 0.0049 0.0058 0.0063
56 -0.0124 -0.0057 -0.0048 -0.0041 0.0039 0.0047 0.0056 0.0062
57 -0.0124 -0.0054 -0.0046 -0.0038 0.0038 0.0046 0.0054 0.0061
58 -0.0124 -0.0053 -0.0045 -0.0037 0.0038 0.0045 0.0054 0.0059
59 -0.0124 -0.0051 -0.0043 -0.0036 0.0037 0.0044 0.0053 0.0058
60 -0.0124 -0.0051 -0.0042 -0.0035 0.0036 0.0043 0.0050 0.0055
Volume II Flood Frequency Estimation
188
Appendix F Additional summary statistics
Station number and name Grade AREA
km2
N Qmean Qmed Qmax Qmax/
Qmean
Qmax/
AREA Qmed/
Qmean
01041 Deele at Sandy Mills B 116.20 32 85.08 82.61 147.16 1.73 1.27 0.97
01055 Mourne Beg at Mourne Beg Weir B 10.80 9 2.92 2.70 4.72 1.61 0.44 0.92
06011 Fane at Moyles Mill A1 234.00 48 15.86 15.39 26.36 1.66 0.11 0.97
06013 Dee at Charleville Weir A1 307.00 30 27.81 27.37 41.84 1.50 0.14 0.98
06014 Glyde at Tallanstown A1 270.00 30 22.56 21.46 39.40 1.75 0.15 0.95
06021 Glyde at Mansfieldstown B 321.00 50 21.54 21.50 33.00 1.53 0.10 1.00
06025 Dee at Burley Bridge A1 176.00 30 18.32 18.69 23.57 1.29 0.13 1.02
06026 Glyde at Aclint Bridge A1 144.00 46 13.87 12.30 24.12 1.74 0.17 0.89
06030 Big at Ballygoly See Box 4.3 B 10.20 30 20.58 10.03 122.00 5.93 11.96 0.49
06031 Flurry at Curralhir A2 45.30 18 13.58 11.70 35.80 2.64 0.79 0.86
06033 White Dee at Coneyburrow Bridge B 57.40 25 27.88 18.60 92.30 3.31 1.61 0.67
06070 Muckno L. at Muckno A1 153.50 24 13.32 13.19 20.93 1.57 0.14 0.99
07006 Moynalty at Fyanstown A2 176.00 19 26.73 27.93 34.05 1.27 0.19 1.05
07009 Boyne at Navan Weir A1 1610.00 29 162.64 134.80 297.60 1.83 0.18 0.83
07033 Blackwater at Virginia Hatchery A2 129.00 25 14.93 14.62 26.58 1.78 0.21 0.98
08002 Delvin at Naul A1 37.00 20 5.62 5.32 8.96 1.59 0.24 0.95
08003 Broadmeadow at Fieldstown B 76.20 18 26.88 22.55 110.00 4.09 1.44 0.84
08005 Sluice at Kinsaley Hall A2 10.10 18 3.04 2.50 7.81 2.57 0.77 0.82
08007 Broadmeadow at Ashbourne B 34.00 15 9.88 8.24 18.79 1.90 0.55 0.83
08008 Broadmeadow at Broadmeadow A2 110.00 25 44.55 40.90 123.69 2.78 1.12 0.92
08009 Ward at Balheary A1 62.00 11 10.38 6.59 53.60 5.17 0.86 0.64
08011 Nanny at Duleek Road Bridge B 181.00 23 31.00 32.22 45.41 1.46 0.25 1.04
08012 Stream at Ballyboghill B 22.10 19 4.21 4.35 8.15 1.93 0.37 1.03
09001 Ryewater at Leixlip A1 215.00 48 38.71 35.46 91.50 2.36 0.43 0.92
09002 Griffeen at Lucan A1 37.00 24 7.24 5.40 23.70 3.28 0.64 0.75
09010 Dodder at Waldron’s Bridge A1 95.20 19 70.15 48.00 269.00 3.83 2.83 0.68
09035 Cammock at Killeen Road B 54.70 9 12.04 11.70 27.80 2.31 0.51 0.97
10002 Avonmore at Rathdrum B 233.00 47 88.19 83.49 266.64 3.02 1.14 0.95
10021 Shanganagh at Common’s Road A1 30.90 24 7.87 7.36 14.30 1.82 0.46 0.94
10022 Cabinteely at Carrickmines A1 10.40 18 3.84 3.85 6.89 1.79 0.66 1.00
10028 Aughrim at Knocknamohill B 204.10 16 56.69 46.95 102.00 1.80 0.50 0.83
11001 Owenavorragh at Boleany B 148.00 33 49.85 47.17 120.70 2.42 0.82 0.95
12001 Slaney at Scarawalsh A2 1036.00 50 169.50 157.00 399.00 2.35 0.39 0.93
12013 Slaney at Rathvilly B 185.00 30 45.16 43.55 72.30 1.60 0.39 0.96
14005 Barrow at Portarlington A2 398.00 48 40.81 38.27 80.42 1.97 0.20 0.94
14006 Barrow at Pass Bridge A1 1096.00 51 83.76 80.52 137.38 1.64 0.13 0.96
Volume II Flood Frequency Estimation
189
Station number and name Grade AREA
km2
N Qmean Qmed Qmax Qmax/
Qmean
Qmax/
AREA Qmed/
Qmean
14007 Stradbally at Derrybrock A1 115.00 25 16.94 16.20 29.30 1.73 0.25 0.96
14009 Cushina at Cushina A2 68.00 25 6.69 6.79 11.19 1.67 0.16 1.02
14011 Slate at Rathangan A1 163.00 26 12.07 12.30 18.70 1.55 0.11 1.02
14013 Burrin at Ballinacarrig A2 154.00 50 16.54 16.05 26.32 1.59 0.17 0.97
14018 Barrow at Royal Oak A1 2415.00 51 141.83 147.98 216.07 1.52 0.09 1.04
14019 Barrow at Levitstown A1 1660.00 51 103.46 102.41 162.86 1.57 0.10 0.99
14029 Barrow at Graiguenamanagh A2 2762.00 47 162.54 160.74 206.21 1.27 0.07 0.99
14033 Owenass at Mountmellick B 185.00 22 22.59 19.50 33.00 1.46 0.18 0.86
14034 Barrow at Bestfield A2 2060.00 14 137.30 125.00 247.00 1.80 0.12 0.91
15001 Kings at Annamult A2 443.00 42 89.39 88.75 151.00 1.69 0.34 0.99
15002 Nore at John’s Bridge A2 1605.00 35 211.98 197.00 393.00 1.85 0.24 0.93
15003 Dinan at Dinan Bridge A2 298.00 50 143.58 150.76 187.52 1.31 0.63 1.05
15004 Nore at McMahons Bridge A2 491.00 51 38.96 37.28 74.96 1.92 0.15 0.96
15005 Erkina at Durrow Foot Bridge B 387.00 50 28.47 27.44 61.85 2.17 0.16 0.96
15012 Nore at Ballyragget B 945.00 16 77.16 77.11 133.00 1.72 0.14 1.00
16001 Drish at Athlummon A2 140.00 33 15.65 15.66 24.49 1.56 0.17 1.00
16002 Suir at Beakstown A2 512.00 51 55.40 52.66 123.88 2.24 0.24 0.95
16003 Clodiagh at Rathkennan A2 246.00 51 31.17 29.98 45.72 1.47 0.19 0.96
16004 Suir at Thurles A2 236.00 48 22.17 21.37 34.89 1.57 0.15 0.96
16005 Multeen at Aughnagross A2 87.00 30 23.11 21.79 34.31 1.48 0.39 0.94
16006 Multeen at Ballinclogh Bridge B 75.00 33 30.37 27.87 58.07 1.91 0.77 0.92
16007 Aherlow at Killardry B 273.00 51 79.18 75.84 138.03 1.74 0.51 0.96
16008 Suir at New Bridge A2 1120.00 51 90.66 92.32 110.91 1.22 0.10 1.02
16009 Suir at Cahir Park A2 1602.00 52 159.29 162.21 206.00 1.29 0.13 1.02
16011 Suir at Clonmel A1 2173.00 52 234.52 223.00 389.00 1.66 0.18 0.95
16012 Tar at Tar Bridge B 228.00 36 55.20 54.57 92.20 1.67 0.40 0.99
16013 Nire at Fourmilewater B 91.00 33 101.69 93.21 207.02 2.04 2.27 0.92
16051 Rossestown at Clobanna B 34.18 13 2.95 2.85 5.67 1.92 0.17 0.96
18001 Bride at Mogeely Bridge B 335.00 48 71.07 71.49 96.93 1.36 0.29 1.01
18002 Ballyduff at Muns Blackwater B 2338.00 49 353.65 344.00 479.00 1.35 0.20 0.97
18003 Blackwater at Killavullen B 1258.00 49 282.76 266.15 502.74 1.78 0.40 0.94
18004 Ballynamona at Awbeg A2 324.00 46 30.96 31.20 52.70 1.70 0.16 1.01
18005 Funshion at Downing Bridge A2 363.00 50 56.69 53.05 109.73 1.94 0.30 0.94
18006 Blackwater at CSET Mallow B 1058.00 27 291.30 286.00 397.00 1.36 0.38 0.98
18016 Blackwater at Duncannon B 113.00 24 80.99 79.65 114.82 1.42 1.02 0.98
18048 Blackwater at Dromcummer B 881.00 23 222.77 220.00 269.00 1.21 0.31 0.99
18050 Blackwater at Duarrigle B 244.60 24 121.96 124.50 175.00 1.43 0.72 1.02
19001 Owenboy at Ballea Upper A2 106.00 48 15.87 15.42 22.03 1.39 0.21 0.97
Volume II Flood Frequency Estimation
190
Station number and name Grade AREA
km2
N Qmean Qmed Qmax Qmax/
Qmean
Qmax/
AREA Qmed/
Qmean
19014 Lee at Dromcarra B 168.00 20 79.69 71.89 157.21 1.97 0.69 0.90
19016 Bride at Ovens Bridge B 123.00 8 28.74 29.58 34.87 1.21 0.60 1.03
19020 Owennacurra at Ballyedmond A2 75.00 28 24.63 22.40 38.70 1.57 0.52 0.91
19031 Sullane at Macroom B 210.00 9 131.09 135.90 201.18 1.53 0.67 1.04
19046 Martin at Station Road B 60.40 9 31.09 29.95 41.95 1.35 0.70 0.96
20002 Bandon at Curranure B 431.00 31 140.60 126.28 287.11 2.04 0.44 0.90
20006 Argideen at Clonakilty WW B 79.30 25 30.25 27.70 55.60 1.84 0.25 0.92
22006 Flesk at Flesk Bridge B 325.00 51 165.89 169.09 282.83 1.70 0.43 1.02
22009 Dreenagh at White Bridge B 37.00 24 11.91 11.47 16.35 1.37 0.63 0.96
22035 Laune at Laune Bridge B 559.65 14 112.81 116.40 141.48 1.25 0.40 1.03
23001 Galey at Inch Bridge A2 196.00 45 97.39 99.05 210.07 2.16 1.07 1.02
23012 Lee at Ballymullen A2 60.00 18 16.87 15.66 31.74 1.88 0.53 0.93
24002 Camogue at Gray’s Bridge A2 231.00 27 24.06 23.49 35.21 1.46 0.15 0.98
24004 Maigue at Bruree B 246.00 52 54.86 50.63 104.55 1.91 0.29 0.92
24008 Maigue at Castleroberts A2 805.00 30 120.96 119.13 194.86 1.61 0.24 0.98
24011 Deel at Deel Bridge B 273.00 33 103.01 104.55 171.74 1.67 0.48 1.01
24012 Deel at Grange Bridge B 359.00 41 110.45 109.99 141.95 1.29 0.16 1.00
24022 Mahore at Hospital A2 39.70 20 9.83 9.80 20.50 2.09 0.52 1.00
24030 Deel at Danganbeg B 248.00 25 52.89 52.00 72.20 1.37 0.45 0.98
24082 Maigue at Islandmore A2 764.00 28 135.47 140.01 206.35 1.52 0.27 1.03
25001 Mulkear at Annacotty See §4.2.1 A2 646.00 49 133.95 132.88 178.58 1.33 0.28 0.99
25002 Newport at Barringtons Bridge §4.2.1 A2 223.00 51 61.15 62.64 74.64 1.22 0.33 1.02
25003 Mulkear at Abington See §4.2.1 A1 397.00 51 69.45 68.98 92.79 1.34 0.23 0.99
25004 Bilboa at Newbridge See §4.2.1 B 125.00 30 41.70 42.30 59.50 1.43 0.49 1.01
25005 Dead at Sunville See §4.2.1 A2 190.00 46 28.73 29.63 33.42 1.16 0.18 1.03
25006 Brosna at Ferbane A1 1207.00 52 86.77 81.91 147.21 1.70 0.12 0.94
25011 Brosna at Moystown B 1227.00 51 85.64 82.02 194.25 2.27 0.41 0.96
25014 Silver at Millbrook Bridge A1 165.00 54 17.67 17.25 27.03 1.53 0.16 0.98
25016 Clodiagh at Rahan A2 274.00 42 23.04 22.57 36.14 1.57 0.13 0.98
25017 Shannon at Banagher A1 7980.00 55 413.25 407.68 596.51 1.44 0.07 0.99
25020 Killimor at Killeen B 197.00 35 46.60 43.65 89.55 1.92 0.32 0.94
25021 Little Brosna at Croghan A2 493.00 44 28.03 28.58 35.80 1.28 0.07 1.02
25023 Little Brosna at Milltown A1 116.00 52 12.14 11.22 20.05 1.65 0.17 0.92
25025 Ballyfinboy at Ballyhooney A1 160.00 31 10.15 10.18 17.40 1.71 0.11 1.00
25027 Ollatrim at Gourdeen Bridge A1 118.00 43 23.32 22.10 40.46 1.73 0.34 0.95
25029 Nenagh at Clarianna A2 301.00 33 54.12 56.48 74.06 1.37 0.25 1.04
25030 Graney at Scarriff Bridge A1 279.00 48 43.80 40.64 87.04 1.99 0.31 0.93
25034 L. Ennell Trib at Rochfort A2 12.00 24 1.50 1.48 2.21 1.47 0.18 0.99
Volume II Flood Frequency Estimation
191
Station number and name Grade AREA
km2
N Qmean Qmed Qmax Qmax/
Qmean
Qmax/
AREA Qmed/
Qmean
25038 Nenagh at Tyone B 139.00 17 42.08 39.30 67.90 1.61 0.21 0.93
25040 Bunow at Roscrea A2 30.00 20 3.78 3.59 6.29 1.67 0.21 0.95
25044 Kilmastulla at Coole A2 98.90 33 25.38 22.70 45.75 1.80 0.46 0.89
25124 Brosna at Ballynagore A2 254.00 18 12.79 13.65 22.50 1.76 0.09 1.07
25158 Bilboa at Cappamore See §4.2.1 A1 116.00 18 47.66 43.88 75.09 1.58 0.65 0.92
26002 Suck at Rookwood A2 626.00 53 56.98 56.56 104.87 1.84 0.17 0.99
26005 Suck at Derrycahill A2 1050.00 51 92.80 93.21 135.94 1.46 0.13 1.00
26006 Suck at Willsbrook A1 182.00 53 26.57 24.23 70.06 2.64 0.38 0.91
26007 Suck at Bellagill Bridge A1 1184.00 53 91.75 88.15 147.84 1.61 0.12 0.96
26008 Rinn at Johnston’s Bridge A1 292.00 49 23.68 22.94 41.02 1.73 0.14 0.97
26009 Black at Bellantra Bridge A2 97.00 35 13.66 13.22 18.76 1.37 0.19 0.97
26010 Cloone at Riverstown B 100.00 35 20.03 17.17 40.80 2.04 0.14 0.86
26014 Lung at Banada Bridge B 222.00 16 44.10 42.82 70.90 1.61 0.77 0.97
26018 Owenure at Bellavahan A2 118.00 49 9.19 8.95 13.98 1.52 0.12 0.97
26019 Camlin at Mullagh A1 260.00 51 22.34 21.18 37.03 1.66 0.14 0.95
26020 Camlin at Argar Bridge A1 128.00 32 11.21 11.27 15.59 1.39 0.12 1.01
26021 Inny at Ballymahon A2 1071.00 30 65.88 66.34 92.51 1.40 0.09 1.01
26022 Fallan at Kilmore A2 950.00 33 6.64 6.49 11.06 1.67 0.01 0.98
26058 Inny Upper at Ballinrink Bridge B 59.00 24 5.98 5.35 12.20 2.04 0.15 0.89
26059 Inny at Finnea Bridge A1 249.00 17 12.98 12.20 16.70 1.29 0.07 0.94
26108 Owenure at Boyle Abbey Bridge B 533.00 15 56.29 57.32 73.07 1.30 0.14 1.02
27001 Claureen at Inch Bridge A2 48.00 30 20.65 20.10 31.70 1.53 0.66 0.97
27002 Fergus at Ballycorey A1 562.00 51 34.22 32.60 59.76 1.75 0.11 0.95
27003 Fergus at Corofin A2 168.00 48 24.01 22.92 40.50 1.69 0.24 0.95
28001 Inagh at Ennistimon B 168.00 17 52.69 47.58 129.28 2.45 0.44 0.90
29001 Raford at Rathgorgin A1 119.00 40 14.17 13.46 19.66 1.39 0.17 0.95
29004 Clarinbridge at Clarinbridge A2 123.00 32 11.39 11.30 14.77 1.30 0.12 0.99
29007 L. Cullaun at Craughwell B 278.00 22 27.83 26.49 42.93 1.54 0.14 0.95
29011 Dunkellin at Kilcolgan Bridge A1 354.00 22 31.94 28.89 66.52 2.08 0.19 0.90
29071 L. Cutra at Cutra A2 123.50 26 16.00 15.70 24.30 1.52 0.20 0.98
30007 Clare at Ballygaddy A2 458.00 31 61.93 62.98 95.98 1.55 0.21 1.02
30012 Clare at Claregalway B 1075.40 9 126.89 126.00 155.00 1.22 0.02 0.99
30021 Robe at Christina’s Bridge B 138.00 26 28.17 27.20 60.70 2.15 0.12 0.97
30031 Cong at Cong Weir B 891.00 24 94.35 93.88 122.28 1.30 1.82 0.99
30037 Robe at Clooncormick B 210.00 21 1.80 1.79 3.19 1.77 1.59 0.99
30061 Corrib Estuary at Wolfe Tone Bridge A2 3111.00 33 274.97 247.97 601.59 2.19 0.19 0.90
31002 Cashla at Cashla A1 72.00 26 12.89 12.16 21.10 1.64 0.29 0.94
31072 Cong at Cong Weir B 891.00 26 49.08 43.20 103.00 2.10 0.57 0.88
Volume II Flood Frequency Estimation
192
Station number and name Grade AREA
km2
N Qmean Qmed Qmax Qmax/
Qmean
Qmax/
AREA Qmed/
Qmean
32011 Bunowen at Louisberg Weir B 68.60 26 74.88 64.87 125.00 1.67 1.28 0.87
32012 Newport at Newport Weir A2 138.30 24 30.06 29.85 36.60 1.22 0.26 0.99
33001 Glenamoy at Glenamoy B 73.00 25 62.11 59.30 116.00 1.87 0.53 0.95
33070 Carrowmore L. at Carrowmore A1 90.00 28 7.90 7.67 11.97 1.52 0.13 0.97
34001 Moy at Rahans A2 1911.00 36 174.76 174.61 286.56 1.64 0.15 1.00
34003 Moy at Foxford A2 1737.00 29 180.42 178.00 282.00 1.56 0.16 0.99
34007 Deel at Ballycarroon B 156.00 53 90.37 84.48 198.91 2.20 0.12 0.93
34009 Owengarve at Curraghbonaun A2 113.00 33 28.37 27.48 38.58 1.36 0.34 0.97
34010 Moy at Cloonacannana B 471.00 12 123.29 113.72 193.07 1.57 0.41 0.92
34011 Manulla at Gneeve Bridge A2 144.00 30 18.80 18.73 26.05 1.39 0.18 1.00
34018 Castlebar at Turlough A1 93.00 27 11.50 11.28 17.33 1.51 0.19 0.98
34024 Pollagh at Kiltimagh A2 128.00 28 20.70 20.80 24.70 1.19 0.19 1.00
35001 Owenmore at Ballynacarrow A2 299.00 29 30.52 31.16 46.04 1.51 0.15 1.02
35002 Owenbeg at Billa Bridge A2 90.00 34 51.78 50.48 69.37 1.34 0.77 0.97
35005 Ballysadare at Ballysadare A2 642.00 55 77.78 75.42 132.71 1.71 0.21 0.97
35011 Bonet at Dromahair B 294.00 36 116.02 115.36 188.01 1.62 1.26 0.99
35071 L. Melvin at Lareen A2 247.20 30 26.95 26.29 37.91 1.41 0.15 0.98
35073 L. Gill at Lough Gill A2 384.00 30 54.81 54.05 78.40 1.43 0.20 0.99
36010 Annalee at Butlers Bridge A1 774.00 50 66.56 66.80 106.62 1.60 0.14 1.00
36011 Erne at Bellahillan B 318.00 49 17.91 18.23 23.60 1.32 0.07 1.02
36012 Erne at Sallaghan A1 263.00 47 14.22 14.12 21.63 1.52 0.08 0.99
36015 Finn at Anlore A1 175.00 33 23.14 22.08 49.99 2.16 0.29 0.95
36018 Dronmore at Ashfield Bridge A1 233.00 50 15.84 16.25 24.43 1.54 0.10 1.03
36019 Erne at Belturbet A2 1501.00 47 89.60 89.95 119.43 1.33 0.08 1.00
36021 Yellow at Kiltybarden A2 23.00 27 24.96 23.37 43.57 1.75 1.89 0.94
36031 Cavan at Lisdarn A2 52.00 30 6.85 6.45 13.70 2.00 0.26 0.94
36071 L. Scur at Gowly B 66.00 20 6.36 6.49 8.13 1.28 3.25 1.02
38001 Owenea at Clonconwal B 109.00 33 70.02 70.63 113.38 1.62 0.44 1.01
39001 New Mills at Swilly B 49.00 30 44.88 44.25 61.50 1.37 0.10 0.99
39008 Leannan at Gartan Bridge A2 78.00 33 28.34 28.18 43.88 1.55 0.56 0.99
39009 Fern O/L at Aghawoney A2 207.00 33 45.91 45.72 76.67 1.67 0.37 1.00
Volume II Flood Frequency Estimation
193
Appendix G Sample probability plots and summary information
Water
-year
AM flow
(m3 s
-1)
Date Station 14018 River Barrow at Royal Oak (2415 km2)
1954 193.75 12/12/1954 Summary statistics
1955 99.88 28/01/1956 N = 51 QMED = 147.98
1956 148.07 31/12/1956 Qmean = 141.83 L-CV = 0.138
1957 114.83 12/02/1958 CV = 0.240 L-skew = 0.036
1958 98.02 21/12/1958 H-skew = 0.218 L-kurt = 0.064
1959 111.04 30/12/1959
1960 161.76 05/12/1960 Summer peaks are tabulated in red and (for EV1 case) plotted in red
1961 83.08 17/01/1962
1962 118.69 07/02/1963
1963 95.83 26/03/1964
1964 166.52 13/12/1964
1965 177.97 19/11/1965
1966 160.19 23/02/1967
1967 147.98 25/12/1968
1968 173.01 10/01/1968
1969 121.31 23/02/1970
1970 113.56 25/11/1970
1971 124.64 03/02/1972
1972 102.09 29/12/1972
1973 183.86 02/02/1974
1974 150.23 28/01/1975
1975 92.58 03/12/1975
1976 138.49 23/02/1977
1977 145.76 05/02/1978
1978 145.76 28/12/1978
1979 162.55 28/12/1979
1980 107.94 27/10/1980
1981 106.72 05/01/1982
1982 135.65 09/11/1982
1983 139.93 26/03/1984
1984 151.73 15/12/1984
1985 102.84 28/08/1986
1986 148.73 11/12/1986
1987 145.76 04/02/1988
1988 90.44 26/10/1988
1989 216.07 08/02/1990
1990 150.23 29/12/1990
1991 97.66 07/01/1992
1992 199.57 15/06/1993
1993 151.73 05/02/1994
1994 212.34 28/01/1995
1995 150.23 09/01/1996
1996 99.86 20/02/1997
1997 178.80 05/01/1998
1998 183.86 31/12/1998
1999 165.72 27/12/1999
2000 192.49 07/11/2000
2001 151.73 26/02/2002
2002 164.13 16/11/2002
2003 109.18 17/01/2004
2004 148.73 29/10/2004
EV1 plot
2 5 10 25 50 100 500
0
50
100
150
200
250
-2 0 2 4 6 8
AM
flo
w (
m3s-1
)
EV1 reduced variate, y
winter peak
summer peak
Return period (years)
2 5 10 25 50 100 500
0
50
100
150
200
250
-6 -4 -2 0 2 4 6
Logistic reduced variate, yL
LO plot
Return period (years)
AM
flo
w (
m3s-1
)
2 5 10 25 50 100 500
1.4
1.6
1.8
2
2.2
2.4
-2.5 -1.5 -0.5 0.5 1.5 2.5
log
10(A
M f
low
)
Standardised Normal variate, yN
LN plot
Return period (years)
Volume II Flood Frequency Estimation
194
Appendix H Probability-plot linear scores and curve patterns
Station number and name Grade #
years
Linear score Curve pattern
EV1 LO LN EV1 LO LN
06011 Fane at Moyles Mill A1 48 4 2 3 L2 U1 L2
06013 Dee at Charleville Weir A1 30 2 3 4 D1 S1 L2
06014 Glyde at Tallanstown A1 30 4 2 3 L1 U1 L2
06025 Dee at Burley Bridge A1 30 2 4 3 D1 L2 S1
06026 Glyde at Aclint Bridge A1 46 3 1 2 S1 S2 S2
06031 Flurry at Curralhir A2 18 2 2 3 U2 U2 U2
06070 Muckno L. at Muckno A1 24 4 4 5 L2 L2 L1
07006 Moynalty at Fyanstown A2 19 2 3 2 D2 D1 D2
07009 Boyne at Navan Weir A1 29 4 1 3 L2 D2X L2
07033 Blackwater at Virginia Hatchery A2 25 4 3 3 S1 U1 S1
08002 Delvin at Naul A1 20 4 1 2 L2 U1 U1
08005 Sluice at Kinsaley Hall A2 18 4 3 5 L2 U1 L1
08008 Broadmeadow at Broadmeadow A2 25 4 2 4 L2 U1 L2
08009 Ward at Balheary A1 11 1 1 2 U2 U2 U1
09001 Ryewater at Leixlip A1 48 5 2 5 L1 U1 L1
09002 Griffeen at Lucan A1 24 1 1 4 S2X U2 L1
09010 Dodder at Waldron’s Bridge A1 19 2 2 4 U2 U2 L2
10021 Shanganagh at Common’s Rd A1 24 4 2 3 L2 S1 L2
10022 Cabinteely at Carrickmines A1 18 4 3 4 L2 L2X L2
12001 Slaney at Scarawalsh A2 50 4 2 3 L2 U1 L2
14005 Barrow at Portarlington A2 48 3 2 2 L2 U1 U1
14006 Barrow at Pass Bridge A1 51 5 2 3 L1 U1 U1
14007 Stradbally at Derrybrock A1 25 4 2 3 L2 U1 S1
14009 Cushina at Cushina A2 25 4 3 4 S1 U1 S1
14011 Slate at Rathangan A1 26 3 3 3 L2 L2 L2
14013 Burrin at Ballinacarrig A2 50 4 3 4 S1 S2 S1
14018 Barrow at Royal Oak A1 51 3 3 4 L2 S1 S1
14019 Barrow at Levitstown A1 51 3 3 4 D1 S2 S1
14029 Barrow at Graiguenamanagh A2 47 3 4 5 D1 S1 L1
14034 Barrow at Bestfield A2 14 4 2 2 L2X U2X U2X
15001 Kings at Annamult A2 42 2 2 3 D2 S2 S1
15002 Nore at John’s Bridge A2 35 4 3 4 L2 U1 L2
15003 Dinan at Dinan Bridge A2 50 2 3 2 D2 S1 D2
15004 Nore at McMahons Bridge A2 51 5 3 5 L2 U1 L2
16001 Drish at Athlummon A2 33 5 4 5 L1 U1 L1
16002 Suir at Beakstown A2 51 3 2 4 L2X U1X L2X
Volume II Flood Frequency Estimation
195
Station number and name Grade #
years
Linear score Curve pattern
EV1 LO LN EV1 LO LN
16003 Clodiagh at Rathkennan A2 51 4 3 3 L2 S1 S1
16004 Suir at Thurles A2 48 5 4 5 L1 U1 L1
16005 Multeen at Aughnagross A2 30 5 2 2 L2 U1 U1
16008 Suir at New Bridge A2 51 2 4 3 D2 L2 D1
16009 Suir at Cahir Park A2 52 2 3 3 D2X D2X D1X
16011 Suir at Clonmel A1 52 3 2 3 D1 S1 S1
18004 Ballynamona at Awbeg A2 46 2 2 2 S2 S2 S2
18005 Funshion at Downing Bridge A2 50 4 2 3 L2X U2 L2X
19001 Owenboy at Ballea Upper A2 48 4 4 4 L2 L2 S1X
19020 Owennacurra at Ballyedmond A2 28 3 4 3 D2 S1 D1
23001 Galey at Inch Bridge A2 45 4 2 3 L2 U1 L2
23012 Lee at Ballymullen A2 18 2 1 1 U2 U2 U2
24002 Camogue at Gray’s Bridge A2 27 4 5 4 L2 L2 L2X
24008 Maigue at Castleroberts A2 30 3 5 4 L2 L2 L2X
24022 Mahore at Hospital A2 20 5 3 5 L1 U1 L2
24082 Maigue at Islandmore A2 28 2 5 3 D1 L2 D1
25006 Brosna at Ferbane A1 52 4 3 4 L2 L2 S1
25014 Silver at Millbrook Bridge A1 54 4 3 4 L2 U1X L2
25016 Clodiagh at Rahan A2 42 4 4 5 L2 L2 L2
25017 Shannon at Banagher A1 55 3 4 5 D1 L2 L1
25021 Little Brosna at Croghan A2 44 3 4 5 D1 L2 L2
25023 Little Brosna at Milltown A1 52 2 2 2 S2 S2 S1X
25025 Ballyfinboy at Ballyhooney A1 31 5 4 5 L1 L2 L2
25027 Ollatrim at Gourdeen Bridge A1 43 4 5 3 D1 L2 D1
25029 Nenagh at Clarianna A2 33 3 3 4 S2 S2 S1
25030 Graney at Scarriff Bridge A1 48 4 2 4 S1X U1 L2
25034 L. Ennell Trib at Rochfort A2 24 2 4 3 D2 L2 D1
25040 Bunow at Roscrea A2 20 4 3 4 L2 U1 L2
25044 Kilmastulla at Coole A2 33 3 3 3 S1 S1 S2
25124 Brosna at Ballynagore A2 18 3 4 2 D1 L2 D1
26002 Suck at Rookwood A2 53 2 1 2 S2X S2X L2X
26005 Suck at Derrycahill A2 51 4 4 5 L1 L2 L1
26006 Suck at Willsbrook A1 53 1 1 1 S2X S2X S2X
26007 Suck at Bellagill Bridge A1 53 5 3 5 L1 U1 L1
26008 Rinn at Johnston’s Bridge A1 49 4 3 3 L2 U1 L2
26009 Black at Bellantra Bridge A2 35 4 3 2 S1 S1 S2
26018 Owenure at Bellavahan A2 49 4 3 4 L2 S1 L2
26019 Camlin at Mullagh A1 51 4 3 3 L2 S1 L2
Volume II Flood Frequency Estimation
196
Station number and name Grade #
years
Linear score Curve pattern
EV1 LO LN EV1 LO LN
26020 Camlin at Argar Bridge A1 32 3 3 2 L2 S1 L2
26021 Inny at Ballymahon A2 30 1 2 1 D2 D1 D1
26022 Fallan at Kilmore A2 33 4 3 4 L2 S1 S1
26059 Inny at Finnea Bridge A1 17 2 2 1 D1 D1 D2X
27001 Claureen at Inch Bridge A2 30 4 3 3 L2 U1 L2
27002 Fergus at Ballycorey A1 51 4 2 4 L2 U1 L2
27003 Fergus at Corofin A2 48 4 4 4 L2 L2 L2
29001 Raford at Rathgorgin A1 40 3 4 4 D1 S1 S1
29004 Clarinbridge at Clarinbridge A2 32 4 3 3 L2 L2 S1
29011 Dunkellin at Kilcolgan Bridge A1 22 2 2 2 U1 S1 S1
29071 L. Cutra at Cutra A2 26 4 3 4 L2 S1 L2
30007 Clare at Ballygaddy A2 31 4 3 4 L2X L2X L2X
30061 Corrib Estuary at Wolfe Tone Bridge A2 33 1 1 2 U2 U2 U1
31002 Cashla at Cashla A1 26 2 1 1 U2X U2X U2X
32012 Newport at Newport Weir A2 24 3 5 3 D1 L2 L2
33070 Carrowmore L. at Carrowmore A1 28 4 3 4 L2X L2X L2X
34001 Moy at Rahans A2 36 4 3 4 L2X L2X L2
34003 Moy at Foxford A2 29 3 3 3 S1 S1 S2
34009 Owengarve at Curraghbonaun A2 33 4 4 5 L2 L2 L1
34011 Manulla at Gneeve Bridge A2 30 4 4 4 S1 L2 L2
34018 Castlebar at Turlough A1 27 4 3 3 L2 U1 U1
34024 Pollagh at Kiltimagh A2 28 3 4 3 D1 L2 L2
35001 Owenmore at Ballynacarrow A2 29 4 5 4 L2 L2 L2
35002 Owenbeg at Billa Bridge A2 34 3 5 4 D1 L2 L2
35005 Ballysadare at Ballysadare A2 55 4 2 4 S1 S1 L2
35071 L. Melvin at Lareen A2 30 4 3 4 L2 U1 L2
35073 L. Gill at Lough Gill A2 30 3 4 4 D1 L2 D1
36010 Annalee at Butlers Bridge A1 50 4 3 4 S1X S1X S1X
36012 Erne at Sallaghan A1 47 3 4 3 D1 L2 L2
36015 Finn at Anlore A1 33 2 2 2 U2 U2 U1
36018 Dronmore at Ashfield Bridge A1 50 4 4 4 L2 L2 L2
36019 Erne at Belturbet A2 47 2 3 3 D1 S1 S1
36021 Yellow at Kiltybarden A2 27 3 3 3 L2 U2 S1
36031 Cavan at Lisdarn A2 30 2 1 1 S2 S2 S2
39008 Leannan at Gartan Bridge A2 33 3 2 3 S1 S1 S1
39009 Fern O/L at Aghawoney A2 33 5 3 5 L1 U1 L1
Volume II Flood Frequency Estimation
197
Appendix I [This page and appendix are intentionally blank]
Volume II Flood Frequency Estimation
198
Appendix J Flood volumes in relation to convex EV1 plots
J1 Flood peaks and volumes for Station 07006 Moynalty at Fyanstown
Table J.1: Basic information for Station 07006 Moynalty at Fyanstown
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 176 1986 – 2004 19 27.93 26.73 0.206 -1.09
EV1 Probability Plot J.1: Station 07006 Moynalty at Fyanstown
Figure J.1: Hydrograph volumes for Station 07006 Rank 1, 2 and 4 AM flood peaks
Notes
Hydrographs available for three of four largest AM flood peaks (figures opposite );
1-day flood volumes are noticeably similar for these three events; at longer duration,
the Rank 2 AM event shows higher volume than the Rank 4 event, with the Rank 1
event giving the least volume among the three; the volume of the Rank 1 event is
only 50% of that of the Rank 2 event across 7, 14 and 30-day durations;
The Rank 2 and Rank 4 events display multiple flood peaks, explaining why their
hydrographs have greater volume than the unimodal Rank 1 event.
EV1
'93'95'04'00
2 5 10 25 50 100 500
0
5
10
15
20
25
30
35
40
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
w inter peak
summer peak
Volume of hydrographs of different year during max peak
0
5
10
15
20
25
1 2 7 14 30Days
Mil
l. c
u.
mete
r
2000
1995
1993
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Oct 1993
Rank 4 AM flood, Nov 2000
Rank 2 AM flood, Nov 1995
Volume II Flood Frequency Estimation
199
Hydrograph during Maximum Flood in the year 1993
0
5
10
15
20
25
30
35
40
02/10/1993 00:00 04/10/1993 00:00 06/10/1993 00:00 08/10/1993 00:00 10/10/1993 00:00 12/10/1993 00:00
Time
Dis
ch
arg
e (
m3/s
)Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1995
0
5
10
15
20
25
30
35
22/11/1995 00:00 24/11/1995 00:00 26/11/1995 00:00 28/11/1995 00:00 30/11/1995 00:00 02/12/1995 00:00 04/12/1995 00:00 06/12/1995 00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 2000
0
5
10
15
20
25
30
35
31/10/2000 00:00 02/11/2000 00:00 04/11/2000 00:00 06/11/2000 00:00 08/11/2000 00:00 10/11/2000 00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Rank 1 AM event,
6 October 1993
Rank 2 AM event,
29 November 1995
Rank 4 AM event,
6 November 2000
Volume II Flood Frequency Estimation
200
J2 Flood peaks and volumes for Station 15003 Dinan at Dinan Bridge
Table J.2: Basic information for Station 15003 Dinan at Dinan Bridge
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 298
1954 – 2004
(2001 missing) 50 150.76 143.57 0.196 -0.78
EV1 Probability Plot J.2: Station 15003 Dinan at Dinan Bridge
Figure J.2: Hydrograph volumes for Station 15003 Rank 1, 2 and 4 AM flood peaks
Notes
Hydrographs available for two largest AM flood peaks (figures opposite );
With only two data samples, it is speculative to infer how hydrograph volumes differ
with flood magnitude. The Rank 1 AM event corresponds to the August 1986 storm,
which was noted for its intensity and short duration.
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Aug 1986
Rank 2 AM flood, Nov 1997
'98 '68 '97 '85
Volume II Flood Frequency Estimation
201
Hydrograph during Maximum Flood in the year 1985
0
50
100
150
200
22/08/1986
00:00
23/08/1986
00:00
24/08/1986
00:00
25/08/1986
00:00
26/08/1986
00:00
27/08/1986
00:00
28/08/1986
00:00
29/08/1986
00:00
30/08/1986
00:00
Time
Dis
char
ge
(m3/
s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 1997
0
50
100
150
200
14/11/1997
00:00
15/11/1997
00:00
16/11/1997
00:00
17/11/1997
00:00
18/11/1997
00:00
19/11/1997
00:00
20/11/1997
00:00
21/11/1997
00:00
22/11/1997
00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Rank 1 AM event,
26 August 1986
Rank 2 AM event,
18 November 1997
Volume II Flood Frequency Estimation
202
J3 Flood peaks and volumes for Station 16008 Suir at New Bridge
Table J.3: Basic information for Station 16008 Suir at New Bridge
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 1120 1954 – 2004 51 92.32 90.66 0.125 -0.32
EV1 Probability Plot J.3: Station 16008 Suir at New Bridge
Figure J.3: Hydrograph volumes for four largest Station 16008 AM flood peaks
Notes
Hydrographs available for four largest AM flood peaks (figures opposite );
The flood hydrographs display similar shape and translate to generally similar flood
volumes;
There is very little difference among flood volumes for these four events, particularly
at the 1, 2 and 7-day durations.
See Section J4 below and analysis of Suir hydrographs in Section 5.8 of Volume III.
'87'83 '68 '60
EV1 Plot
2 5 10 25 50 100 500
0
20
40
60
80
100
120
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
w inter peak
summer peak
Volume of hydrographs of different years during max peak
0
50
100
150
200
250
1 2 7 14 30Days
Mil
l. c
u.
mete
r
1987
1983
1968
1960
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Dec 1960
Rank 4 AM flood, Feb 1988
Rank 2 AM flood, Dec 1968
Rank 3 AM flood, Feb 1984
Volume II Flood Frequency Estimation
203
Hydrograph during Maximum Flood in the year 1960
0
20
40
60
80
100
120
26/11/1960
00:00
28/11/1960
00:00
30/11/1960
00:00
02/12/1960
00:00
04/12/1960
00:00
06/12/1960
00:00
08/12/1960
00:00
10/12/1960
00:00
12/12/1960
00:00
14/12/1960
00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 1968
0
20
40
60
80
100
120
140
16/12/1968
00:00
18/12/1968
00:00
20/12/1968
00:00
22/12/1968
00:00
24/12/1968
00:00
26/12/1968
00:00
28/12/1968
00:00
30/12/1968
00:00
01/01/1969
00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 1983
0
20
40
60
80
100
120
29/01/1984
00:00
31/01/1984
00:00
02/02/1984
00:00
04/02/1984
00:00
06/02/1984
00:00
08/02/1984
00:00
10/02/1984
00:00
12/02/1984
00:00
14/02/1984
00:00
16/02/1984
00:00
Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 1987
0
20
40
60
80
100
120
26/01/1988
00:00
28/01/1988
00:00
30/01/1988
00:00
01/02/1988
00:00
03/02/1988
00:00
05/02/1988
00:00
07/02/1988
00:00
09/02/1988
00:00
11/02/1988
00:00
Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Rank 1 AM event,
4 December 1960
Rank 2 AM event,
24 December 1968
Rank 3 AM event,
6 February 1984
Rank 4 AM event,
4 February 1988
Volume II Flood Frequency Estimation
204
J4 Flood peaks and volumes for Station 16009 Suir at Cahir Park
Table J.4: Basic information for Station 16009 Suir at Cahir Park
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 1602 1953 – 2004 52 162.21 159.29 0.172 -0.41
EV1 Probability Plot J.4: Station 16009 Suir at Cahir Park
Figure J.4: Hydrograph volumes for four of five largest Station 16009 AM flood peaks
Notes
Hydrographs available for four of five largest AM flood peaks (figures opposite );
The hydrograph shapes are relatively similar, especially in their rising limbs;
It is possible to compare the hydrograph shapes with those typically seen at Station
16008 upstream (see Section J3); three flood events (Dec 1960, Dec 1968 and Feb
1990) are among the five largest floods at both stations;
The flood volumes are relatively similar in magnitude, particularly at the 1, 2 and 7-
day durations; at 30-day duration, the Rank 1 flood peak gives the largest flood
volume; the Rank 5 flood in Oct 2004 followed a notably dry autumn.
See Section J3 above and analysis of Suir hydrographs in Section 5.8 of Volume III.
'04'68 '00 '60 '89
EV1 Plot
2 5 10 25 50 100 500
0
50
100
150
200
250
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
Volume of hydrographs of different year during max peak
0
50
100
150
200
250
300
350
1 2 7 14 30Days
Mill cu
. m
ete
r
2004
2000
1960
1989
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Feb 1990
Rank 5 AM flood, Oct 2004
Rank 2 AM flood, Dec 1960
Rank 3 AM flood, Nov 2000
'87'83 '68 '60
EV1 Plot
2 5 10 25 50 100 500
0
20
40
60
80
100
120
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
w inter peak
summer peak
Volume II Flood Frequency Estimation
205
Hydrograph during Maximum Flood in the year 1989
0
50
100
150
200
250
30/01/1990
00:00
01/02/1990
00:00
03/02/1990
00:00
05/02/1990
00:00
07/02/1990
00:00
09/02/1990
00:00
11/02/1990
00:00
13/02/1990
00:00
15/02/1990
00:00
Time
Dis
char
ge
(m3/
s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 1960
0
50
100
150
200
250
25/11/1960
00:00
27/11/1960
00:00
29/11/1960
00:00
01/12/1960
00:00
03/12/1960
00:00
05/12/1960
00:00
07/12/1960
00:00
09/12/1960
00:00
11/12/1960
00:00
13/12/1960
00:00
15/12/1960
00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 2000
0
50
100
150
200
250
29/10/2000
00:00
31/10/2000
00:00
02/11/2000
00:00
04/11/2000
00:00
06/11/2000
00:00
08/11/2000
00:00
10/11/2000
00:00
12/11/2000
00:00
Time
Dis
char
ge
(m3/
s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Hydrograph during Maximum Flood in the year 2004
0
50
100
150
200
250
20/10/2004
00:00
22/10/2004
00:00
24/10/2004
00:00
26/10/2004
00:00
28/10/2004
00:00
30/10/2004
00:00
01/11/2004
00:00
03/11/2004
00:00
05/11/2004
00:00
Time
Dis
char
ge
(m3/
s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1w eek
Rank 1 AM event,
7 February 1990
Rank 2 AM event,
4 December 1960
Rank 3 AM event,
6 November 2000
Rank 5 AM event,
29 October 2004
Volume II Flood Frequency Estimation
206
J5 Flood peaks and volumes for Station 24013 Deel at Rathkeale
Table J.5: Basic information for Station 24013 Deel at Rathkeale (post-drainage)
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 426 1969 – 2004 36 109.60 108.31 0.152 0.05
EV1 Probability Plot J.5: Station 24013 Deel at Rathkeale (post-drainage)
Figure J.5: Hydrograph volumes for four largest Station 24013 AM flood peaks
Notes
Hydrographs available for four largest AM flood peaks (figures opposite );
Although the crest segments of the hydrographs have a characteristic shape, their
overall structure differs according to the complexity (i.e. multimodality) of the event;
The Rank 1 and Rank 2 AM floods have greater volume than the Rank 3 and Rank 4
AM floods;
At 1 and 2-day duration, the difference is minor; however, at 7-day duration, the
difference in volume is considerable [reflecting the multimodality of the hydrographs
in the Rank 1 and Rank 2 AM floods].
Volume of hydrograph of different years during max peak
0
10
20
30
40
50
60
70
80
1 2 7 14 30Days
Milli
on c
u.m
1988
1980
1998
1973
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Dec 1973
Rank 4 AM flood, Oct 1988
Rank 2 AM flood, Dec 1998
Rank 3 AM flood, Nov 1980
'73 '98 '80
'88 '94
Volume II Flood Frequency Estimation
207
Hydrograph during Maximum Flood in the year 1973
0
20
40
60
80
100
120
140
160
26/11/1973
00:00
27/11/1973
00:00
28/11/1973
00:00
29/11/1973
00:00
30/11/1973
00:00
01/12/1973
00:00
02/12/1973
00:00
03/12/1973
00:00
04/12/1973
00:00
05/12/1973
00:00
06/12/1973
00:00
Time
Dis
ch
arg
e (
m3/s
)Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1998
0
20
40
60
80
100
120
140
160
25/12/1998 00:00 27/12/1998 00:00 29/12/1998 00:00 31/12/1998 00:00 02/01/1999 00:00 04/01/1999 00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1980
0
20
40
60
80
100
120
140
160
28/10/1980 00:00 30/10/1980 00:00 01/11/1980 00:00 03/11/1980 00:00 05/11/1980 00:00 07/11/1980 00:00 09/11/1980 00:00
Time
Dis
char
ge
(m3/
s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1988
0
20
40
60
80
100
120
140
15/10/1988 00:00 17/10/1988 00:00 19/10/1988 00:00 21/10/1988 00:00 23/10/1988 00:00 25/10/1988 00:00 27/10/1988 00:00 29/10/1988 00:00
Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Rank 1 AM event,
1 December 1973
Rank 2 AM event,
30 December 1998
Rank 3 AM event,
2 November 1980
Rank 4 AM event,
22 October 1988
Volume II Flood Frequency Estimation
208
J6 Flood peaks and volumes for Station 24082 Maigue at Islandmore
Table J.6: Basic information for Station 24082 Maigue at Islandmore
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 764 1977 – 2004 28 140.01 135.47 0.264 -0.22
EV1 Probability Plot J.6: Station 24082 Maigue at Islandmore
Figure J.6: Hydrograph volumes for four largest Station 24082 AM flood peaks
Notes
Hydrographs available for four largest AM flood peaks (figures opposite );
Not much difference is shown among the volumes of hydrographs of these four
events for the 1 and 2-day volume;
The Rank 1 AM flood event yields the largest volume at all durations, but especially
at 7, 14 and 30-day duration.
'98 '88 '00
'89
EV1 Plot
2 5 10 25 50 100 500
0
50
100
150
200
250
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
Volume of hydrographs of different year during max peak
0
20
40
60
80
100
120
140
160
1 2 7 14 30Days
Mill
cu
. met
er
1998
1988
2000
1989
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Feb 1990
Rank 4 AM flood, Dec 1998
Rank 2 AM flood, Nov 2000
Rank 3 AM flood, Oct 1988
Volume II Flood Frequency Estimation
209
Hydrograph during Maximum Flood in the year 1989
0
50
100
150
200
250
01/02/1990 00:00 03/02/1990 00:00 05/02/1990 00:00 07/02/1990 00:00 09/02/1990 00:00 11/02/1990 00:00
Time
Dis
charg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 2000
0
20
40
60
80
100
120
140
160
180
200
31/10/2000 00:00 02/11/2000 00:00 04/11/2000 00:00 06/11/2000 00:00 08/11/2000 00:00 10/11/2000 00:00 12/11/2000 00:00
Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1988
0
20
40
60
80
100
120
140
160
180
200
14/10/1988
00:00
16/10/1988
00:00
18/10/1988
00:00
20/10/1988
00:00
22/10/1988
00:00
24/10/1988
00:00
26/10/1988
00:00
28/10/1988
00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1998
0
20
40
60
80
100
120
140
160
180
200
23/12/1998 00:00 25/12/1998 00:00 27/12/1998 00:00 29/12/1998 00:00 31/12/1998 00:00 02/01/1999 00:00 04/01/1999 00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Rank 1 AM event,
6 February 1990
Rank 2 AM event,
6 November 2000
Rank 3 AM event,
21 October 1988
Rank 4 AM event,
30 December 1998
Volume II Flood Frequency Estimation
210
J7 Flood peaks and volumes for Station 25017 Shannon at Banagher
Table J.7: Basic information for Station 25017 Shannon at Banagher
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 7989 1950 – 2004 55 407.68 413.25 0.203 0.18
EV1 Probability Plot J.7: Station 25017 Shannon at Banagher
Figure J.7: Hydrograph volumes for four of five largest Station 25017 AM flood peaks
Notes
Hydrograph unavailable for Rank 1 AM flood peak;
Hydrographs available for next four largest AM flood peaks (figures opposite );
note that Rank 5 AM flood event on 2 Feb 1995 is equalled by one on 28 Dec 1959;
Hydrographs are severely attenuated for this large catchment; there is little difference
in the volumes of the four events across any of the durations.
'94 '01 '89 '99'54
EV1 Plot
2 5 10 25 50 100 500
0
100
200
300
400
500
600
700
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
Volume of hydrograph of different years during max peak
0
200
400
600
800
1000
1200
1400
1 2 7 14 30Days
Mill. c
u. m
ete
r
1994
2001
1989
1999
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 2 AM flood, Dec 1999
Rank 5= AM flood, Feb 1995
Rank 3 AM flood, Feb 1990
Rank 4 AM flood, Feb 2002
Volume II Flood Frequency Estimation
211
Hydrograph during Maximum Flood in the year 1999
0
100
200
300
400
500
600
16/12/1999 00:00 21/12/1999 00:00 26/12/1999 00:00 31/12/1999 00:00 05/01/2000 00:00 10/01/2000 00:00
Time
Dis
ch
arg
e (
m3/s
)Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1989
0
100
200
300
400
500
600
22/01/1990 00:00 27/01/1990 00:00 01/02/1990 00:00 06/02/1990 00:00 11/02/1990 00:00 16/02/1990 00:00 21/02/1990 00:00 26/02/1990 00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 2001
0
100
200
300
400
500
600
03/02/2002
00:00
05/02/2002
00:00
07/02/2002
00:00
09/02/2002
00:00
11/02/2002
00:00
13/02/2002
00:00
15/02/2002
00:00
17/02/2002
00:00
19/02/2002
00:00
21/02/2002
00:00
23/02/2002
00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1994
0
100
200
300
400
500
600
21/01/1995 00:00 26/01/1995 00:00 31/01/1995 00:00 05/02/1995 00:00 10/02/1995 00:00 15/02/1995 00:00
Time
Dis
ch
arg
e (
m3/s
)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week Rank 5= AM event,
2 February 1995
Rank 3 AM event,
9 February 1990
Rank 2 AM event,
27 December 1999
Rank 4 AM event,
13 February 2002
Volume II Flood Frequency Estimation
212
J8 Flood peaks and volumes for Station 25021 Little Brosna at Croghan
Table J.8: Basic information for Station 25021 Little Brosna at Croghan
Nominal area Period of AM
flow data # years
Median Mean CV Hazen skewness
km2 m
3s
-1 m
3s
-1
2 493 1961 – 2004 44 28.58 28.03 0.141 -0.13
EV1 Probability Plot J.8: Station 25021 Little Brosna at Croghan
Figure J.8: Hydrograph volumes for four largest Station 25021 AM flood peaks
Notes
Hydrographs available for four largest AM flood peaks (figures opposite );
The hydrographs are rather intricate for this station, with shorter-term fluctuations
superposed on the main body of the hydrograph;
Although the Rank 1 AM flood has the largest peak discharge, it gives a considerably
smaller flood volume than the other high-ranking AM floods, most notably at 7, 14
and 30-day durations. [Editorial note: It appears that a rating curve adopted in the
FSU has been wrongly applied to the 1961 and 1962 AM water levels: ignoring a
rating change on 14 June 1963. It is likely that the FSU flood peak of 35.8 m3s
-1 on
5 Nov 1962 is greatly exaggerated. This would account for the anomalous shape of
the crest segment of the hydrograph shown opposite for the Rank 1 AM event.]
'99'00 '62
EV1 Plot
2 5 10 25 50 100 500
'94
0
5
10
15
20
25
30
35
40
45
-2 -1 0 1 2 3 4 5 6 7EV1 y
AM
F(m
3/s
)
w inter peak
summer peak
Volume of hydrographs of different year during max peak
0
10
20
30
40
50
60
70
1 2 7 14 30Days
Mill cu
. m
ete
r
1994
1999
2000
1962
Flo
od v
olu
me
in 1
06 m
3
Duration across which flood volume evaluated
Rank 1 AM flood, Nov 1962
Rank 4 AM flood, Feb 1995
Rank 2 AM flood, Nov 2000
Rank 3 AM flood, Feb 2002
Volume II Flood Frequency Estimation
213
Hydrograph during Maximum Flood in the year 1962
0
5
10
15
20
25
30
35
40
31/10/1962 00:00 02/11/1962 00:00 04/11/1962 00:00 06/11/1962 00:00 08/11/1962 00:00 10/11/1962 00:00 12/11/1962 00:00
Time
Dis
char
ge (
m3/
s)Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 2000
0
5
10
15
20
25
30
35
40
29/10/2000
00:00
31/10/2000
00:00
02/11/2000
00:00
04/11/2000
00:00
06/11/2000
00:00
08/11/2000
00:00
10/11/2000
00:00
12/11/2000
00:00
14/11/2000
00:00Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1999
0
5
10
15
20
25
30
35
40
16/12/1999
00:00
18/12/1999
00:00
20/12/1999
00:00
22/12/1999
00:00
24/12/1999
00:00
26/12/1999
00:00
28/12/1999
00:00
30/12/1999
00:00
01/01/2000
00:00
03/01/2000
00:00
Time
Disc
harg
e (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Hydrograph during Maximum Flood in the year 1994
0
5
10
15
20
25
30
35
40
21/01/1995
00:00
23/01/1995
00:00
25/01/1995
00:00
27/01/1995
00:00
29/01/1995
00:00
31/01/1995
00:00
02/02/1995
00:00
04/02/1995
00:00
Time
Dis
char
ge (m
3/s)
Red zone= Vol. of 1day
Yellow zone= Vol. of 2day
Blue zone= Vol. of 1week
Rank 1 AM event,
5 November 1962
Rank 2 AM event,
6 November 2000
Rank 3 AM event,
25 December 1999
Rank 4 AM event,
28 January 1995
[This may not be Rank 1 AM flood.
See notes on previous page.]
Volume II Flood Frequency Estimation
214
Appendix K Seasonal distribution of annual maximum floods
Table K.1: Percentage of AM floods occurring in winter half-year (Oct-Mar)
Station
number
%
AMs
Oct–
Mar
Station
number
%
AMs
Oct–
Mar
Station
number
%
AMs
Oct–
Mar
Station
number
%
AMs
Oct–
Mar
Station
number
%
AMs
Oct–
Mar
Station
number
%
AMs
Oct–
Mar
01041 91 09002 68 16009 90 24012 88 26010 91 32012 88
01055 89 09010 63 16011 90 24013 87 26012 98 33001 68
03051 100 09035 44 16012 92 24022 100 26014 100 33070 89
06011 100 10002 83 16013 88 24030 88 26017 94 34001 97
06013 90 10021 63 16051 85 24082 82 26018 96 34003 97
06014 97 10022 60 18001 90 25001 73 26019 88 34004 92
06021 94 10028 69 18002 94 25002 76 26020 88 34007 85
06025 80 11001 91 18003 96 25003 78 26021 90 34009 91
06026 91 12001 90 18004 85 25004 87 26022 88 34010 58
06030 63 12013 73 18005 84 25005 87 26058 83 34011 97
06031 94 14005 90 18006 93 25006 92 26059 96 34018 100
06033 80 14006 88 18016 83 25011 78 26108 100 34024 90
06070 85 14007 76 18048 83 25014 81 27001 73 34029 86
07002 96 14009 80 18050 92 25016 93 27002 96 35001 90
07003 83 14011 81 19001 92 25017 98 27003 88 35005 93
07004 98 14013 82 19014 90 25020 91 27070 79 35011 89
07005 91 14018 96 19016 100 25021 91 28001 82 35071 90
07006 84 14019 92 19020 89 25023 71 29001 85 35073 87
07007 91 14029 98 19031 89 25025 90 29004 81 36010 96
07009 90 14033 82 19046 67 25027 88 29007 82 36011 96
07010 93 14034 82 20001 93 25029 85 29011 86 36012 98
07011 98 15001 88 20002 94 25030 92 29071 83 36015 94
07012 94 15003 84 20006 84 25034 83 30001 94 36018 92
07033 84 15004 92 22006 86 25038 82 30004 97 36019 96
07041 100 15005 90 22009 79 25040 75 30005 86 36021 73
08002 76 15012 75 22035 100 25044 83 30007 94 36027 93
08003 78 16001 88 23001 87 25124 94 30012 100 36031 90
08005 67 16002 90 23002 80 25158 72 30021 92 36071 75
08007 73 16003 82 23012 83 26002 91 30031 100 38001 73
08008 84 16004 90 24001 88 26005 94 30037 57 39001 83
08009 67 16005 87 24002 81 26006 94 30061 91 39008 82
08011 87 16006 82 24004 87 26007 94 31002 81 39009 88
08012 68 16007 84 24008 87 26008 92 31072 69 202 FSU
catchments 09001 77 16008 90 24011 85 26009 89 32011 65
Volume II Flood Frequency Estimation
215
Table K.2: Month of maximum recorded flood in AM series
Station
number
Month
of
series
max
Station
number
Month
of
series
max
Station
number
Month
of
series
max
Station
number
Month
of
series
max
Station
number
Month
of
series
max
Station
number
Month
of
series
max
01041 Dec 09002 Dec 16009 Dec 24012 Dec 26010 Nov 32012 Dec
01055 Dec 09010 Dec 16011 Nov 24013 Dec 26012 Feb 33001 Sep
03051 Dec 09035 Oct 16012 Nov 24022 Dec 26014 Nov 33070 Oct
06011 Dec 10002 Nov 16013 Nov 24030 Aug 26017 Dec 34001 Oct
06013 Dec 10021 May 16051 Jan 24082 Dec 26018 Dec 34003 Oct
06014 Dec 10022 May 18001 Oct 25001 Dec 26019 Oct 34004 Oct
06021 Dec 10028 Oct 18002 Oct 25002 Oct 26020 Dec 34007 Oct
06025 Nov 11001 Aug 18003 Dec 25003 Dec 26021 Dec 34009 Nov
06026 Dec 12001 Nov 18004 Jan 25004 Dec 26022 Dec 34010 Jun
06030 Nov 12013 Nov 18005 Dec 25005 Nov 26058 Jan 34011 Oct
06031 Dec 14005 Dec 18006 Nov 25006 Dec 26059 Jan 34018 Dec
06033 Dec 14006 Dec 18016 Jan 25011 Dec 26108 Jan 34024 Nov
06070 Nov 14007 Feb 18048 Dec 25014 Dec 27001 Jan 34029 Jan
07002 Dec 14009 Feb 18050 Dec 25016 Dec 27002 Dec 35001 Oct
07003 Nov 14011 Feb 19001 Jan 25017 Dec 27003 Dec 35005 Oct
07004 Dec 14013 Dec 19014 Dec 25020 Jan 27070 Dec 35011 Oct
07005 Oct 14018 Feb 19016 Jan 25021 Dec 28001 Dec 35071 Jan
07006 Dec 14019 Feb 19020 Dec 25023 Nov 29001 Jan 35073 Dec
07007 Dec 14029 Dec 19031 Feb 25025 Jan 29004 Jan 36010 Dec
07009 Nov 14033 Dec 19046 Mar 25027 Dec 29007 Jan 36011 Dec
07010 Nov 14034 Feb 20001 Oct 25029 Feb 29011 Jan 36012 Dec
07011 Dec 15001 Dec 20002 Dec 25030 Dec 29071 Dec 36015 Oct
07012 Jan 15003 Aug 20006 Dec 25034 Dec 30001 Dec 36018 Dec
07033 Jan 15004 Feb 22006 Dec 25038 Feb 30004 Nov 36019 Dec
07041 Nov 15005 Dec 22009 Dec 25040 Oct 30005 Dec 36021 Oct
08002 Aug 15012 Feb 22035 Jan 25044 Dec 30007 Feb 36027 Jan
08003 Nov 16001 Dec 23001 Dec 25124 Feb 30012 Dec 36031 Oct
08005 Aug 16002 Dec 23002 Aug 25158 Jan 30021 Jan 36071 Apr
08007 Nov 16003 Dec 23012 Aug 26002 Oct 30031 Dec 38001 Sep
08008 Jun 16004 Dec 24001 Oct 26005 Dec 30037 Nov 39001 Sep
08009 Dec 16005 Sep 24002 Feb 26006 Dec 30061 Jan 39008 Dec
08011 Aug 16006 Dec 24004 Dec 26007 Nov 31002 Oct 39009 Dec
08012 Dec 16007 Sep 24008 Dec 26008 Nov 31072 Jul 202 FSU
catchments 09001 Dec 16008 Dec 24011 Dec 26009 Oct 32011 Sep
Volume II Flood Frequency Estimation
216
Appendix L Distance metrics for pooling-group construction
L1 Introduction
This appendix describes work underlying the distance metric recommended in Section 7.3 for
pooling-group formation:
2
lnBFI
ji
2
lnSAAR
ji
2
lnAREA
ji
ijσ
lnBFIlnBFI
σ
lnSAARlnSAAR
σ
lnAREAlnAREAd
[Equation 7.2]
Under the Region Of Influence (ROI) approach (Burn, 1990), the aim is to select catchments
that are hydrologically similar to the subject catchment. Catchments are recruited to the
pooling group (Reed et al., 1999) using a distance metric which represents their closeness to
the subject catchment in “catchment descriptor” space. There are many possible
arrangements and variations when constructing the catchment-descriptor space in which to do
the pooling, e.g.
Descriptors other than AREA, SAAR and BFI might have been used;
A different number of descriptors could have been used (e.g. two or four descriptors
rather than three);
Different transformations might have been applied;
Components in the distance metric could have been assigned different weights.
L2 Notation
The general form of the distance metric used for selecting members of a pooling group is:
n
1k
2
jk,ik,kij XXWd L.1
where n is the number of catchment descriptors, Xk,i is the normalised value of the
kth
catchment descriptor at the ith
site and Wk is the weight applied to descriptor k reflecting
the relative importance of that catchment descriptor. The subscript j applies to the subject
site and the subscript i applies to the available gauged sites.
In choosing a distance measure dij, a decision has to be made about which catchment
descriptors are to be included in the distance measure, whether logarithms or other
transformations are to be used, and what weightings are to be applied.
Following Jakob et al. (1999), catchment descriptors are normalised by dividing by the
sample standard deviation after any transformation. Thus, the term used to represent
catchment size is Xk,i = ℓnAREAi/σℓnAREA.
L3 Selecting variables to define the distance metric
The research considered use of the physical catchment descriptors AREA, SAAR, BFI and
FARL. The first three were found useful by Jakob et al. (1999). The fourth descriptor
responds to a specific finding in Section 8.2 that pooled growth factors in Ireland appear
sensitive to FARL. Some UK work (Kjeldsen et al., 2008) has also found FARL relevant in
this context.
Volume II Flood Frequency Estimation
217
L4 Statistics to help in choosing a good distance metric
The objective is to find a distance metric which leads to pooling groups which are most
homogeneous (i.e. least heterogeneous), thereby exploiting the national resource of AM flow
data effectively. The search is supported by three statistics:
H1, heterogeneity measure based on dissimilarity of L-CVs across pooling group;
H2, heterogeneity measure based on dissimilarity of L-CVs and L-skewnesses;
PUM, the pooled uncertainty measure (see next section).
The heterogeneity measures H1 and H2 are those put forward by Hosking and Wallis (1997).
L5 Pooled uncertainty measure, PUM
The pooled uncertainty measure (PUM) is a weighted average of the differences between site
and pooled growth factors measured on a logarithmic scale (Jakob et al., 1999):
2
M
1i
i
M
1i
Tp
Ti
Tlong
long
ii
n
lnxlnxn
PUM
L.2
where Mlong is the number of long-record sites and (for the ith
site): ni denotes record length,
iTx is the T-year single-site growth factor and iTpx is the T-year pooled growth factor.
PUM is a measure of how effective a pooling method is at identifying a homogeneous region.
A good pooling method will yield low values of PUM.
L6 Application to flood data at 90 A1 + A2 stations
Ninety stations graded A1 or A2 were used in the analysis. PUM has been evaluated at
100-year return period for those 85 (of the 90) stations having a record of 20 years or longer.
[Editorial note: The pooling group for each of the 85 stations used the distance metric under
test to select from the set of 90 catchments.] Stations were added to the pooling group
according to the 5T rule of Section 7.3.3, i.e. until there are at least 5 100 = 500
station-years of AM flow data in the pooling group.
The Generalised Extreme Value (GEV) distribution was used to calculate the pooled and
single-site growth factors. In making such experiments, Jakob et al. (1999) exclude the
subject site from its own pooling group when evaluating PUM. Here, PUM was evaluated
both with and without the subject site. However, only results for the latter case are reported.
Eight combinations of the four variables AREA, SAAR, BFI and FARL have been tested for
use in Equation L.1. The eight pooling schemes are:
Scheme 1: ℓnAREA
Scheme 2: ℓnAREA, ℓnSAAR
Scheme 3: ℓnAREA, ℓnSAAR, BFI
Scheme 4: ℓnAREA, ℓnSAAR, BFI, FARL
Volume II Flood Frequency Estimation
218
Scheme 5: ℓnSAAR
Scheme 6: BFI
Scheme 7: ℓnAREA, BFI
Scheme 8: ℓnAREA, ℓnSAAR, ℓnBFI
Initially, all weights Wk in Equation L.1 were set to unity. The dataset is summarised in
Table L.1. [Editorial note: It would appear that the dataset is rather similar to that used in
Chapter 8 and includes five relatively heavily urbanised catchments (URBEXT > 0.20 and
highlighted in Table 8.1) that might have warranted omission.]
Table L.1: Summary of AM flow dataset used in dij study
#
stations
Shortest record
length (years)
Longest record
length (years)
Mean record
length (years)
# station-years of
AM flow data
90 18 55 37.47 3372
Table L.2 summarises the mean variation in PUM100, H1 and H2 values achieved with the
eight different pooling schemes. Small values of PUM100, H1 and H2 indicate superior
performance. The numerical measures vary relatively little between methods except that
Scheme 6 (using BFI alone) is consistently the least-effective.
Table L.2: Mean values of PUM100, H1 and H2 for various pooling schemes
Sch
eme
Variables in distance metric dij
Mean value of
PUM100 H1 H2
1 ℓnAREA 0.1956 6.167 2.770
2 ℓnAREA, ℓnSAAR 0.1929 5.228 2.768
3 ℓnAREA, ℓnSAAR, BFI 0.1958 5.438 2.823
4 ℓnAREA, ℓnSAAR, BFI, FARL 0.1966 5.016 2.958
5 ℓnSAAR 0.1913 5.528 2.726
6 BFI 0.2053 7.141 3.214
7 ℓnAREA, BFI 0.2044 6.135 2.950
8 ℓnAREA, ℓnSAAR, ℓnBFI 0.1958 5.558 2.877
[Editorial note: Box-plots L.1 show the 100-year PUM values. Comparison with the results
presented in Table L.2 indicates that the ranking of methods would be appreciably altered
were this based on median rather than mean values of PUM100. Corresponding box-plots (not
included) indicate that rankings in terms of H1 and H2 would also change. Were median
values of the three measures used to rank the methods overall, the performance of Scheme 3
– which is recommended below – would in fact be comparable with or better than the other
seven schemes.]
Volume II Flood Frequency Estimation
219
Box-plots L.1: 100-year PUM values for eight formulations of distance metric dij
L7 Discussion
The use of ℓnAREA and ℓnSAAR (Scheme 2) might be warranted in Irish conditions.
However, any margin by which it outperforms (e.g.) Scheme 3 is minor and may not be
statistically significant. From a hydrological standpoint there is merit in including a third
component – i.e. additional to ℓnAREA and ℓnSAAR – that reflects a catchment feature
known to influence flood growth rates. Candidates include catchment permeability (indexed
by BFI) and storage attenuation (indexed by FARL). An extended investigation explored this
and sought to optimise the terms and weights (Wk) in Equation L.1.
L8 Alternative weightings of the recommended distance metric
It was found that weights of 1.7, 1.0 and 0.2 applied respectively to ℓnAREA, ℓnSAAR and
BFI in Scheme 3 offered a small but useful improvement over the unweighted schemes
reported in Table L.2. An alternative distance metric is therefore:
2
BFI
ji
2
SAAR
ji
2
lnAREA
ji
ijσ
BFIBFI0.2
σ
lnSAARlnSAAR
σ
lnAREAlnAREA1.7d
L.3
This led to values of mean PUM100, H1 and H2 of 0.1889, 5.20 and 2.62 respectively.
[Editorial note: Comparison with the values in Table L.2 confirms that introduction of the
additional weights leads to improved performance. It should be noted that this has been
achieved at the expense of an additional two parameters. Further analysis might confirm the
significance of this improvement. Any new recommendation on the choice of distance metric
is likely to be influenced by feedback from practitioners on the use of Equation 7.2 in
practical cases.]
Further details and additional research are presented by Das and Cunnane (2011 and 2012),
who favour the same distance metric but with a further-altered set of weights:
2
BFI
ji
2
SAAR
ji
2
lnAREA
ji
ijσ
BFIBFI0.1
σ
lnSAARlnSAAR
σ
lnAREAlnAREA1.5d
L.4
Pooled Uncertainty Measure(PUM)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
lnA lnA+lnS lnA+lnS+B lnA+lnS+B+F lnS B lnA+B lnA+lnS+lnB
PU
M100
ℓnAREA ℓnAREA ℓnAREA ℓnAREA ℓnSAAR BFI ℓnAREA ℓnAREA
ℓnSAAR ℓnSAAR ℓnSAAR BFI ℓnSAAR
BFI BFI BFI
FARL
Scheme Scheme Scheme Scheme Scheme Scheme Scheme Scheme
1 2 3 4 5 6 7 8
PU
M10
0
Volume II Flood Frequency Estimation
220
[This page is intentionally blank]