rood

os

c School of Geography, University of Nottingham, University Park, Nottingham NG7 2RD, UK

a r t i c l e i n f o

Article history:Received 18 July 2012Received in revised form 10 December 2012

and aquatic ecosystems (e.g. Poff et al., 2002; Matthews andQuesne, 2009). Evaluation of the hydrological impacts of climatechange is most commonly based upon driving a hydrologicalmodel with climatic projections derived from general circulationmodels (GCMs) forced with alternative emissions scenarios. This

rasson et al., 2004), regional (e.g. Arnell, 1999a) and globalscales (e.g. Arnell, 2003; Nohara et al., 2006; Gosling et al.,2010).

There are a range of uncertainties that are introduced through-out these climate change hydrological impact assessments (Nawazand Adeloye, 2006; Gosling et al., 2011a). Uncertainty is rstly re-lated to the denition of the greenhouse gas emissions scenariosused to force the GCMs. Secondly, uncertainty is associated withthese GCMs, with climate model structural uncertainty causing dif-ferent models to produce different climate projections for the same

Corresponding author. Tel.: +44 207 679 0589; fax: +44 0207 679 0565.E-mail addresses: j.r.thompson@ucl.ac.uk (J.R. Thompson), amanda.green.09@

ucl.ac.uk (A.J. Green), daniel.kingston@geography.otago.ac.nz (D.G. Kingston),

Journal of Hydrology 486 (2013) 130

Contents lists available at

Journal of H

.e lssimon.gosling@nottingham.ac.uk (S.N. Gosling).structures, process representations and PET methods (Linacre for MIKE SHE and SLURP, PenmanMon-teith for Mac-PDM.09).

2013 Elsevier B.V. All rights reserved.

1. Introduction

It is widely acknowledged that climate change will impact theglobal hydrological cycle (Kundzewicz et al., 2007; Arnell andGosling, in press), with implications for human use of water re-sources (Bates et al., 2008; Gosling et al., 2011b; Gosling, 2012)

approach has been used in the assessment of climate change im-pacts for hydrological systems that vary in scale from small wet-lands located within wider catchments (e.g. Thompson et al.,2009), through small to medium sized catchments (e.g. Chunet al., 2009; Thompson, 2012), major river basins (e.g. Conwayand Hulme, 1996; Nijssen et al., 2001), to the national (e.g. And-Accepted 13 January 2013Available online 4 February 2013This manuscript was handled byKonstantine P. Georgakakos, Editor-in-Chief,with the assistance of Raymond Najjar,Associate Editor

Keywords:MekongMIKE SHESLURPMac-PDMClimate changeUncertainty0022-1694/$ - see front matter 2013 Elsevier B.V. Ahttp://dx.doi.org/10.1016/j.jhydrol.2013.01.029s u m m a r y

Hydrological model-related uncertainty is often ignored within climate change hydrological impactassessments. A MIKE SHE model is developed for the Mekong using the same data as an earlier semi-dis-tributed, conceptual model (SLURP). The model is calibrated and validated using discharge at 12 gaugingstations. Two sets of climate change scenarios are investigated. The rst is based on a 2 C increase in glo-bal mean temperature (the hypothesised threshold of dangerous climate change), as simulated by sevenGCMs. There are considerable differences in scenario discharge between GCMs, ranging from catchment-wide increases in mean discharge (up to 12.7%; CCCMA CGCM31, NCAR CCSM30), decreases (up to 21.6%in the upper catchments; CSIRO Mk30, IPSL CM4), and spatially varying responses (UKMO HadCM3 andHadGEM1, MPI ECHAM5). Inter-GCM differences are largely driven by differences in precipitation. Thesecond scenario set (HadCM3, increases in global mean temperature of 16 C) shows consistentlygreater discharge (maximum: 28.7%) in the upper catchment as global temperature increases, primarilydue to increasing precipitation. Further downstream, discharge is strongly inuenced by increasing PET,which outweighs impacts of elevated upstream precipitation and causes consistent discharge reductionsfor higher temperatures (maximum: 5.3% for the main Mekong). MIKE SHE results for all scenarios arecompared with those from the SLURP catchment model and the Mac-PDM.09 global hydrological model.Although hydrological model-related uncertainty is evident, its magnitude is smaller than that associatedwith choice of GCM. In most cases, the three hydrological models simulate the same direction of changein mean discharge. Mac-PDM.09 simulates the largest discharge increases when they occur, which isresponsible for some differences in direction of change at downstream gauging stations for some scenar-ios, especially HadCM3. Inter-hydrological model differences are likely attributed to alternative modelaWetland Research Unit, UCL Department of Geography, University College London, Gower Street, London WC1E 6BT, UKbDepartment of Geography, University of Otago, PO Box 56, Dunedin, New ZealandAssessment of uncertainty in river ow pusing multiple GCMs and hydrological m

J.R. Thompson a,, A.J. Green a, D.G. Kingston b, S.N. G

journal homepage: wwwll rights reserved.jections for the Mekong Riverels

ling c

SciVerse ScienceDirect

ydrology

evier .com/ locate / jhydrol

under 1000 mm on the Korat Plateau to over 3200 mm in moun-tainous parts of Laos. In the upper parts of the basin within the Ti-

of Hemissions scenario. Downscaling of GCM projections to ner spa-tial and temporal scales appropriate for hydrological modellingpresents a third source of uncertainty.

The nal source of uncertainty in climate change hydrologicalimpact assessments is associated with the hydrological modelsused to translate climatological changes to hydrological impacts.Alternative hydrological models that produce acceptable resultsfor an observed baseline period may respond differently whenforced with the same climate change scenario (Gosling and Arnell,2011; Haddeland et al., 2011). Hydrological models include fullydistributed, physically based models (e.g. Refsgaard et al., 2010),which can detail a range of processes, potentially at a very ne spa-tial scale (e.g. Thompson et al., 2004; Hammersmark et al., 2008),but which require an extensive range of data. Semi-distributed orlumped models (e.g. Arnold et al., 1998) adopt a more conceptualapproach for process description, whilst global hydrological mod-els (e.g. Dll et al., 2003; Gosling and Arnell, 2011) employ largemodel grid sizes and simplied process descriptions.

Investigation of uncertainty within climate change hydrologicalimpact assessments has often focused on GCM uncertainty. Meth-ods have included perturbed physics ensembles, in which pertur-bation of the parameterisations within a single or a series ofGCMs are employed to provide many realizations of climate thatcan be used within hydrological modelling studies. The 2009 UKClimate Projections (UKCP09, Jenkins et al., 2009), for example,are based on a large perturbed physics ensemble using the Met Of-ce Hadley Centres HadCM3 GCM and results from another 12GCMs. Simple conceptual hydrological models of different wetlandtypes have been driven with all the realizations from the UKCP09projections to provide regionalized frequency distributions of thehydrological and, in turn, ecological impacts of climate change(Acreman et al., 2012). Similarly, Prudhomme et al. (2003) used aMonte Carlo approach to dene 25,000 climate scenarios for simu-lation of a number of UK catchments using a conceptual rainfallrunoff model. However, these approaches are computationallyintensive, especially with complex hydrological models. A morecommon approach to investigating GCM uncertainty is to use arange of projections for the same emissions scenario derived froman ensemble of GCMs (e.g. Meehl et al., 2007).

This approach was used within the UK Natural Environment Re-search Council QUEST-GSI (Global Scale Impacts) project (http://www.met.reading.ac.uk/research/quest-gsi/), in which hydrologi-cal models for catchments around the world were employed to as-sess the impacts of a consistent set of climate change scenarios. Foreach catchment, only one hydrological model was utilised, givingno indication of the impact of choice or structure of hydrologicalmodel upon climate change impacts. Prudhomme and Davies(2009) suggest that this is relatively common, with hydrologicalmodel uncertainty often being ignored within impact studies.Where different hydrological models of the same catchment exist,they have invariably been developed by different institutions andfor different purposes, and are therefore rarely used to assess thesame climate change scenarios. The impact of hydrological mod-el-derived uncertainty on climate change impacts may not, how-ever, be negligible (e.g. Dibike and Coulibaly, 2005; Haddelandet al., 2011; Hagemann et al., 2012). The QUEST-GSI project pro-vided an initial assessment of some of these issues through a com-parison of catchment hydrological model results with those of theMac-PDM.09 global hydrological model (Gosling and Arnell, 2011)for the same GCM/global mean temperature change scenarios(Gosling et al., 2011a). The comparison was, however, limited tothe catchment outlet, whereas in some of the larger catchments,hydrological models provided distributed responses in river ow

2 J.R. Thompson et al. / Journalfrom major sub-catchments.The current study builds upon this earlier work, with a particu-

lar focus on one of the QUEST-GSI catchments. The Mekong is onebetan Highlands and Yunnan, precipitation falls both as rain andsnow, the latter in particular during the relatively dry Novem-berMarch period when snow covers approximately 5% of the totalMekong catchment (Kiem et al., 2005). Snowmelt is responsible forthe initial rise in the annual ood season within the Lancang sub-catchment. The Mekong River begins to rise in May and peak dis-charges are attained between August and October, after which theydecline, reaching their lowest levels in MarchApril.

The Mekong has been identied as a hotspot of global change(Takeuchi, 2008). Rapid development and population growth haveresulted in the previously discussed deforestation. Linked to thesedevelopments are increasing competition for water, contaminationofwater by agriculture, industry and settlements, andunsustainableuse of resources such as sheries, which currently sustain aroundof the worlds major rivers. It ows through a catchment with spa-tially variable climate, topography and land cover and supports alarge and growing human population. Climate change may exacer-bate the already signicant changes resulting from developmentwithin the catchment. A model of the Mekong is created usingMIKE SHE, a fully distributed hydrological modelling system com-bining both physically based and conceptual components (see be-low). As far as possible, this model employs the same data as themodel of the basin used within the QUEST-GSI project (Kingstonet al., 2011). A more robust calibration enables the investigationof the response to the QUEST-GSI climate change scenariosthroughout the catchment. These results are subsequently com-pared to those that are available for the earlier catchment modelas well as those from the Mac-PDM.09 global hydrological model.

2. Materials and methods

2.1. The Mekong catchment

The Mekong is the largest river in southeast Asia. It is theworlds eighth largest in terms of annual discharge (475 km3),12th longest (c. 4,350 km) and 21st largest by drainage area(795,000 km2) (e.g. Kiem et al., 2008). Rising in the Tibetan High-lands at an elevation of over 5100 m, it passes through China, Bur-ma, Laos, Thailand, Cambodia and Vietnam (Fig. 1). Majortributaries include the Chi and Mun, which drain the Korat Plateauof eastern Thailand and join the Mekong upstream of Pakse, andthe Se Kong and Sre Pok, which rise in Vietnams Central Highlandsand ow into the Mekong at Stung Treng. Further downstream, theriver both provides water to and drains the Tonle Sap Lake,depending upon the season, before discharging into the South Chi-na Sea via the distributaries of the Mekong Delta.

In the upper catchment (the Lancang) the river and its tributar-ies ow through narrow, steep gorges. Land cover consists of tun-dra and montane semi-desert. Further downstream, below ChiangSaen, the river becomes largely navigable except for a few water-falls. Natural vegetation is dominated by evergreen and deciduousforest (Ishidaira et al., 2008). Rapid economic development and agrowing population (currently 60 million and projected to increaseto 90 million by 2025 (MRC, 2003)) have driven the expansion ofagriculture and consequent deforestation, leading to a large reduc-tion in forest extent (Nobuhiro et al., 2008).

The dominant climatic inuence on the Mekong is the Asianmonsoon. The rainy southwest monsoon begins in mid-May andextends into early-October. Over 90% of annual precipitation fallswithin this period (Kite, 2001). Annual precipitation ranges from

ydrology 486 (2013) 130300 million people within and outside the catchment (MRC, 2003).Dams inupstreamparts of the catchment,mostnotably theManwan(constructed in 1993) and the Dachaoshan (constructed in 2001),

l of HJ.R. Thompson et al. / Journahave been implicated in changes in ow regime, sediment ows andsheries (Hapuarachchi et al., 2008; Li and He, 2008; Kummu et al.,2010; Wang et al., 2011) and more are planned which will likelyexacerbate these changes (Stone, 2010).

Whilst useful and informative modelling studies have beenundertaken for the Mekong, some have only adopted future cli-mate projections from a single GCM or ensemble means, thereby

Fig. 1. The Mekong catchment and its representation within the MIKE SHE model inclmeteorological inputs. The gauging stations within the MIKE 11 river network that werydrology 486 (2013) 130 3not addressing climate model structural uncertainty (Kingstonet al., 2011). For example, Kiem et al. (2008) only used the JapaneseMeteorological Agency GCM and the IPCC SRES A1b scenario with-in a gridded hydrological model of the catchment, whilst Ishidairaet al. (2008) employed the mean of the Tyndall Centre v2.03scenario set within a distributed hydrological model. To addressthese issues, Kingston et al. (2011) used the model of the Mekong

uding the distribution of linear reservoir sub-catchments, interow reservoirs ande used for calibration and validation are also indicated.

of Hdeveloped by Kite (2000, 2001) using the Semi-distributed LandUse-based Runoff Processes (SLURP, v.12.7) model (Kite, 1995).This model was used to assess the impacts of the consistent setof climate change scenarios used throughout the QUEST-GSI pro-ject (Todd et al., 2011). The data employed within the SLURP modelprovided the starting point for the development of the MIKE SHEmodel of the Mekong.

2.2. MIKE SHE model development

MIKE SHE can be described as a deterministic, fully distributedand physically based hydrological modelling system (Graham andButts, 2005). It is a comprehensive model for simulating the majorprocesses of the land phase of the hydrological cycle and has beenused in environments ranging from major international river ba-sins (Andersen et al., 2001; Stisen et al., 2008), through catchmentsof hundreds or thousands of km2 (Feyen et al., 2000; Huang et al.,2010; Singh et al., 2010, 2011), to small (

l of Hconsidered appropriate to maintain the original distribution ofmeteorological inputs. Monthly precipitation totals and mean tem-perature were obtained from the 0.5 resolution University of Del-aware global precipitation dataset (UDel; Legates and Willmott,1990) and the CRU TS 3.0 dataset (Mitchell and Jones, 2005),respectively. Monthly data from the 268 grid cells covering the Me-kong catchment were then stochastically disaggregated to dailyresolution following the procedures of Arnell (2003) as describedby Todd et al. (2011). This required the coefcient of variationfor precipitation and standard deviation for temperature from sta-tion-based data provided by the US National Climate Data Centre(NCDC) global surface summary of the day (GSOD) meteorologicalstations previously employed by Kite (2001). Mean daily precipita-tion and temperature were subsequently evaluated for each sub-catchment used to distribute meteorological inputs.

Precipitation lapse rates for those sub-catchments with a largerange in elevation were modied during model calibration. Simi-larly, a temperature lapse rate was included over the Lancang,the one sub-catchment in which snow cover is a regular feature.This lapse rate was adjusted so that the range of temperaturesand number of months when temperature was below 0 C withinspecic MIKE SHE grid squares approximated those at GSOD mete-orological stations located within these grid squares. Monthly PETwas evaluated for each of the sub-catchments employed to distrib-ute meteorological data using the Linacre method, the same PETscheme employed by Kingston et al. (2011). Monthly totals weredistributed evenly throughout the month, following initial experi-ments that showed model results to be insensitive to this methodcompared to the application of daily estimates of PET.

Although all gridded data were specied using a 1 km 1 kmgrid, the computational grid employed by the model was increasedin size to 10 km 10 km following a series of experimental runswhich showed little change in simulated river discharge for gridsizes between 1 km and 20 km (see Vzquez et al., 2002; Thomp-son, 2012). The larger grid size resulted in logistically appropriatecomputation times for the application of the autocalibration rou-tines discussed below. This balance between computation timeand representation of catchment attributes is common with dis-tributed hydrological models (McMichael et al., 2006). WithinMIKE SHE, all input data were automatically resampled to the lar-ger grid. Hypsometric curves derived from the resampled and ori-ginal topography are very similar, as are the relative importance ofthe different soil and land use categories, suggesting that the largergrid retains a good representation of catchment characteristics.

2.3. Model calibration and validation

In common with Kingston et al. (2011), a baseline period of19611990 was used for calibration and the shorter 19911998period for validation. Whereas SLURP was only calibrated usingthree stations (Chiang Saen, Pakse and Ubon, Fig. 1), data from afurther nine gauging stations were used for MIKE SHE (Fig. 1). Eightof these are on the main Mekong and the other (Yasothon) on atributary, the Chi. Although daily discharge data were availablefor the full calibration period for the majority of these stations,the records for Stung Treng, Kompong Cham and Phnom Penh werelimited to January 1961December 1969, January 1964March1974 and January 1961March 1974, respectively. The dischargedata for Kratie, which were available for the 30 year calibrationperiod, were derived by Kite (2000) using Pakse discharge andmethods developed by the Institute of Hydrology (1988).

Calibration was undertaken using the time constants of the sat-urated zone linear reservoirs, the dead storage proportion for the

J.R. Thompson et al. / Journalower baseow reservoirs, and, in sub-catchments with a wide ele-vation range, precipitation lapse rate. The snowmelt degree-daycoefcient was varied during calibration of the Lancang at ChiangSaen, this being the one sub-catchment with snow cover. Followingcalibration, a review of the simulated snow cover was undertakento conrm that extensive snow was only present during the periodNovemberMarch and that its maximum extent was approxi-mately equivalent to 5% of the total catchment area (e.g. Kiemet al., 2005). Calibration was undertaken for each gauging stationin a downstream sequence beginning at Chiang Saen and progress-ing to Mukdahan. The Chi and Mun sub-catchments (Yasothon andUbon gauging stations) were then calibrated before continuing thecalibration for the gauging stations between Pakse and PhnomPenh. In each case, only those model parameters for sub-catch-ments between the previously calibrated upstream gauging stationand the current station were varied.

Initially, an automatic multiple objective calibration was under-taken based on the shufed complex evolution method (Duanet al., 1992; Madsen, 2000, 2003). Two equally weighted calibra-tion criteria, the absolute value of the average error and the rootmean square error, were employed (Butts et al., 2004), and thesewere aggregated into one measure using a transformation thatcompensates for differences in the magnitudes of the criteria(Madsen, 2003). The autocalibration routine evaluated the two cal-ibration criteria at the model time step (dened as a maximum of48 h). However, as acknowledged by Kingston et al. (2011), there isa disconnect between the daily meteorological input data and dis-charge as a result of generating the daily meteorological data usinga stochastic weather generator.

Whilst model performance following autocalibration was gen-erally good, it was possible to improve it through manual modi-cation of model parameters, with observed and simulateddischarge being aggregated to mean monthly ow (Kingstonet al., 2011). Model performance at each gauging station was as-sessed using the NashSutcliffe coefcient (NSE, Nash and Sutcliffe,1970), the Pearson correlation coefcient (r) and the percentagedeviation in simulated mean ow from the observed mean ow(Dv; Henriksen et al., 2003). The scheme of Henriksen et al.(2008) was used to classify the model performance as indicatedby NSE and Dv.

Following calibration, the implications of using the stochasticweather generator to disaggregate monthly precipitation and tem-perature to a daily time step was investigated using eight differentoutputs from the weather generator for both meteorological in-puts. All possible combinations of these time series were employed(i.e. 64 runs). Calibration statistics were evaluated for each run.Subsequently, the model was run for the shorter 19911998 periodfor validation. Data for two of the gauging stations used for calibra-tion, Kompong Cham and Phnom Penh, were not available for thisperiod, whilst the length of records for the remaining 10 stationsvaried from the complete 8 years to only 3 years.

2.4. Simulation of climate change

The same revised meteorological inputs as used by Kingstonet al. (2011) were employed to simulate potential impacts of, anduncertainty associated with, climate change. Future (monthly res-olution) climate scenarios for temperature and precipitation werederived using the ClimGen pattern-scaling technique (Arnell andOsborn, 2006) for a 30 year period. ClimGen, a spatial scenario gen-erator, employs the assumption that the spatial pattern of climatechange, expressed as change per unit of global mean temperature,is constant for a given GCM. In this way, it is possible for the pat-tern of climate change from a GCM to be scaled up and down inmagnitude. This enables the impacts of specic thresholds of glo-bal climate change to be investigated (Todd et al., 2011). Scenarios

ydrology 486 (2013) 130 5were generated for a prescribed warming of 2 C, the hypothesisedthreshold for dangerous climate change (Todd et al., 2011), for se-ven GCMs: CCCMA CGCM31, CSIRO Mk30, IPSL CM4, MPI ECHAM5,

NCAR CCSM30, UKMO HadGEM1 and UKMO HadCM3. These wereselected from the CMIP-3 database (Meehl et al., 2007) as exemplarGCMs representing different future representations of global cli-mate system features. In addition, the UKMO HadCM3, a widelyemployed GCM previously used for model uncertainty analysis(e.g. Murphy et al., 2004), was selected to derive scenarios for pre-scribed warming of global mean temperature of 1, 2, 3, 4, 5, and6 C.

Monthly scenario Linacre PET was calculated for each GCM gridsquare and, as for baseline meteorological data, distributed on adaily basis evenly throughout each month. Scenario precipitationand temperature for each GCM grid square were downscaled todaily resolution using the stochastic weather generator. Precipita-tion, temperature and PET were averaged for each meteorologicalsub-catchment. Separation of the river discharge climate changesignal into that attributable to modications in precipitation, PETand temperature was investigated by in turn specifying scenariotime series for one of the three meteorological inputs whilstemploying baseline time series for the other two (Kingston et al.,2011; Thompson, 2012). This was undertaken for both sets of cli-mate change scenarios.

2.5. Inter-hydrological model comparison

As discussed above, a less frequently considered source ofuncertainty in climate change hydrological impact assessments ishydrological model uncertainty (Prudhomme and Davies, 2009).

distributed, physically based and semi-distributed, conceptual pro-cess descriptions. The SLURP model of the Mekong developed byKite (2000, 2001) and rened by Kingston et al. (2011) providesthe second hydrological model. SLURP is a semi-distributed, con-ceptual model that uses a combination of sub-basins and land cov-er classications to dene model elements in an approach similarto the hydrological response units used within models such asSWAT (e.g. Arnold et al., 1998). Vertical water balances are subse-quently evaluated for each of these elements, with results beingaggregated at the sub-catchment scale. Of the 12 gauging stationsused in the current study to present MIKE SHE results (Fig. 1), re-sults from SLURP for the same climate change scenarios were avail-able for three stations: Chiang Saen, Pakse and Ubon (althoughresults for the latter station were not presented in the earlierstudy).

The third hydrological model is Mac-PDM.09 (Mac formacro-scale and PDM for probability distributed moisturemodel), an established global hydrological model that simulatesrunoff across the global land surface domain (Arnell, 1999b,2003; Gosling and Arnell, 2011). It calculates the daily water bal-ance in each of the 67,420 0.5 0.5 cells across the global landsurface, treating each cell as an independent catchment. Runoff isgenerated from precipitation falling on the portion of a cell thatis saturated, and by drainage from water stored in the soil. Mac-PDM.09 was not specically calibrated for the Mekong; instead,the model was calibrated at the continental scale by tuning it tohelp dene parameter values. This involved tests of precipitation

dels

eptu

rom

one:

b-and

eed

n tim

6 J.R. Thompson et al. / Journal of Hydrology 486 (2013) 130In this study, an assessment of such uncertainty was undertakenusing results from three alternative hydrological models for thesame two sets of climate change scenarios. Table 1 summarisesthe key attributes of these three hydrological models, includingthe type of model, the meteorological inputs employed and thespatial distribution of meteorological inputs, catchment character-istics and process simulation.

The rst hydrological model is the MIKE SHE model developedin the current study that comprises a combination of spatially

Table 1Summary of key attributes of the MIKE SHE, SLURP and Mac-PDM.09 hydrological mo

Attribute MIKE SHE

Model type Distributed, physically based model, with conclinear reservoir saturated zone component

River routing Kinematic routing (MIKE 11)Time step Variable max. 48 hMeteorological inputsa P, T, PETPET method Linacre PET calculated externally

Snow scheme Degree-dayMeteorological inputs spatial

distributionDistributed according to 13 sub-catchments fSLURP (Fig. 1)

Spatial distribution of catchmentcharacteristics

Topography, land cover and soil: based on1 km 1 km gridded data resampled to a10 km 10 km computational grid

Spatial resolution of processcomputation

All model components except the saturated zdistributed according to 10 km 10 km grid.Saturated zone: distributed according to 17 sucatchments, each comprising three interowtwo baseow reservoirs (Fig. 1)

Calibration parametersb ki, kp, kb, DZfrac, Plapse, Tlapse, DD

a P: precipitation, T: air temperature, PET: potential evapotranspiration, W: wind spux (downward).

b ki: interow time constants for saturated zone interow reservoirs, kp: percolatio

reservoirs, DZfrac: dead storage in the baseow reservoirs, Plapse: precipitation lapse rate,constants and capacities of the fast and slow soil stores, M: Mannings n roughness coecomputation of Linacre PET.datasets and potential evaporation calculations and comparisonswith long-term average runoff and within-year runoff patternsfor a small number of major river basins and for a large numberof small basins (see Arnell, 1999b). The performance of Mac-PDM.09 was recently evaluated by validating simulated runoffagainst observed runoff for 50 catchments across the globe, andthe model was found to perform well (Gosling and Arnell, 2011).Moreover, a recent inter-hydrological model comparison exerciseshowed that Mac-PDM.09 performed as well as other global

of the Mekong.

SLURP Mac-PDM.09

al Semi-distributed vertical water balance model Semi-distributedconceptual waterbalance globalhydrological model

Muskingum NoneDaily DailyP, T P, T, W, SH, LWnet, SWLinacre PET calculated within the model PenmanMonteith

PET calculated withinthe model

Degree-day Degree-dayDistributed according to 13 sub-catchments (Fig. 1) 0.5 0.5 grid

Topography, land cover and soil: based on a1 km 1 km gridded data

Land cover and soil:0.5 0.5 grid

13 sub-catchments divided into elements for waterbalance calculations based on land cover, with 98elements for the Mekong catchment. Results foreach element aggregated based on relative coverwithin each sub-catchment

0.5 0.5 grid

RC, M, FC, U Not calibrated for theMekong (see text)

, SH: specic humidity, LWnet: net longwave radiation ux, SW: shortwave radiation

e constants for saturated zone interow reservoirs, kb: time constants for baseow

Tlapse: temperature lapse rate, DD: snow melt degree-day coefcient, RC: retentionfcient for overland ow, FC: soil eld capacity coefcients, U: wind factor used in

hydrological models in terms of simulating the patterns and mag-nitudes of observed global runoff (Haddeland et al., 2011).

For comparison with results from MIKE SHE and SLURP, simu-lated runoff from Mac-PDM.09 for each climate change scenariowas aggregated at a monthly time step for all the model grid cellswithin the boundaries of the catchments of six gauging stations:Chiang Saen, Vientiane, Nakhon Phanom, Pakse, Phnom Penh andUbon (Fig. 1). These were selected in order to provide a comparisonof results for the three gauging stations for which results are avail-able for all three hydrological models (Chiang Saen, Pakse andUbon). Vientiane and Nakhon Phanom were selected as stationsin the middle of the catchment upstream of the twomajor tributar-ies draining the Korat Plateau (the Chi and Mun), whilst PhnomPenh is the lowest station on the river and upstream of the ecolog-ically and economically important Mekong Delta. Comparisons ofresults for these stations are limited to MIKE SHE and Mac-PDM.09. This analysis extends the preliminary inter-model com-parison undertaken by Gosling et al. (2011a) that was limited toa comparison of SLURP and Mac-PDM.09 for Pakse alone.

3. Results

3.1. Model calibration and validation

Model performance statistics derived for the 19611990 periodfor the 12 gauging stations used in model calibration are providedin Table 2. As indicated above, these are derived from meanmonthly discharges. For Stung Treng, Kompong Cham and Phnom

Penh, statistics are based upon the reduced periods of observeddischarge. In the case of Chiang Saen, Pakse and Ubon, the valuesof Dv and NSE reported by Kingston et al. (2011) from the cali-brated SLURP model of the Mekong are also provided. Accordingto the classication scheme of Henriksen et al. (2008), the perfor-mance of the MIKE SHE model, as indicated by the values of Dvand NSE, can in general be classed as excellent (20 out of the24 model performance statistics). Although an equivalent classi-cation for the correlation coefcient is not employed, the value ofthis statistic is above or very close to 0.95 for 10 of the 12 gaugingstations.

Fig. 2 shows monthly mean observed and simulated dischargefor ve gauging stations along the main Mekong for the calibrationperiod. These stations are selected as representative of results forthe main stem of the Mekong above Kratie, the latter being thelowest gauging station for which ow data (albeit in this case de-rived from another station) are available for the full simulationperiod. Flow duration curves derived from observed and simulatedmean monthly discharges at all gauging stations (not shown) con-rm the good performance of the MIKE SHE model. When com-pared to SLURP, model performance at Chiang Saen is superior,with MIKE SHE Dv and NSE values exceeding those obtained usingthe earlier model (Table 2). Dv for SLURP falls in the very good asopposed to excellent category. Fig. 3, which shows observed andsimulated river regimes for all 12 gauging stations including thosefrom SLURP for the stations where they are available, demonstratesa considerable overestimation by SLURP in recession discharge fol-lowing the annual peak at Chiang Saen. In contrast, discharges at

Table 2MIKE SHE model performance statistics for 12 gauging stations within the Mekong Basin for the calibration period (19611990 unless stated otherwise). Letters after gaugingstation names refer to the labels used in Fig. 1. Corresponding statistics from Kingston et al. (2011) for SLURP are shown in brackets for three stations. Model performanceindicators are taken from Henriksen et al. (2008).

Station Obs. mean (m3 s1) Sim. mean (m3 s1) Dv (%)a NSEb rc

J.R. Thompson et al. / Journal of Hydrology 486 (2013) 130 7Mekong at Chiang Saen (a) 2711.3

Mekong at Luang Prabang (b) 3980.2

Mekong at Vientiane (c) 4521.1

Mekong at Nakhon Phanom (d) 7031.6

Mekong at Mukdahan (e) 7602.4

Mekong at Pakse (f) 9836.8

Mekong at Stung Treng (19611969) (g) 13381.0

Mekong at Kratie (h) 13418.9

Mekong at Kompong Cham (1964March 1974) (i) 13409.5

Mekong at Phnom Penh (1961March 1974) (j) 13022.3

Chi at Yasothon (k) 202.3

Mun at Ubon (l) 636.3

Performance indicator ExcellentIIIII

Dv 0.85a Percentage deviation in simulated mean ow from observed mean ow (Henriksenb NashSutcliffe coefcient (Nash and Sutcliffe, 1970).c Pearson correlation coefcient.2735.3 +0.88 0.888 0.943IIIII IIIII

(2921.1) (+8.20) (0.78)(IIII) (IIII)

4132.4 +3.82 0.892 0.947IIIII IIIII

4740.9 +4.86 0.900 0.951IIIII IIIII

7322.7 +4.14 0.910 0.955IIIII IIIII

7874.4 +3.58 0.907 0.953IIIII IIIII

10144.8 +3.13 0.901 0.951IIIII IIIII

(9876.7) (+0.90) (0.890)(IIIII) (IIIII)

13911.8 +3.97 0.924 0.963IIIII IIIII

13579.8 +1.20 0.901 0.950IIIII IIIII

14237.8 +6.18 0.904 0.954IIII IIIII

14719.4 +13.03 0.866 0.951III IIIII

204.1 +0.88 0.494 0.712IIIII II

638.4 +0.34 0.550 0.750IIIII III

(899.5) (+41.90) (0.44)(I) (II)

Very good Fair Poor Very poorIIII III II I

510% 1020% 2040% >40%0.650.85 0.500.65 0.200.50

of H8 J.R. Thompson et al. / Journalthis time of year simulated by MIKE SHE more closely follow theobserved. Similarly good reproduction of the river regimes at otherstations between Chiang Saen and Pakse by MIKE SHE is also

0

2

4

6

8

10

12

01/6

1

01/6

2

01/6

3

01/6

4

01/6

5

01/6

6

01/6

7

01/6

8

01/6

9

01/7

0

01/7

1

01/7

2

01/7

3

01/7

4

01/7

5Dis

char

ge (1

03m

3 s-1)

Chiang Saen (a)

0

4

8

12

16

20

01/6

1

01/6

2

01/6

3

01/6

4

01/6

5

01/6

6

01/6

7

01/6

8

01/6

9

01/7

0

01/7

1

01/7

2

01/7

3

01/7

4

01/7

5Dis

char

ge (1

03 m

3 s-1)

Vientiane (c)

05

101520253035

01/6

1

01/6

2

01/6

3

01/6

4

01/6

5

01/6

6

01/6

7

01/6

8

01/6

9

01/7

0

01/7

1

01/7

2

01/7

3

01/7

4

01/7

5Dis

char

ge (1

03 m

3 s-1)

Mukdahan (e)

05

1015202530354045

01/6

1

01/6

2

01/6

3

01/6

4

01/6

5

01/6

6

01/6

7

01/6

8

01/6

9

01/7

0

01/7

1

01/7

2

01/7

3

01/7

4

01/7

5Dis

char

ge (1

03 m

3 s-1)

Pakse (f)

0

10

20

30

40

50

60

01/6

1

01/6

2

01/6

3

01/6

4

01/6

5

01/6

6

01/6

7

01/6

8

01/6

9

01/7

0

01/7

1

01/7

2

01/7

3

01/7

4

01/7

5Dis

char

ge (1

03 m

3 s-1)

Kratie (h)

Fig. 2. Monthly mean observed and MIKE SHE simulated discharge for ve gauging statiorefer to the gauging station labels used in Fig. 1).ydrology 486 (2013) 130demonstrated. At the latter station, the performance statistics arecomparable for MIKE SHE and SLURP. A higher NSE value is ob-tained for the former, whilst simulated mean ow is closer to the

01/7

6

01/7

7

01/7

8

01/7

9

01/8

0

01/8

1

01/8

2

01/8

3

01/8

4

01/8

5

01/8

6

01/8

7

01/8

8

01/8

9

01/9

0

Obs Sim

01/7

6

01/7

7

01/7

8

01/7

9

01/8

0

01/8

1

01/8

2

01/8

3

01/8

4

01/8

5

01/8

6

01/8

7

01/8

8

01/8

9

01/9

0

Obs Sim

01/7

6

01/7

7

01/7

8

01/7

9

01/8

0

01/8

1

01/8

2

01/8

3

01/8

4

01/8

5

01/8

6

01/8

7

01/8

8

01/8

9

01/9

0

Obs Sim

01/7

6

01/7

7

01/7

8

01/7

9

01/8

0

01/8

1

01/8

2

01/8

3

01/8

4

01/8

5

01/8

6

01/8

7

01/8

8

01/8

9

01/9

0

Obs Sim

01/7

6

01/7

7

01/7

8

01/7

9

01/8

0

01/8

1

01/8

2

01/8

3

01/8

4

01/8

5

01/8

6

01/8

7

01/8

8

01/8

9

01/9

0

Obs Sim

n along the Mekong River for the calibration period (19611990). (Letters in brackets

ObsSim

)

l of H3

4

5

6

7

harg

e (1

03m

3 s-1

)ObsSimSLURP

Chiang Saen (a)

4

6

8

10

12Luang Prabang (b

harg

e (1

03m

3 s-1

)

J.R. Thompson et al. / Journaobserved for SLURP. However, Fig. 3 demonstrates that the sea-sonal peak discharge (August) is underestimated by SLURP by onaverage nearly 2640 m3 s1 (10%) compared to under 300 m3 s1

(1.1%) for MIKE SHE (although in September underestimation byMIKE SHE increases to 4.6%).

Further downstream on the main Mekong, the values of Dv forKompong Cham and Phnom Penh, for which discharge data is lim-ited to the rst half of the calibration period, fall below the excel-lent category and are instead classied as very good and fair,respectively (Table 2). The NSE values for these two stations arestill, however, classied as excellent. Reduction in model

0

1

2

J F M A M J J A S O N D

Dis

c

0

2

J F M A M J

02468

101214161820

J F M A M J J A S O N D

ObsSim

Nakhon Phanom (d)

0

5

10

15

20

25

J F M A M J

ObsSim

Mukdahan (e)

0

5

10

15

20

25

30

35

40

J F M A M J J A S O N D

ObsSim

Stung Treng (1961-1969) (g)

0

5

10

15

20

25

30

35

40

J F M A M J

ObsSim

Kratie (h)

05

1015202530354045

J F M A M J J A S O N D

ObsSim

Phnom Penh (1961-March 1974) (j)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

J F M A M J

ObsSim

Chi at Yasothon (k)

Dis

char

ge (1

03m

3 s-1

)D

isch

arge

(103

m3 s

-1)

Dis

char

ge (1

03m

3 s-1

)

Dis

cD

isch

arge

(103

m3 s

-1)

Dis

char

ge (1

03m

3 s-1

)D

isch

arge

(103

m3 s

-1)

Fig. 3. Observed and MIKE SHE simulated river regimes for all 12 gauging stations witotherwise). Regimes simulated by SLURP for three gauging stations are also shown. (Let4

6

8

10

12 ObsSim

Vientiane (c)

harg

e (1

03m

3 s-1

)

ydrology 486 (2013) 130 9performance at these two stations is evident in Fig. 3, as simulateddischarge leads observed more than at the other stations on themain Mekong for which data are available for the full calibrationperiod. This feature is also evident (although to a smaller extent)at Stung Treng, where observed data are also limited in duration,suggesting that model performance may be inuenced by thisreduction in the length of observed discharge.

Model performance is relatively weak for the Chi and Mun sub-catchments (Table 2), with peak seasonal discharge being underes-timated (Fig. 3). In comparison to other gauging stations, simulatedows deviate further from the observed. During calibration it was

J A S O N D0

2

J F M A M J J A S O N D

J A S O N D0

5

10

15

20

25

30

J F M A M J J A S O N D

ObsSimSLURP

Pakse (f)

J A S O N D05

1015202530354045

J F M A M J J A S O N D

ObsSim

Kompong Cham (1964-March 1974) (i)

J A S O N D0.0

0.5

1.0

1.5

2.0

2.5

J F M A M J J A S O N D

ObsSimSLURP

Mun at Ubon (l)

Dis

char

ge (1

03m

3 s-1

)D

isch

arge

(103

m3 s

-1)

Dis

char

ge (1

03m

3 s-1

)D

isc

hin the Mekong catchment for the calibration period (19611990 unless indicatedters in brackets refer to the gauging station labels used in Fig. 1).

hmeton

1)

sen

of HTable 3MIKE SHE model performance statistics for 10 gauging stations within the Mekong catcstation names refer to the labels used in Fig. 1. Corresponding statistics from Kingsindicators are taken from Henriksen et al. (2008).

Station Obs. mean (m 3s

Mekong at Chiang Saen (1991June 1997) (a) 2490.3

Mekong at Luang Prabang (19911997) (b) 3749.7

Mekong at Vientiane (19911996) (c) 4241.7

Mekong at Nakhon Phanom (1991November 1995) (d) 7063.2

Mekong at Mukdahan (19911995) (e) 7434.4

Mekong at Pakse (19911998) (f) 9168.4

Mekong at Stung Treng (19911993) (g) 12569.5

Mekong at Kratie (19911998) (h) 12505.7

Chi at Yasothon (19911995) (k) 200.4

Mun at Ubon (19911993) (l) 486.8

Performance indicator ExcellentIIIII

Dv 0.85

a Percentage deviation in simulated mean ow from observed mean ow (Henrik

10 J.R. Thompson et al. / Journalnot possible to raise peak discharges without increasing dischargeduring the annual rise and recession, which are reasonably wellreproduced. These changes would result in overestimation of meandischarge causing Dv, currently classied as excellent, to in-crease. This trade-off is evident in SLURP results for Ubon. Whilstthe magnitude of the simulated peak corresponds well with the ob-served (Fig. 3), overestimation of discharge through most of therest of the year causes Dv to be classied as very poor (Table 2).The NSE value for this gauging station for MIKE SHE is a class abovethat of SLURP (fair compared to poor). At Yasothon, the NSE va-lue just falls short of the fair category. Therefore, although MIKESHE does not perform as well for these two sub-catchments as itdoes for gauging stations on the main Mekong, results are consid-ered to be superior to those of SLURP.

Results from the 64 runs over the same 30 year calibration per-iod using the calibrated model and eight alternative precipitationand temperature time series from the stochastic weather generatorshowed that the model was not sensitive to the disaggregationprocedure. NSE and r for Chiang Saen varied by only 0.03 and0.02, respectively, whilst at Pakse the corresponding ranges were0.02 and 0.01. Further downstream at Kratie, the ranges of bothstatistics were only 0.01. Very similar results were obtained forthe other gauging stations. These ndings concur with the resultsof Kingston et al. (2011), who employed a smaller set of 10 alterna-tive model runs to assess the sensitivity of the SLURP model toweather generator inputs.

Table 3 shows performance statistics for the shorter validationperiod (19911998). As indicated, the length of observed dischargerecords varies between gauging stations. The corresponding valuesfor Chiang Saen and Pakse reported by Kingston et al. (2011) arealso provided (values for Ubon were not reported in the earlierstudy). Good performance of the MIKE SHE model is indicated,

b NashSutcliffe coefcient (Nash and Sutcliffe, 1970).c Pearson correlation coefcient.nt for the validation period (19911998 unless stated otherwise). Letters after gauginget al. (2011) for SLURP are shown in brackets for two stations. Model performance

Sim. mean (m3 s1) Dv (%)a NSEb rc

2258.7 9.30 0.813 0.850IIII IIII

(2550.0) (+2.40) (0.810)(IIIII) (IIII)

3448.9 8.02 0.887 0.904IIII IIIII

3969.1 6.43 0.901 0.922IIII IIIII

6404.3 9.33 0.791 0.853IIII IIII

6866.9 7.63 0.812 0.853IIII IIII

8400.5 8.38 0.858 0.877IIII IIIII

(8783.4) (4.20) (0.770)(IIIII) (IIII)

11139.8 11.37 0.508 0.536III III

11700.8 6.44 0.734 0.759IIII IIII

166.8 16.73 0.581 0.650III III

440.0 9.60 0.820 0.847IIII IIII

Very good Fair Poor Very poorIIII III II I

510% 1020% 2040% >40%0.650.85 0.500.65 0.200.50

m)sub

l of HTable 4Mean annual precipitation and potential evapotranspiration (PET) for the baseline (mwithin the Mekong catchment. (Numbers in brackets refer to the meteorological inputsthe baseline).

Parameter Scenario Lancang (1) Mek. 1 (4) Chi (5)

Precipitation Baseline 1052.8 1855.8 1272.3CCCMA 10.1 10.2 12.3CSIRO 4.6 4.6 3.3HadCM3 10.1 1.0 0.1HadGEM1 5.9 3.7 6.1IPSL 5.2 1.1 0.1MPI 3.6 7.0 10.2NCAR 8.5 9.1 5.0

PET Baseline 1765.6 1923.0 2363.6CCCMA 11.7 12.3 13.1CSIRO 14.6 15.7 15.9HadCM3 12.9 13.9 13.3HadGEM1 12.4 12.1 10.3IPSL 15.9 15.7 15.3MPI 13.6 13.6 13.3NCAR 11.3 10.9 11.1

Precipitation Baseline 1052.8 1855.8 1272.31 C 5.0 0.4 0.12 C 10.1 1.0 0.13 C 15.2 1.6 0.14 C 20.2 2.4 0.45 C 25.3 3.3 0.96 C 30.2 4.3 1.5

PET Baseline 1765.6 1923.0 2363.61 C 6.3 6.8 6.42 C 12.9 13.9 13.33 C 19.8 21.3 20.64 C 26.9 29.1 28.25 C 34.4 37.3 36.36 C 42.2 45.9 44.8

J.R. Thompson et al. / Journato 0.90 for the calibration period (validation: 0.73) in the currentstudy.

3.2. Climate change scenarios: 2 C increase using seven GCMs

3.2.1. Changes in climateBaseline annual precipitation and PET, and percentage changes

from these totals, for each of the 2 C, seven GCM scenarios areshown for eight sub-catchments in Table 4 (top half). Results forthe other sub-catchments are not shown as they are relativelysmall and changes are represented by one or more of those in Ta-ble 4. Mean monthly precipitation and PET for the baseline andeach scenario are shown for four sub-catchments in Fig. 4. Again,these are representative of changes over the remaining sub-catch-ments. Changes in temperature are not presented here since, inmost sub-catchments, changes in temperature are reected inmodied PET. The exception is the Lancang, the one sub-catchmentin which snow and its seasonal melt inuences runoff. Changes inmean annual temperature over this sub-catchment range from+2.3 C (CCCMA) to +2.9 C (IPSL). Temperatures increase through-out the year in all scenarios with the largest changes (up to +3 C)between October and March. Summer temperature increases areon average lower (+2.02.3 C).

Wide variation in the magnitude and direction of annualchanges in precipitation occurs between GCMs. These were de-scribed by Kingston et al. (2011) and are summarised here and inTable 4. CCCMA, MPI and NCAR exhibit increasing annual precipi-tation for all sub-catchments. Increases are greater in upstreamsub-catchments for CCCMA and NCAR, and in downstream sub-catchments for MPI. CSIRO simulates reduced annual precipitationacross all sub-catchments, whilst reductions occur across all butthree south-central sub-catchments (Chi-Mun, Mekong 2 and SrePok) for IPSL. HadCM3 shows increased precipitation over the fourand changes (%) for the climate change scenarios for representative sub-catchments-catchments identied in Fig. 1. Italicised cells indicate negative changes compared to

Mun (6) Mek. 2 (8) Se Kong (9) Sre Pok (10) Mek. 3 (11)

1313.6 2213.2 2432.5 2055.3 1870.310.2 8.4 5.2 1.9 5.32.9 2.8 2.8 2.9 1.30.4 1.1 2.1 4.5 3.04.8 1.2 2.9 3.9 1.00.1 0.6 0.4 1.3 0.410.3 8.8 6.6 7.6 12.23.5 1.9 3.5 3.7 5.3

2336.5 1813.0 1728.5 1695.9 1770.312.7 12.5 12.7 12.3 12.515.2 15.2 14.9 14.2 14.313.2 14.7 14.8 14.8 15.110.3 12.4 13.0 12.7 12.514.2 14.3 13.9 12.8 13.212.9 13.4 13.5 13.1 13.210.6 11.4 11.1 10.7 10.3

1313.6 2213.2 2432.5 2055.3 1870.30.3 0.6 1.3 2.7 1.70.4 1.1 2.1 4.5 3.00.4 1.3 2.6 5.6 4.20.2 1.3 3.0 6.3 5.10.1 1.3 3.1 6.7 6.00.5 1.1 3.2 6.9 6.7

2336.5 1813.0 1728.5 1695.9 1770.36.2 7.3 7.4 7.3 7.2

13.2 14.7 14.8 14.8 15.120.6 22.4 22.5 22.6 23.528.3 30.5 30.5 30.8 32.236.5 39.0 39.0 39.3 41.445.1 47.9 47.8 48.3 51.1

ydrology 486 (2013) 130 11most northern sub-catchments (Lancang to Mekong 1). Increasesin this part of the catchment are restricted to the top two sub-catchments for HadGEM1, whilst precipitation also increases overthe lower three sub-catchments.

Seasonal changes in precipitation also vary widely betweenGCMs (Kingston et al., 2011; Fig. 4). CCCMA and NCAR experiencepeak increases in April and September, with decreases concen-trated in NovemberFebruary/March. A notable exception is Lanc-ang for the CCCMA GCM, which experiences increasingprecipitation in every month. Inter-seasonal changes in precipita-tion for the MPI GCM are, in contrast, unimodal, with increasesbeing concentrated between May and November. For CSIRO, in-creases in precipitation are limited to September over upstreamsub-catchments (e.g. Lancang and Mekong 1, Fig. 4). This periodexpands to JuneSeptember in the far south (e.g. Mekong 3). Thedistribution of change through the year for IPSL is also unimodal,with increases being limited to AugustSeptember and, in the farsouth, SeptemberOctober (Fig. 4). HadCM3 and HadGEM1 exhibita bimodal distribution in the seasonal pattern of changes in precip-itation, with the largest increases occurring in May and September(HadCM3)/OctoberNovember (HadGEM1).

Differences in the PET climate change signals between the sevenGCMs are smaller than those for precipitation (Table 4, Fig. 4). An-nual PET increases across the Mekong for all GCMs. With theexception of three sub-catchments (Chi, Mun and Chi-Mun), thesmallest increases in annual PET are associated with NCAR (onaverage +10.9%). In the three sub-catchments where this is notthe case, the lowest PET increases result from HadGEM1, althoughthese are only slightly lower than those of NCAR. There is a system-atic geographical pattern for the GCMs that produce the largest in-creases in annual PET. In the four most northerly sub-catchments(14, Fig. 1), IPSL produces the largest changes (mean: +16.0%).In the middle Mekong (sub-catchments 59), the largest changes

of H12 J.R. Thompson et al. / Journalare associated with CSIRO (mean: +15.7%), followed by IPSL orHadCM3, whilst in the lower part of the catchment (sub-catch-ments 1013), HadCM3 (CSIRO for sub-catchment 9) produces

0

50

100

150

200

250

J F M A M J J A S O N D

Prec

ipita

tion

(mm

)

Lancang (1)

0

50

100

150

200

250

300

350

400

450

500

J F M A M J J A S O N D

Prec

ipita

tion

(mm

)

Mekong 1 (4)

0

50

100

150

200

250

300

350

400

J F M A M J J A S O N D

Prec

ipita

tion

(mm

)

Mun (6)

0

50

100

150

200

250

300

350

400

450

J F M A M J J A S O N D

Prec

ipita

tion

(mm

)

Mekong 3 (11)

Fig. 4. Mean monthly precipitation and PET for the baseline and the 2 C, seven GCM climscales. Numbers in brackets refer to the meteorological inputs sub-catchments identieydrology 486 (2013) 130the largest increase in PET (mean:+15.5%), followed by CSIRO.Many of the GCMs show a relatively constant climate change signalthroughout the year (Fig. 4). Notable exceptions are the peaks for

80

100

120

140

160

180

200

220

240

J F M A M J J A S O N D

PET

(mm

)

100

120

140

160

180

200

220

240

J F M A M J J A S O N D

PET

(mm

)

150

170

190

210

230

250

270

290

J F M A M J J A S O N D

PET

(mm

)

120

130

140

150

160

170

180

190

200

210

J F M A M J J A S O N D

PET

(mm

)

ate change scenarios for four representative sub-catchments. (Note different y-axisd in Fig. 1).

CSIRO in April and May and for MPI in April, which are most clearlyevident in central and lower sub-catchments (e.g. Mekong 1, Munand Mekong 3; Fig. 4).

3.2.2. Changes in river owTable 5 presents the values of the mean, Q5 and Q95 (discharges

exceeded 5% and 95% of the time, respectively) discharges for thebaseline and the percentage changes in these discharges for eachof the 2 C, seven GCM scenarios. These are provided for eightgauging stations which are representative of changes at the otherfour stations used in model calibration/validation. Mukdahan rep-resents the changes at Nakhon Phanom approximately 100 km up-stream. Similarly, discharge at Stung Treng, which is not shown,responds in the same way to Kratie (c. 150 km further down-stream), whilst Phnom Penh is representative of the changes insimulated discharge at Kompong Cham (c. 90 km upstream). Re-sults for the Mun at Ubon represent those in the smaller catchmentof the Chi at Yasothon. The simulated baseline and scenario riverregimes for the same eight gauging stations are shown in Fig. 5.

Of the three GCMs for which precipitation increases in all sub-catchments, two (CCCMA and NCAR) result in increases in meandischarge for all gauging stations (Table 5). The magnitude of thechanges is relatively uniform along the main stem of the Mekong

the opposite for baseline river regimes. This is due to the large Sep-tember increases in precipitation.

A similar pattern of change occurs for NCAR, although in mostcases changes are larger than those associated with CCCMA (partic-ularly in the middle section of the catchment between Luang Pra-bang and Pakse; Table 5). Table 4 shows that whilst increases inprecipitation over this part of the catchment are smaller for NCARcompared to CCCMA, PET rises by smaller amounts for NCAR,accounting for the enhanced river ow. A relatively consistentchange in the river regime at all stations is shown in Fig. 5. Base-ows are higher, especially in the upstream part of the catchment,and the seasonal rise begins slightly earlier, although for themajority of this period, deviations from the baseline are small.The month of highest mean discharge shifts from August to Sep-tember, although (with the exception of Chiang Saen) mean dis-charge in August also increases. Following the peak, dischargesduring the recession are higher than the baseline, especially inupper and middle parts of the catchment.

Precipitation increases in all sub-catchments for MPI. However,the magnitude of these changes is relatively small in upstreamparts of the catchment compared to CCCMA and NCAR, whilst in-creases in PET are larger than for these two GCMs (Table 4). Meanannual ows at Chiang Saen and Luang Prabang therefore decline(by 2.3% and 1%, respectively, Table 5). Below these stations, gains

angised

e (c)

Q95 Baseline 767.9 1092.6 1234.0

J.R. Thompson et al. / Journal of Hydrology 486 (2013) 130 13CCCMA 7.5 13.0 13.4CSIRO 23.2 23.5 21.4HadCM3 12.9 11.0 9.0HadGEM1 1.6 2.6 2.9IPSL 21.9 18.3 16.9for CCCMA, albeit with a small increase at stations in middlereaches (Mukdahan and Pakse). The much larger increase in meandischarge of the Mun at Ubon is indicative of other sub-catchmentsin this part of the catchment, which results in enhanced ows fromthese tributaries to the Mekong, although absolute magnitudes arerelatively small. In the upper catchment (Chiang Saen), increasingannual discharge is a result of higher ows during the initial rise indischarge, which could be partially attributable to enhanced snow-melt, and then higher discharge during the recession and low owperiod (Fig. 5). Peak annual discharge is, however, slightly (

of H14 J.R. Thompson et al. / Journalin Q5 for ISPL, with relatively small reductions at Pakse due to in-creased wet season and mean annual precipitation in this part ofthe catchment. The river regimes from both GCMs display a

0

1

2

3

4

5

6

7

J F M A M J J A S O N D

Dis

char

ge (1

06m

3 s-1

)

Chiang Saen (a)

0

2

4

6

8

10

12

14

J F M A M J J A S O N D

Vientiane (c)

0

5

10

15

20

25

30

35

J F M A M J J A S O N D

Pakse (f)

0

5

10

15

20

25

30

35

40

45

J F M A M J J A S O N D

Phnom Penh (j)

Dis

char

ge (1

06m

3 s-1

)D

isch

arge

(106

m3 s

-1)

Dis

char

ge (1

06m

3 s-1

)

Fig. 5. River regimes simulated by MIKE SHE for the baseline and 2 C, seven GCM climatin brackets refer to the gauging station labels used in Fig. 1).ydrology 486 (2013) 130delayed response in the annual rise in river discharge, whilst dis-charge during the post-peak recession and the dry season are rel-atively unaffected (Fig. 5). On average, mean discharge for each

0

2

4

6

8

10

12

J F M A M J J A S O N D

Luang Prabang (b)

0

5

10

15

20

25

J F M A M J J A S O N D

Mukdahan (e)

0

5

10

15

20

25

30

35

40

45

J F M A M J J A S O N D

Kratie (h)

0.0

0.5

1.0

1.5

2.0

2.5

J F M A M J J A S O N D

Mun at Ubon (l)

Dis

char

ge (1

06m

3 s-1

)D

isch

arge

(106

m3 s

-1)

Dis

char

ge (1

06m

3 s-1

)D

isch

arge

(106

m3 s

-1)

e change scenarios for eight gauging stations within the Mekong catchment. (Letters

month declines at stations on the Mekong for CSIRO and there is ashift from August to September for peak ows. On the Mun at Ubonpeak discharge, which occurs in October, exceeds the baseline. Re-sults for ISPL show a gradual downstream shift from peak owsoccurring in August (Chiang Saen and Luang Prabang), throughapproximately equal mean discharge in August and September(Vientiane), to September discharge exceeding ows in August(Nakhom Phanom and downstream). On the Mun and Chi tributar-ies, mean discharges in September and October exceed those of thebaseline.

Increases in annual precipitation for the Lancang and Nam Oufor HadGEM1 result in a small (

inter-GCM differences in PET. In contrast, the much larger differ-ences in precipitation between the GCMs ensure that the specica-tion of scenario precipitation with baseline PET and temperatureenhances the inter-GCM differences in discharge. The smallestrange of change in mean discharge for gauging stations on themain Mekong (21.1%, between3.2% and 17.9%) is for Phnom Penhcompared to the largest (32.5%, between 6.8% and 25.7%) forChiang Saen. Inter-GCM differences are larger on the two tributar-ies (e.g. 42.1%, between 10.0% and 32.1% for the Mun at Ubon).

Fig. 6 demonstrates very small changes from the baseline whenscenario temperature is employed with baseline precipitation andPET. Within the MIKE SHE model, changes in temperature are al-ready incorporated within the alternative scenario PET, so thattemperature inuences snowmelt alone. As a result, the largest(but still small) changes in mean ow occur at Chiang Saen (reduc-tions in mean discharge of between 0.04% and 1.2%), closest toparts of the catchment that experience snow cover. The magnitudeof these changes declines downstream, and for tributaries in whichsnow is not a feature, changes in temperature alone have no im-pact on discharge (e.g. the Mun, Fig. 6). Small changes in the riverregime at Chiang Saen are associated with a slightly earlierincrease in discharge due to earlier snowmelt and lower peak dis-charges, but variability between the different GCMs is small. Forexample, the increase in May discharge ranges from 19.9% to25.0% (although absolute discharge is small), whilst August dis-charge declines by between 4.4% and 5.4%.

3.2.3. Comparison of MIKE SHE results with SLURP and Mac-PDM.09Fig. 7 shows percentage changes in mean discharge (runoff for

Mac-PDM.09) at six gauging stations for each of the 2 C, sevenGCM scenarios, as simulated by the three hydrological models.As described above, results are only available for three stationsfor SLURP. Direction of change in mean discharge (runoff) for a gi-ven GCM is predominantly the same for all the hydrological mod-els. Of the 42 gauging station/GCM combinations (six stations/seven GCMs), only three exhibit changes in mean discharge (run-off) which differ in sign between hydrological models. These arefor stations towards the southern part of the catchment. Resultsfor HadCM3 for Pakse show increases in discharge (runoff) forMIKE SHE and Mac-PDM.09. For MIKE SHE, the climate change sig-nal (mean discharge: +1.4%) is the smallest of all the GCMs at thisstation, and SLURP shows a reduction in mean discharge of a com-parable magnitude (1.6%). In contrast, Mac-PDM.09 runoff in-creases by 8.4%. At Phnom Penh, this same GCM is associatedwith a reduction in MIKE SHE mean discharge of 2.1%, whilstMac-PDM.09 runoff increases by 2.5% (SLURP results are not avail-able). Finally, at Ubon, both MIKE SHE and SLURP simulate reduc-tions in mean discharge (of 3.0% and 5.1%, respectively) for the IPSLGCM, whilst mean runoff from Mac-PDM.09 increases by a verysmall amount (0.5%). Beyond these differences there is generalagreement in the order of magnitude of changes for the sevenGCMs for the different hydrological models. At most stations, whenlisted in order of increasing change in mean discharge, the GCMs

10

20

30

40

50

n di

scha

rge

(%)

Chiang Saen (a)

0

10

20

30

3

n di

scha

rge

(%)

Vientiane (c)

0

5

10

15

20

25n

disc

harg

e (%

)Nakhon Phanom (d)

3

SL

16 J.R. Thompson et al. / Journal of Hydrology 486 (2013) 130-30

-20

-10

0

CC CS H3 H1 I M N

Cha

nge

in m

ea

-30

-20

-10

CC CS H

Cha

nge

in m

ea

-20

-15

-10

-5

0

5

10

15

20

CC CS H3 H1 I M N

Cha

nge

in m

ean

disc

harg

e (%

)

Pakse (f)

-15

-10

-5

0

5

10

15

CC CS H

Cha

nge

in m

ean

disc

harg

e (%

)

Phnom Penh (j)

MIKE SHEFig. 7. Change from baseline mean annual discharge (runoff for Mac-PDM.09) for the 2 catchment, as simulated by the three hydrological models. (CC: CCCMA; CS: CSIRO; H3:gauging station labels used in Fig. 1).H1 I M N-25

-20

-15

-10

-5

CC CS H3 H1 I M N

Cha

nge

in m

ea

H1 I M N-25

-20

-15

-10

-5

0

5

10

15

20

CC CS H3 H1 I M N

Cha

nge

in m

ean

disc

harg

e (%

)

Ubon (l)

URP Mac-PDM.09C, seven GCM climate change scenarios for six gauging stations within the MekongHadCM3; H1: HadGEM1; I: IPSL; M: MPI; N: NCAR. Letters in brackets refer to the

l of Happear in the same order for both MIKE SHE and Mac-PDM.09, withthe exception, in most cases, of a single pair of GCMs.

Where mean discharge (runoff) at a gauging station increasesfor an individual GCM for all three hydrological models (MIKESHE and Mac-PDM.09 where SLURP results are not available), thesmallest changes are for SLURP, followed by MIKE SHE and thenMac-PDM.09 (Fig. 7). The greater increases for Mac-PDM.09 areparticularly apparent at upstream stations. At Chiang Saen, thepercentage increase in mean runoff for Mac-PDM.09 for the fourGCMs associated with increased mean river ow is on average5.7 and 3.8 times as large as those of SLURP and MIKE SHE, respec-tively. These values are skewed by large changes for HadCM3 andlarge inter-hydrological model differences (but small absolutechanges) for HadGEM1. HadCM3 stands out as a GCM for whichdifferences between the catchment and global hydrological modelsare particularly large, especially in upstream parts of the Mekong.Percentage increases in mean runoff at Chiang Saen for Mac-PDM.09 for HadCM3 are 9.4 and 4.1 times as great as those forSLURP and MIKE SHE, respectively. Further downstream, inter-hydrological model differences in the magnitude of increases indischarge/runoff (when they occur for all the models) are smaller.This is exemplied in results for Ubon, where the increases in run-off for Mac-PDM.09 for the three GCMs with higher river ow forall three hydrological models are, on average, less than 1.4 timesas large as the increases in MIKE SHE discharge. Mac-PDM.09 in-creases are still 4.6 times as large as those simulated by SLURP.

The dominant trend for GCMs associated with declines in an-nual ow at gauging stations on the main Mekong for all (both)hydrological models is for the largest changes to result for Mac-PDM.09, followed by MIKE SHE and SLURP (Fig. 7). Exceptionsare MPI at Chiang Saen and CSIRO at Pakse, where the reverse orderof magnitude of changes occurs whilst, as noted above, at PhnomPenh MIKE SHE simulates a decline in mean discharge for HadCM3whilst mean runoff increases for Mac-PDM.09. Inter-model differ-ences (in particular between MIKE SHE and Mac-PDM.09 and espe-cially downstream of Chiang Saen) in the magnitude of the declinesare smaller than for those GCMs where river ow increases. Forthree GCMs, reductions in the mean discharge at Ubon are largerfor MIKE SHE than for Mac-PDM.09 (and, as discussed above, forIPSL MIKE SHE mean discharge declines whilst mean runoff forMac-PDM.09 increases slightly). Results for SLURP for Ubon showthat, in most cases, reductions in mean discharge are larger (whenthey occur) than for the other two models. Differences betweenMIKE SHE and SLURP are relatively small. The exception is Had-GEM1, where the decline in SLURP discharge is of a similar magni-tude to that of Mac-PDM.09.

Following the approach of Gosling et al. (2011a), Fig. 8 showsmean monthly total discharge (runoff for Mac-PDM.09) expressedas a percentage of the annual total for ve (three in the case ofSLURP) gauging stations simulated by the three hydrological mod-els for the baseline and each of the 2 C, seven GCM scenarios. Re-sults for Vientiane, which are not shown in the interests of clarity,are similar to those for Nakhon Phanom. Given the dominant sea-sonal precipitation signal of the Asian monsoon, it is unsurprisingthat all three hydrological models simulate large seasonality in riv-er ow. The amplitude of this cycle is, however, greater for Mac-PDM.09. Beyond this, the most obvious differences in the seasonalcycle simulated by the three hydrological models are associatedwith the scenario results for Chiang Saen. SLURP simulates a con-sistent earlier rise in the annual hydrograph, which Kingstonet al. (2011) attributed to earlier snowmelt and a subsequent smal-ler proportion of the annual total discharge occurring in peakmonths for all the scenarios. This is not the case for MIKE SHE or

J.R. Thompson et al. / JournaMac-PDM.09. MIKE SHE simulates some modest increases in earlyseason (May) discharge for some scenarios, but both this modeland Mac-PDM.09 simulate a reduced signicance of ows duringsubsequent months (June and July) for the same scenarios. Thegreater concentration of annual ows in August and Septemberfor some scenarios is evident for MIKE SHE and Mac-PDM.09. Theincreased signicance of ows in the rst of these months for IPSLis clearly evident, although the signicance of September owsalso increases for CSIRO and NCAR. This consistency between MIKESHE and Mac-PDM.09 is repeated for Nakon Phanom.

The broad agreement between MIKE SHE and Mac-PDM.09 isrepeated for the other two gauging stations further downstreamon the main Mekong. Fig. 8 shows that for Pakse the inuence ofthe earlier rise in discharge simulated by SLURP has diminished,although scenario results still suggest greater signicance of dis-charge at this time of year for many for the GCMs, which is con-trary to MIKE SHE and Mac-PDM.09 (with the exception of MPI).MIKE SHE does show some consistency with the results of theother two models, such as the shift of peak ows from August toSeptember for GCMs including CSIRO, IPSL and NCAR. Results forUbon show a general agreement between the three hydrologicalmodels. Both MIKE SHE and SLURP simulate a greater concentra-tion of the annual total discharge in October for all the GCMs withthe exception of HadGEM1 for SLURP (an almost negligible de-cline). Mac-PDM.09 simulates similar increases for ve GCMsalthough for CCCMA and MPI runoff in this month decreasesslightly in signicance. Other inter-model similarities include theincrease in the signicance of discharge during the latter part ofthe annual rise (August) for the MPI GCM.

3.3. Climate change scenarios: 16 C increase using HadCM3

3.3.1. Changes in climateTable 4 (lower half) presents mean annual precipitation and PET

for eight representative meteorological input sub-catchments foreach of the 16 C, HadCM3 scenarios. Mean monthly precipitationand PET for the baseline and each scenario are shown for four sub-catchments in Fig. 9. Temperature changes are not presented andthe patterns of these changes are reected in the modicationsto PET. Over the Lancang (where snow is a feature) changes inmean annual temperatures are above the prescribed increase inglobal mean temperature (for example 1.3 C, 5.1 C and 7.7 Cfor 1 C, 4 C and 6 C, respectively). In common with the 2 C, se-ven GCM scenarios, largest temperature increases occur betweenOctober and March.

Changes in annual precipitation exhibit a distinct geographicalpattern. Over the four sub-catchments that are furthest upstream(sub-catchments 14, Fig. 1), annual precipitation increases in aconsistent linear pattern with increasing temperature (Table 4).The magnitude of these increases declines in a downstream direc-tion (cf. Lancang versus Mekong 1; Table 4). In the downstreampart of the Mekong (sub-catchments 913), annual precipitationdeclines for all scenarios. The magnitude of these reductions in-creases with prescribed warming. There is also a downstreamtrend, so that whilst the maximum change in annual precipitationover the Se Kong (sub-catchment 9) is 3.2%, for Mekong 3 it is6.7%. In the central Mekong (sub-catchments 58), annual pre-cipitation responds in a non-linear way to increased prescribedwarming (as a result of differing linear seasonal trends see fol-lowing paragraph and Fig. 9). As Table 4 illustrates for the Mun(sub-catchment 6), annual precipitation initially declines with ris-ing temperature, but later increases, such that for the 5 C and 6 Cscenarios annual precipitation is above the baseline. However, inall the scenarios, the magnitude of the changes is small (

n (a

of H25

al) Chiang Saen (a):

MIKE SHE25

al) Chiang Sae

SLURP

18 J.R. Thompson et al. / Journalincreases declines from north to south. For example, over the Lanc-ang, precipitation increases in every month except April (Fig. 9).The largest percentage changes occur in MayJune and Septem-berOctober, either side of the wettest months. By Mekong 1, in-creases in precipitation are limited to 5 months (in particular

0

5

10

15

20D

isch

arge

(% o

f ann

u

0

5

10

15

20

Dis

char

ge (%

of a

nnu

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Nakhon Phanom (d):MIKE SHE

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Pakse (f):MIKE SHE

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Pakse (f):SLURP

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Phnom Penh (j):MIKE SHE

0

5

10

15

20

25

30

35

Dis

char

ge (%

of a

nnua

l) Ubon (l):MIKE SHE

0

5

10

15

20

25

30

35

Dis

char

ge (%

of a

nnua

l) Ubon (l):SLURP

J F M A M J J A S O N D J F M A M J

J F M A M J J A S O N D

J F M A M J J A S O N D J F M A M J

J F M A M J J A S O N D

J F M A M J J A S O N D J F M A M J

Fig. 8. Mean monthly discharge (runoff for Mac-PDM.09) as a percentage of the mean anas simulated by the three hydrological models. (Letters in brackets refer to the gauging):25

ual) Chiang Saen (a):Mac-PDM.09

ydrology 486 (2013) 130MayJune but also SeptemberOctober and December). Precipita-tion in the wettest baseline months (JuneJuly) is reduced,although the extension of the wet season on either side of these2 months results in the overall increase in annual precipitation, al-beit of a smaller magnitude to the Lancang. For the four central

0

5

10

15

20

Dis

char

ge (%

of a

nn

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Nakhon Phanom (d):Mac-PDM.09

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Pakse (f):Mac-PDM.09

0

5

10

15

20

25

30

Dis

char

ge (%

of a

nnua

l) Phnom Penh (j):Mac-PDM.09

0

5

10

15

20

25

30

35

J A S O N D J F M A M J J A S O N D

J F M A M J J A S O N D

J A S O N D J F M A M J J A S O N D

J F M A M J J A S O N D

J A S O N D J F M A M J J A S O N D

Dis

char

ge (%

of a

nnua

l) Ubon (l):Mac-PDM.09

nual total for the 2 C, seven GCM climate change scenarios for ve gauging stations,station labels used in Fig. 1).

l of HJ.R. Thompson et al. / Journasub-catchments, increasing precipitation is restricted to 4 months.Relatively large gains in precipitation are concentrated in the earlypart of the monsoon season. In the southernmost sub-catchments

0

50

100

150

200

250

300

Prec

ipita

tion

(mm

)

Lancang (1)

0

50

100

150

200

250

300

350

400

Prec

ipita

tion

(mm

)

Mekong 1 (4)

0

50

100

150

200

250

300

Prec

ipita

tion

(mm

)

Mun (6)

0

50

100

150

200

250

300

350

Prec

ipita

tion

(mm

)

Mekong 3 (11)

J F M A M J J A S O N D

J F M A M J J A S O N D

J F M A M J J A S O N D

J F M A M J J A S O N D

Fig. 9. Mean monthly precipitation and PET for the baseline and 16 C, HadCM3 climatemeteorological inputs sub-catchments identied in Fig. 1).ydrology 486 (2013) 130 19(e.g. Mekong 3), late monsoon precipitation declines, and increasesin monthly totals are limited to May and June. The magnitudes ofthese increases are considerably smaller than those experienced

80

100

120

140

160

180

200

220

240

260

280

PET

(mm

)

100

120

140

160

180

200

220

240

260

280

300

PET

(mm

)

150

170

190

210

230

250

270

290

310

330

350

PET

(mm

)

120

140

160

180

200

220

240

260

J F M A M J J A S O N D

J F M A M J J A S O N D

J F M A M J J A S O N D

J F M A M J J A S O N D

PET

(mm

)

change scenarios. (Note the different y-axis scales. Numbers in brackets refer to the

in these months further upstream (e.g. May increase for the 1 Cand 6 C scenarios are 2.3% and 13.8%, respectively).

Throughout all the sub-catchments, annual PET increases line-arly with prescribed warming (Table 4). Magnitudes of thesechanges are, in general, larger in the southern (warmer) part ofthe catchment compared to the northern (cooler) sub-catchments.PET increases throughout the year with, in most cases, a relativelyconstant climate change signal for each month (Fig. 9). A notableexception is the elevated PET for April in some sub-catchments(e.g. Mekong 1), especially for the larger increases in temperature.In addition, percentage changes in PET early and late in the yearover some northern sub-catchments (e.g. the Lancang) are largerthan those in summer due to the larger changes in temperaturein these months. Absolute values of PET at this time of year remain,however, relatively low.

3.3.2. Changes in river owValues of the mean, Q5 and Q95 discharges for eight gauging

stations for the baseline and the percentage changes in these dis-charges for each of the 16 C, HadCM3 scenarios are shown in Ta-ble 6. Fig. 10 provides the corresponding baseline and scenarioriver regimes. The eight gauging stations are the same as those

August, although for the 4 C scenario at all three stations and the6 C scenario for Vientiane, the highestmeanmonthly discharge oc-curs a month later. The higher dry season ows simulated causeQ95 to increase with the magnitude of prescribed warming,although the size of these increases declines with movementdownstream.

Results for Mukdahan (and Nakhon Phanom, not shown) indi-cate larger mean ows as the magnitude of prescribed warming in-creases although, unlike further upstream, there is not a linearresponse to higher temperatures and changes for the 2 C and3 C scenarios are similar (Table 6). Whilst discharges during theannual rise and recession increase with prescribed warming, peakows in August and September are lower (Fig. 10). These reduc-tions are not related linearly to degree of warming, with the largestassociated with the 4 C scenario. This reects the balance betweenlower precipitation and higher PET in these months. Whilst Q5 de-clines for the 14 C scenarios, very modest increases result fromthe warmest scenarios. Higher dry season ows cause Q95 to in-crease with prescribed warming. These increases are sustainedby ows from upstream rather than local runoff, as precipitationover the Mekong 1 sub-catchment declines as PET increases(Fig. 9). The downstream reduction in the magnitude of increases

angised

(c)

20 J.R. Thompson et al. / Journal of Hydrology 486 (2013) 130used to present results for the 2 C, seven GCM scenarios, and arerepresentative of gauging stations for which results are not shown.

Changes in mean discharge for the 16 C, HadCM3 scenariosfollow the same geographical pattern as changes in annual precip-itation. Mean discharge at the three upstream gauging stations in-creases linearly with prescribed warming (Table 6). This suggeststhat increases in precipitation more than compensate for higherPET. In most cases, mean monthly discharge increases throughoutthe year, with the magnitude of the increase rising with degree ofprescribed warming (Fig. 10). Exceptions are August for the 2 Cand 4 C scenarios at all three stations and the 1 C scenario atLuang Prabang and Vientiane, for which discharge declines slightly(

l of HJ.R. Thompson et al. / JournaPET over this part of the catchment. Mean ow declines consis-tently with degree of prescribed warming (Table 6). Changes are,in percentage terms, as large as those experienced at upstream

0

1

2

3

4

5

6

7

8

J F M A M J J A S O N D

Dis

char

ge (1

06m

3 s-1

)

Chiang Saen (a)

0

2

4

6

8

10

12

14

J F M A M J J A S O N D

Vientiane (c)

0

5

10

15

20

25

30

J F M A M J J A S O N D

Pakse (f)

0

5

10

15

20

25

30

35

40

J F M A M J J A S O N D

Phnom Penh (j)

Dis

char

ge (1

06m

3 s-1

)D

isch

arge

(106

m3 s

-1)

Dis

char

ge (1

06m

3 s-1

)

Fig. 10. River regimes simulated by MIKE SHE for the baseline and 16 C, HadCM3 c(Letters in brackets refer to the gauging station labels used in Fig. 1).ydrology 486 (2013) 130 21gauging stations, albeit of an opposite direction. Particularly largereductions in discharge occur in months with the highest baselineows (Fig. 10), accounting for the further declines in peak ows in

0

2

4

6

8

10

12

J F M A M J J A S O N D

Luang Prabang (b)

0

5

10

15

20

25

J F M A M J J A S O N D

Mukdahan (e)

0

5

10

15

20

25

30

35

40

J F M A M J J A S O N D

Kratie (h)

0.0

0.5

1.0

1.5

2.0

J F M A M J J A S O N D

Mun at Ubon (l)

Dis

char

ge (1

06m

3 s-1

)D

isch

arge

(106

m3 s

-1)

Dis

char

ge (1

06m

3 s-1

)D

isch

arge

(106

m3 s

-1)

limate change scenarios for eight gauging stations within the Mekong catchment.

the Mekong at Pakse. Small increases in mean monthly dischargeare limited to August (the 1 C scenario) and JuneJuly (3 C6 Cscenarios). Declines in discharge throughout most of the year, inparticular during the annual recession and subsequent dry season,demonstrate that increases in discharge within the main Mekongat this time are dependent upon enhanced ows from upstream.

For gauging stations in the lower parts of the Mekong, a consis-tent pattern of changes in discharge is evident. Contributions fromlower tributaries decline in a similar way to those described forUbon. Mean discharge at the four stations on the lower Mekongtherefore declines with prescribed warming and movement down-stream (Table 6). Changes in the river regime are characterised byfurther reductions in peak discharges that are very slightly largeras the magnitude of prescribed warming increases (Fig. 10). Whilstfor the 1 C and 2 C scenarios mean monthly discharge declines inJune and then between August and October, declines occurthroughout the period JuneOctober for the warmer scenarios.The month with the highest mean monthly discharge shifts fromAugust to September for the most extreme (6 C) scenario at allfour gauging stations (and for the 4 C scenario for Phnom Penh).The Q5 discharges decline at all four gauging stations, with themagnitude of these reductions initially increasing with prescribedwarming, but showing less variability between the 3 C and 6 Cscenarios (Table 6). Fig. 10 shows very modest increases in dry sea-son ows at Kratie and Phnom Penh (repeated at Stung Treng andKompong Cham) and as a result the values of Q95 increase. Theseincreases are smaller compared to those for gauging stations

upstream and are again the result of enhanced runoff in upstreamparts of the catchment as opposed to local runoff at this time ofyear. There is less variability in these changes between prescribedwarming scenarios except for the two extremes cases (the 1 C and6 C scenarios).

Fig. 11 shows percentages changes in mean annual dischargefor four gauging stations resulting from the 16 C, HadCM3 sce-narios as well as those which result when one of each of the threemeteorological inputs are modied in turn whilst retaining base-line values for the other two. It conrms the dominant inuenceof change in precipitation over upstream parts of the catchment(e.g. Chiang Saen). Consistent increases in precipitation with pre-scribed warming far outweigh increases in PET and are responsiblefor progressive increases in mean discharge. Further downstream,changes in discharge due to precipitation are smaller and beginto approximate those due to PET (e.g. Mukdahan and especiallyPhnom Penh). For tributaries in the south of the catchment inwhich discharge is not dominated by ows from upstream partsof the main Mekong, changes in PET exert a much larger inuence(e.g. the Mun at Ubon). Results show that mean discharge is rela-tively insensitive to changes in temperature (excluding its inu-ence upon PET), especially in lower parts of the catchment.

3.3.3. Comparison of MIKE SHE results with SLURP and Mac-PDM.09Percentage changes in mean discharge (runoff for Mac-PDM.09)

at six gauging stations (three for SLURP) for each of the 16 C,HadCM3 scenarios, as simulated by the three hydrological models,

60

80

100

char

ge (%

)

All Precip. PET Temp.Chiang Saen (a)

30

40

50

60

char

ge (%

)

All Precip. PET Temp.Mukdahan (e)

M

22 J.R. Thompson et al. / Journal of Hydrology 486 (2013) 130-40

-20

0

20

40

1C 2C 3C 4C 5C 6C

Cha

nge

in m

ean

dis

-40

-30

-20

-10

0

10

20

30

40

1C 2C 3C 4C 5C 6C

Cha

nge

in m

ean

disc

harg

e (%

)

All Precip. PET Temp.Phnom Penh (j)Fig. 11. Percentage change in mean annual discharge simulated by MIKE SHE