Hydrologic Modeling of an Eastern Pennsylvania …kalinla/papers/JHE2006.pdfHydrologic Modeling of...

Hydrologic Modeling of an Eastern Pennsylvania Watershedwith NEXRAD and Rain Gauge Data

Latif Kalin1 and Mohamed M. Hantush2

Abstract: This paper applies the Soil Water Assessment Tool �SWAT� to model the hydrology in the Pocono Creek watershed located inMonroe County, Pa. The calibrated model will be used in a subsequent study to examine the impact of population growth and rapidurbanization in the watershed on the base flow and peak runoff. Of particular interest in this paper is the exploration of potential use ofNext Generation Weather Radar �NEXRAD� technology as an alternative source of precipitation data to the conventional surface raingauges. NEXRAD estimated areal average precipitations are shown to compare well with the gauge measured ones at two climate stationsin the study area. Investigation of the spatially distributed NEXRAD precipitation estimates revealed that average annual precipitation canvary spatially as much as 12% in the Pocono Creek watershed. The SWAT model is calibrated and validated for monthly stream flow, baseflow, and surface runoff. Hydrographs generated from both gauge and NEXRAD driven model simulations compared well with observedflow hydrographs. Although little effort is spent on daily calibration, model simulations and observed flows were in good agreement at thedaily scale as well. Almost similar model efficiency statistics, i.e., mass balance error �MBE�, coefficient of determination �R2�, andNash-Sutcliffe efficiency �ENS�, were obtained during the calibration period in the gauge and NEXRAD driven simulations. In thevalidation period, NEXRAD simulations generated higher model efficiencies at the monthly scale. On the other hand, simulations withgauge precipitations resulted in slightly better model efficiencies at the daily time scale. The spatial representation of precipitation did notcontribute much to model performance when stream flow at the watershed outlet was the required output. However, the use of NEXRADtechnology appears to offer a promising source of precipitation data in addition to currently existing surface gauge measurements.Discussions on new directions in radar-rainfall technology are provided.

DOI: 10.1061/�ASCE�1084-0699�2006�11:6�555�

CE Database subject headings: Watershed management; Radar; Rain gages; Hydrologic models; Calibration; Spatial distribution;Pennsylvania.

Introduction

Distributed hydrologic models are effective tools for environmen-tal decision making and water resources planning and manage-ment. They are helpful not only in making predictions of futureflow conditions, but also in assessment of hydrologic impacts ofchanges in management scenarios, land cover, and climate. Thisstudy calibrates and validates a watershed-scale distributedhydrologic model to the Pocono Creek, which drains 120 km2 ofan area in Monroe County, Pa. Pocono Creek has very good waterand biological quality and has been designated as Special Protec-tion Waters by the State and the Delaware River Basin Commis-

1Assistant Professor of Forest Hydrology, School of Forestry andWildlife Sciences, Auburn University, Auburn, AL 36849; formerly,ORISE Postdoctoral Researcher, U.S. Environmental Protection Agency,National Risk Management Research Laboratory, 26 W. Martin LutherKing Dr., Cincinnati, OH 45268. E-mail: [email protected]

2Research Hydrologist, U.S. Environmental Protection Agency,National Risk Management Research Laboratory, 26 W. Martin LutherKing Dr., Cincinnati, OH 45268 �corresponding author�. E-mail:[email protected]

Note. Discussion open until April 1, 2007. Separate discussions mustbe submitted for individual papers. To extend the closing date by onemonth, a written request must be filed with the ASCE Managing Editor.The manuscript for this paper was submitted for review and possiblepublication on July 28, 2005; approved on April 4, 2006. This paper ispart of the Journal of Hydrologic Engineering, Vol. 11, No. 6,
November 1, 2006. ©ASCE, ISSN 1084-0699/2006/6-555–569/$25.00.
JOURNAL OF HYDROL

sion �DRBC�. The Creek has a wild brown trout population, sig-nificant to outdoor recreation, which is the largest economicgenerator of the region. Population growth and projected urban-ization threaten to stress the natural resources, including thebrown trout habitats in the watershed. The anticipated increase inthe demand for groundwater coupled with the expected decreasein groundwater recharge caused by increased imperviousness areexpected to reduce base flow and increase peak runoff in thePocono Creek.

This paper focuses on a comparative assessment of the use ofNext Generation Weather Radar �NEXRAD� estimated precipita-tion as an alternative to gauge-measured precipitation data. Theimpact on the hydrology of projected future land use changes dueto urbanization in the Pocono Creek watershed will be addressedin a future effort.

The distributed hydrologic Soil Water Assessment Tool�SWAT� �Neitsch et al. 2002a,b�, was chosen to fulfill the pro-ject objectives. The SWAT model was originally developed toquantify the impact of land management practices in large, com-plex watersheds with varying soils, land use, and managementconditions over a long period of time, on the order of years. Ithas been developed and maintained by U.S. Department of Agri-culture �USDA� scientists and is freely available from �http://www.brc.tamus.edu/swat�. It is widely accepted and used, andnumerous applications can be found in the peer-reviewed litera-ture. For instance, as of March 2005, the SWAT model Web sitecited 156 peer-reviewed publications in the form of journal papers
or book chapters ��http://www.brc.tamus.edu/swat/swat-peer-
OGIC ENGINEERING © ASCE / NOVEMBER/DECEMBER 2006 / 555

reviewed.pdf��. Muleta and Nicklow �2005� applied the SWATmodel coupled with automated calibration to estimate daily flowand sediment in a 133 km2 southern Illinois watershed. Eckhardtand Arnold �2001� developed a version of SWAT with a globaloptimization algorithm �SWAT-G� to model daily flow in an81 km2 watershed in Germany. Fohrer et al. �2002� used SWAT-Gin conjunction with two other geographic information system�GIS� based agricultural economy and ecology models in a moun-tainous 60 km2 watershed in Germany to analyze the effect ofland use changes. Sophocleaous and Perkins �2000� integratedSWAT with MODFLOW and applied it to three different Kansaswatersheds to demonstrate different aspects of the integratedmodel. Tripathi et al. �2004� showed on a 92.5 km2 Indian water-shed that SWAT can successfully simulate average annual andmonthly flow and sediment yield even if the weather input isobtained through SWAT’s weather generator.

Precipitation in the form of either rainfall or snowfall is themajor driving mechanism of any hydrologic model. Traditionally,measurements from climate stations or raingauges have been usedas the only reliable source of precipitation in watershed modeling.Such measurements are in fact point measurements, and estima-tion of mean areal precipitation requires a network of numerousrain gauges. The spatial distribution of rainfall may play a crucialrole in some applications, which necessitates a dense network ofrainfall gauges. Mountains influence climate not only on regionaland global scales but also on local scales due to orographic effects�Mukhopadhyay et al. 2003�. As a result, the distributions ofrainfall are more heterogeneous in mountainous regions.Radar-generated precipitation estimates such as from NEXRADproducts have found increasing usage in the hydrologic commu-nity lately as an alternative source to gauge data. NEXRAD datacan provide the much-needed information on spatial distributionof the precipitation pattern.

Although the use of NEXRAD generated precipitation seemsto have clear advantages over the use of gauge-measured precipi-tation due to its capability of capturing the spatial variation ofprecipitation, it is subject to several sources of errors. Precipita-tion �R� is not directly measured, but rather estimated from radarreflectivity �Z� measurements through Z-R relationships. Theexact form of the Z-R relationship is obtained after calibration,yet this introduces significant uncertainty into the estimated pre-cipitation. Further, reflectivity is not a function of volume; it is afunction of surface area, which varies to a much greater extentthan volume when a particle makes its journey to the ground.Other sources of errors in radar-rainfall estimation are groundclutters, anomalous beam propagation, radar beam overshooting,returns from nonweather echoes, etc. �Harrison et al. 2000�.Hence, there is a tradeoff between the advantages, such as spa-tially distributed data, and disadvantages due to possible errors inusing NEXRAD estimated precipitation data.

Several studies employed NEXRAD precipitation data inhydrologic modeling �Smith et al. 2004; Reed et al. 2004�. Nearyet al. �2004� predicted stream flow in two basins in centralTennessee with HEC-HMS using both gauge precipitation andNEXRAD. They concluded that the use of NEXRAD over gaugedata did not offer any improvement. Mukhopadhyay et al. �2003�used NEXRAD generated rainfall data in prediction of the flashflood hazards on coupled alluvial fan-piedmont plain landformswith kinematic wave theory. Using HEC-1, Bedient et al. �2000�computed outflow hydrographs from three major storms and com-pared them to measured hydrographs utilizing NEXRAD andgauge data. Radar proved to be as accurate, and in some cases
more accurate, than the raingauge model. Building on these re-
556 / JOURNAL OF HYDROLOGIC ENGINEERING © ASCE / NOVEMBER/D

search results, Bedient et al. �2003� developed an advanced floodwarning system that relies on NEXRAD for hydrologic predictionin Houston, Tex. Carpenter et al. �2001� employed NEXRAD datain an operational hydrological modeling framework. Results fromthese studies are not conclusive. In other words, it is hard to reacha conclusion based on these studies to promote one precipitationdata source over the other. A dense network of rain gauges mayprovide better spatial representation than NEXRAD. The latter,however, is more accessible by the public and much less costly.

Few studies utilized NEXRAD data within the SWAT model.Jayakrishnan et al. �2005� compared the effect of precipitationinput, i.e., NEXRAD versus rain gauges, on the mean monthlystream flow estimations by SWAT without calibrating the model.They compared simulations with observed stream flows and ob-tained better results with NEXRAD. Di Luzio and Arnold �2004�used NEXRAD precipitation data in modeling 24 storm events bymodifying the SWAT model for hourly simulations. The workpresented in this paper differs from these studies in that the hy-drologic model is calibrated twice, once using rain gauge data asinput, and once with NEXRAD being the input.

The objectives of this paper are: �1� to calibrate and validate awatershed scale model to quantify runoff, base flow, and streamflow in the Pocono Creek; and �2� to explore the potential use ofNEXRAD data as an alternative precipitation data source. Thecalibrated model will later be used to assess future long-termimpacts of the projected land use changes on the sustainability ofthe Pocono Creek stream flow.

In the next section, a summary of the flow component of theSWAT model is presented, followed by information on the studyarea and various data employed in the study. The subsequent sec-tion discusses the methodology employed, specifically base flowseparation, NEXRAD data processing, and calibration and vali-dation of the SWAT model, and it is followed by the conclusionsection.

Summary of SWAT Model

SWAT is a distributed, deterministic process-based hydrologicmodel �Neitsch et al. 2002a,b�. The AVSWAT �Di Luzio et al.2002� graphical user interface �GUI�, which runs under ArcViewGIS, is used to preprocess model data, run the SWAT model, andpostprocess model outputs. SWAT uses readily available inputsand has the capability of routing runoff and chemicals throughstreams and reservoirs, adding flows and input measured datafrom point sources. It is also capable of simulating long periodsfor computing the effect of management changes. Major compo-nents of the model include weather, surface runoff, return flow,percolation, evapotranspiration �ET�, transmission losses, pondand reservoir storage, crop growth and irrigation, groundwaterflow, reach routing, nutrient and pesticide loading, and watertransfer.

The input data needed to run the SWAT model include soil,land use, weather, rainfall, management conditions, stream net-work, and watershed configuration data. AVSWAT has the capa-bility of extracting most of these model parameters from readilyavailable GIS maps such as digital elevation models �DEM�, landuse maps, STATSGO soil maps, etc. Below is a short summary ofthe SWAT model from the model manual and theoretical docu-mentation, version 2000 �Neitsch et al. 2002a,b�.

SWAT partitions the watershed into subunits including sub-basins, reach/main channel segments, impoundments on main
channel network, and point sources to set up a watershed. Sub-
ECEMBER 2006

basins are divided into hydrologic response units �HRUs�, whichare portions of subbasins with unique land use/management/soilattributes. AVSWAT enables extraction of input parameters easily.It uses digital elevation models �DEM� as input to extract thechannel network and delineate the watershed and subwatersheds,the resolution of which depends on the user-provided thresholdarea, which is the area required to initiate a first-order channel.The threshold area can be chosen in such a way that the resultantchannel network resembles the one provided in topographic maps.The user needs to provide two threshold values to create HRUs,one for land use and one for soil. Land uses that cover a percent-age of the subbasin area less than the threshold level are consid-ered minor and thus eliminated. After the elimination process, thearea of the remaining land uses is reapportioned so that 100% ofthe land area in the subbasin is modeled. The soil threshold isapplied next in a similar fashion to eliminate minor soil types thatoccupy negligible portions of the HRUs.

In SWAT, the land phase of the hydrologic cycle is based onthe water balance equation

SWt = SW0 + �i=1

t

�Rday,i − Qsurf,i − Ea,i − wseep,i − Qgw,i� �1�

where SWt=final soil water content �mm�; SW0= initial soilwater content �mm�; t=time �days�; Rday,i=amount of precipi-tation on day i �mm�; Qsurf,i=amount of surface runoff on dayi �mm�; Ea,i=amount of evapotranspiration on day i �mm�;wseep,i=amount of percolation and bypass flow exiting the soilprofile bottom on day i �mm�; and Qgw,i=amount of return flowon day i �mm�.

Snowmelt is included with rainfall in the calculation of runoffand percolation; it is controlled by the air and snow pack tem-perature, the melting rate, and the areal coverage of snow. Themass balance for the snow pack is given by

SNOt = SNO0 + Rday − Esub − SNOmlt �2�

where SNOt=water content of the snow pack at the end of aday �mm�; SNO0=initial water content of snow pack �mm�;Rday=amount of precipitation on a given day �mm�;Esub=amount of sublimation on a given day �mm�; andSNOmlt=amount of snowmelt on a given day �mm�. This equationassumes that the water released from snowmelt is evenly distrib-uted over the 24 h of the day.

SWAT uses the USDA Soil Conservation Service �SCS� curvenumber method �USDA 1972� or the Green and Ampt infiltrationmethod �Green and Ampt 1911� to compute surface runoff vol-ume for each HRU. The former option is utilized in this study.The SCS runoff equation is an empirical model that was devel-oped to provide a consistent basis for estimating the amounts ofrunoff under varying land use, soil types, and antecedent moistureconditions �Rallison and Miller 1981�. The SCS curve numberequation is

Qsurf =�Rday − 0.2S�2

�Rday + 0.8S��3�

where S=retention parameter �mm�. In this equation the initialabstractions, which include surface storage, interception, and in-filtration prior to runoff, are approximated as 0.2S. The retention
parameter is defined as
JOURNAL OF HYDROL

S = 25.4�1,000

CN− 10� �4�

in which CN=curve number for the day, which is a function ofthe soil’s permeability, land use, and antecedent soil waterconditions.

Evapotranspiration �ET� is the primary mechanism by whichthe water is removed from a watershed. It includes all processesby which water at the earth’s surface is converted to water vapor:evaporation from the plant canopy, transpiration, sublimation,and evaporation from the soil. SWAT calculates actual ET frompotential evapotranspiration �PET�. The latter is estimated bythree methods in SWAT: �1� the Penman-Monteith method; �2� thePriestly-Taylor method; and �3� the Hargreaves method. ThePenman-Monteith method requires solar radiation, air tempera-ture, relative humidity, and wind speed. The Priestly-Taylormethod requires solar radiation, air temperature, and relative hu-midity. The Hargreaves method requires air temperature only.Penman-Monteith is used in this study.

Once surface runoff is calculated, the amount of surface runoffreleased to the main channel is computed from

Qch,i = �Qsurf,i + Qstor,i−1� · �1 − e−surlag/tconc� �5�

where Qch,i=amount of surface runoff discharged to the mainchannel on day i �mm�; Qsurf,i=amount of surface runoff gener-ated in the subbasin on day i �mm�; Qstor,i−1=surface runoff storedor lagged from day i−1 �mm�; surlag=surface runoff lag coeffi-cient; and tconc= time of concentration for the subbasin �hs�. Thelast expression within the bracket on the right-hand side of Eq. �5�represents the fraction of total available water allowed to enter thereach on a given day. Remaining water becomes available waterfor the next day �Qstor,i�.

The movement of water through the channel network of thewatershed to the outlet is routed in the main channel and reser-voirs. Flow is routed through the channel using either a variablestorage coefficient method developed by Williams �1969� or theMuskingum routing method, the latter of which is employed inthis study. Transmission losses, which reduce runoff volume asthe flood wave travels downstream, are also accounted for by themodel. SWAT is a continuous time model, i.e., a long-term yieldmodel. The model is not designed to simulate detailed, single-event flood routing.

The water balance for the shallow aquifer is

aqsh,i = aqsh,i−1 + wrchrg,i − Qgw,i − wrevap,i − wdeep,i − wpump,sh,i

�6�

in which aqsh,i=amount of water stored in the shallow aquifer onday i �mm�; aqsh,i−1=amount of water stored in the shallow aqui-fer on day i−1 �mm�; wrchrg,i=amount of recharge entering theaquifer on day i �mm�; Qgw,i=groundwater discharge, or baseflow, into the main channel on day i �mm�; wrevap,i=amount ofwater moving into the soil zone in response to water deficiencieson day i �mm�; wdeep,i=amount of water percolating from theshallow aquifer into the deep aquifer on day i �mm�; andwpump,sh,i=amount of water removed from the shallow aquifer bypumping on day i �mm�. The recharge to the aquifer on a givenday is calculated by

wrchrg,i = �1 − e−1/�gw� · wseep,i + e−1/�gw · wrchrg,i−1 �7�

where �gw=delay time or drainage time of the overlying geologic
formations �days�; wseep,i=total amount of water exiting the bot-

tom of the soil profile on day i �mm�; and wrchg,i−1=amount ofrecharge entering the aquifer on day i−1 �mm�.

Base flow is computed by SWAT using the following equation:

Qgw,i = e−�gw�t · Qgw,i−1 + �1 − e−�gw�t� · wrchrg,i �8�

where Qgw,i−1=groundwater flow into the main channel on dayi−1 �mm�; �gw=base flow recession constant; and �t=time step

Fig. 1. Pocono Creek watershed: �a� location, topography, climate s�STATSGO�

�1 day�.


Study Area and Data

The 120 km2 Pocono Creek watershed is located between lati-tudes 40°59�N–41°06�N and longitudes 75°14�W–75°26�W inMonroe County, eastern Pennsylvania, near the New Jersey stateborder �Fig. 1�a��, within the Delaware River Basin. PoconoCreek’s 26-km-long watershed valley drains from the Pocono Pla-

s, and NEXRAD cell centroids; �b� land use map; and �c� soil map
tation
teau in its headwaters eventually into the Brodhead Creek, a tribu-

ECEMBER 2006

tary into the Delaware River. The model is constructed for thearea above the U.S. Geological Survey �USGS� stream gauge,which is located about 6.4 km upstream from the mouth near thecity of Stroudsburg, Pa. �Fig. 1�a��, and drains an area about98.87 km2.

Two physiographic provinces encompass the Pocono Creekbasin: the Appalachian Plateau in the northern part of the water-shed and the Valley and Ridge in the southern part. The Appala-chian Plateau province is categorized by gently folded rocks ofDevonian age, and more than 75% of the watershed lies within itsboundaries. The Appalachian Plateau province is subdivided bythe Pocono Escarpment into the Pocono Plateau section in thenorthwest and the Glaciated Low Plateau section to the southeast.The remaining 25% of the basin is occupied by the AppalachianMountain section of the Valley and Ridge province. More in-tensely deformed sedimentary rocks also of Devonian age char-acterize this part of the catchment �Pocono Creek Pilot Project2001�.

Table 1 summarizes the data used in this study along with theirsources and formats. Precipitation, temperature, humidity, andwind speed data were obtained from two National WeatherService �NWS� climate stations: Mount Pocono to the north andStroudsburg to the east. As can be seen from Fig. 1�a�, bothstations are outside the watershed boundary. To study the poten-tial effects of this on model performance, NEXRAD data werealso utilized during the study as an alternative precipitationdata source. Specifically, the XMRG products produced by theMiddle Atlantic River Forecast Center �MARFC: �http://www.erh.noaa.gov/er/marfc�� were used. XMRG precipitationfiles are generated in a specific file format after analyses fromboth gauges and radar with some manual quality control and areavailable at approximately 4 km cell resolutions. The smallsquares with dots inside in Fig. 1�a� represent the locations ofthe centroids of the NEXRAD precipitation cells. The XMRGfiles for the MARFC region can be downloaded from �http://dipper.nws.noaa.gov/hdsb/data/nexrad/marfc�mpe.php�.

The current land cover �Fig. 1�b�� is dominantly forest �89%�.Pasture constitutes about 3.5% and there is almost no agriculturalactivity ��0.2% �. Residential, commercial, and transportationareas comprise about 5.8% of the watershed, including the com-mercially developed Route 611 corridor, Big Pocono State Park,Camelback Ski Area, the Nature Conservancy’s TannersvilleCranberry Bog, and state gamelands. Silt loam is the major soiltype in the watershed covering about 84.9%. Two other soil typesin the watershed are sandy loam �11.4%� and loam �3.7%�. Theelevation in the watershed changes from 183 m at the outlet to

Table 1. Data Used in Study and Their Sources

Data Source

Soil SWAT build in

Land use DRBC

Elevation DRBC provided, also available at http://www.pasda.

Climate NOAA National Data Center, http://nndc.noaa.gov

Radar NWS Hydrologic Data Systems Group, http://dippernoaa.gov/hdsb/data/nexrad/nexrad.html

Stream flow http://waterdata. usgs.gov/pa/nwis/uv?dd cd=02&for=pre&period=16&site�no=01441495

648 m near the Camel Back Ski Area. A 30 m resolution digital

JOURNAL OF HYDROL

elevation model �DEM� is used in extraction of the stream net-work and delineation of the watershed and subwatersheds. Thewatershed is divided into 29 subbasins. The STATSGO soildatabase was used to acquire any soil-related model parameters�Fig. 1�c��. Hydrologic response units �HRU� are generated byusing 10 and 0% thresholds for land use and soil maps that re-sulted in a total number of 129 HRUs.

Methodology

Base Flow Separation

Stream flow is usually partitioned into two parts, the fast and theslow response components, the latter of which is due to the baseflow contribution. Any other contribution to stream flow by vari-ous mechanisms can be deemed as the fast response component.SWAT computes the base flow and surface runoff components ofthe stream flow separately. Although some parameters play aninteractive role, such as the CN, some parameters only affect oneindividual component of the flow. For instance, Manning’s rough-ness only affects surface runoff, whereas the base flow recessionconstant only affects base flow. Hence, to better calibrate modelparameters, it is necessary to partition the observed stream flowhydrograph into base flow and surface runoff components, both ofwhich then become estimated quantities rather than observed.

We estimated base flow using the method described in Arnoldet al. �1995� and Arnold and Allen �1999�. This is a recursivedigital filter technique that was originally used in signal process-ing and filtering. The filter can be passed over the stream flowdata several times, with each pass resulting in less base flow.Various other methods of base flow separation are also available,and each can give significantly different base-flow estimateswhich clearly affect not only model calibrated parameters but alsomodel performance.

Next Generation Weather Radar

Both the Mount Pocono and Stroudsburg climate stations are lo-cated outside the watershed boundary �Fig. 1�a�� and providepoint measurements. This raises the question whether the datafrom these two gauge stations are really representative of spatialrainfall pattern in the Pocono Creek watershed, which is partlycovered with mountains. It is well known that precipitation maychange significantly in mountainous areas. Precipitation is usuallyhigher in upper elevations. To be able to realistically answer this

Additional information

USDA STATSGO soil database

National Land Cover Data �NLCD�

/ 30 m resolution DEM

Stations: Mount Pocono, 41°08�N/75°23�W; Stroudsburg,41°01�N/75°11�W; hourly �daily for Stroudsburg�precipitation, temperature, wind speed, relative humidity

NEXRAD, Multisensor Precipitation Estimator �MPE� Data inXMRG format; hourly precipitation at approximately 4 kmresolution

USGS 01441495 Pocono Creek at Wigwam Run nearStroudsburg, Pa.; latitude 40°59�27�, longitude 75°15�20�,data collected every 15 min

psu.edu

. nws.

mat

question, several rain gauges inside the watershed boundary are


needed so that spatial distribution of the precipitation pattern canbe attained. The NEXRAD precipitation data do provide a spatialrepresentation of the precipitation distribution over the entire wa-tershed. They also provide a finer temporal resolution �hourly� ascompared to daily values from the climate stations that are avail-able. In any case, hourly precipitation is not required for thisstudy and daily precipitations suffice for the current model appli-cation. Because our ultimate goal is modeling base flow andstream flow in the Pocono Creek watershed, we compared gauge-driven model performances to NEXRAD driven ones. This shedslight on whether consideration of the distributed nature of precipi-tation improves the model performance or not and whetherNEXRAD can be relied on as an alternative to the surface raingauge measurements.

There are four levels of NEXRAD precipitation products, cat-egorized based on the amount of preprocessing, calibration, andquality control performed �Xie et al. 2005�. The Stage I product isa digital precipitation array �DPA�, also called hourly digital pre-cipitation �HDP�, and it is derived from reflectivity measurementsusing a Z-R �reflectivity-precipitation� relationship after the appli-cation of several quality control algorithms. The Stage II productis HDP combined with hourly rain gauge observations for a singleradar site and has mean field bias correction. Therefore, Stage IIprecipitation estimates are superior to Stage I estimates. Stage IIIdata are obtained by combining Stage II data from multipleweather radars covering an entire River Forecast Center �RFC�and provides estimates for each Hydrologic Rainfall AnalysisProject �HRAP� cell. Stage III was created specifically for theNWS river forecast centers, which need rainfall estimates over amuch larger area than covered by an individual radar. Stage IIIproducts are also mosaicked to cover the entire ContinentalUnited States, and the end product is named Stage IV �Xie et al.2005�. Among these products, Stage III is the most widely usedone.

NEXRAD Data Processing

The data type we used in this study is called MPE �multisensorprecipitation estimate�, which is roughly equivalent to Stage IIIlevel data. They come in compressed and in the XMRG format,which is hourly precipitation estimates on an HRAP grid gener-ated by MPE and Stage III. There is one tar file for each month,and within each there is one tar file for each day, then multiplebinary files, each with 1 h precipitation, also compressed. Theprocess requires multiple unzipping and untarring. For instance,from 6/21/2002 to 4/30/2005 �this is the time period of flow mea-surements at the Pocono Creek USGS stream gauge�, there aremore than 25,000 individual hourly compressed files. To be ableto use them, they all need to be first uncompressed and thenconverted from binary to ASCII file format. We used two scriptswritten in PERL, which is a high-level programming languagecommonly used in file processing and text manipulation, tohandle this situation. After uncompressing all the hourly files withthe first script, the second script converted them all from binaryto ASCII format by calling the C program “xmrgtoasc.c.” Onceuncompressed and converted to ASCII, they occupied a spaceof approximately 5 GB. Hourly precipitation files contain 200columns and 200 rows, with each entry representing the arealaverage precipitation of an area approximately 4�4 km. TheseXMRG files come in a coordinate system called HRAP �Hydro-logic Rainfall Analysis Project�, which is basically a polar stereo-graphic map projection that is formed on a plane intersecting a
spherical earth datum at 60° N and is not recognized by many

GIS software packages, including ArcView. Reed and Maidment�1999� present details about how the HRAP coordinate is definedand provide insights as to why displaying HRAP grids with GISdata has been a point of confusion. The link at �http://www.nws.noaa.gov/oh/hrl/distmodel/hrap.htm� is an excellentreference explaining how to perform coordinate transformationsand provides the tools for doing so. Readers can also refer to therecent paper of Xie et al. �2005� for a detailed explanation of anautomated procedure to process NEXRAD data.

As mentioned in the previous paragraph, there are 200�200=40,000 data entries in each hourly precipitation file. Amongthese we only need the ones covering the Pocono Creek water-shed. Therefore, we want to determine which HRAP cellsfall within the Pocono Creek watershed. This can be done inArcInfo using the Arc Macro Language �AML� script described inXie et al. �2005� or alternatively, as we did, in ArcView using theextension “coord.avx.” This script gives users the option to createa Shapefile of HRAP center points in HRAP coordinates, polarstereographic coordinates, or geographic coordinates �lon-lat�,and it can be downloaded from the link provided in the previousparagraph. As can be seen from Fig. 1�a�, merely 12 HRAP cellsfall within or near the boundary of the study watershed. Thus, weextracted the hourly precipitation estimates only for those 12cells.

There are two key points worth mentioning. First, the timesgiven for the XMRG files are in Universal Time �5 h ahead ofU.S. Eastern Standard Time�, not the local time. For instance, theXMRG filenames look like “xmrg�12172004�07z�MA,” and thisrepresents the total precipitation occurring on 12/17/2004 from7:00 to 8:00 AM in Universal Time �2:00–3:00 AM in EasternTime�. Secondly, the C code “xmrgtoasc.c” may not work withsome of the older XMRG files downloaded from NWS �prior to2/1/2004�, which may be due to formatting inconsistencies. Theconverted ASCII files for that time period were obtained fromNWS through personal communication.

Model Calibration

Models are only simplified representations of natural processes.Even fully physically based models cannot avoid simplifications.Adding into that parameter uncertainties and measurement errors,model calibration becomes inevitable in any modeling exercise. Asplit data set approach is implemented in this study to calibrateand validate the SWAT model. The period from 7/1/02 to 5/31/04of the daily flow data, aggregated into monthly flows, is used forcalibration and the remaining data from 6/1/04 to 4/30/05 is usedfor validation. Three to five years of data are typically required incalibration, although with a good set of data and the proper ob-jective function, a single year of data has been shown to be ad-equate �Sorooshian et al. 1983�. The statistical measures of massbalance error �MBE�, coefficient of determination �R2� and Nash-Sutcliffe �Nash and Sutcliffe 1970� efficiency �ENS� are used asindicators of model performance, which are defined as

MBE =

�i=1

N

Qsim,i − �i=1

N

Qobs,i

�N

Qobs,i

=Qsim − Qobs

Qobs

�9�

i=1

ECEMBER 2006

R2 =

��i=1

N

�Oobs,i − Oobs��Osim,i − Osim�2

��i=1

N

�Oobs,i − Oobs�2��i=1

N

�Osim,i − Osim�2 �10�

ENS = 1.0 −

��i=1

N

�Osim,i − Oobs,i�2��

i=1

N

�Oobs,i − Oobs�2 �11�

where Qsim,i and Qobs,i=simulated and observed �or estimated�flows at the ith observation, respectively; and N=number of

observations. Similarly, Osim and Oobs=average simulated andobserved flows over the simulation period. The coefficient of de-termination describes the proportion of the total variances in theobserved data that can be explained by the model and ranges from0 to 1, whereas Nash-Sutcliffe efficiency is a measure of how wellthe plot of observed versus predicted values fit the 1:1 line andcan vary from −� to 1. A negative ENS indicates that model pre-dictions are not better than the average of the observed data.

The calibration process followed is adapted from Santhi et al.�2001� and is summarized in Fig. 2. The threshold values setfor MBE, R2, and ENS in the figure are at monthly time scales.Curve numbers �CN� of each soil/land use combination are cali-brated first to meet the criteria set for surface runoff �SR�.At the next stage the, GW_REVAP coefficient, which is a limitingfactor for the maximum amount of water that can be removedfrom the aquifer to the overlying unsaturated zone due to mois-ture deficit, threshold depth of water in shallow aquifer requiredfor base flow to occur �GWQMN�, delay time for aquifer recharge�GW_DELAY�, and soil evaporation compensation factor�ESCO� are calibrated to meet the stream flow �SF� and base flow�BF� criteria. Once the model is calibrated using measured stream

Fig. 2. Calibration procedure in the SW

flows for the calibration period, the model is validated using the

JOURNAL OF HYDROL

data for the previously stated validation period, and MBE, R2, andENS are checked against the threshold values shown in Fig. 2. Thethreshold values set for the MBE, R2, and ENS statistical measuresmust be satisfied for a successfully calibrated model.

Before calibrating the model parameters, we performed anautomated sensitivity analysis with a version of SWAT calledAVSWATX that is based on Latin hypercube �LH� and one factorat a time �OAT� sampling. Unfortunately, AVSWATX performssensitivity analysis only for SF volume or average SF. Fig. 3depicts the sensitivity of average stream flow to various modelparameters. It is clear that CN is the most sensitive model param-eter. An interesting observation from the figure is that the order ofsensitivity of model parameters may change with the precipitationdata. Fig. 3 shows parameter sensitivities for 2, 5, and 24 years ofprecipitation data, with shorter durations being the subset of thelonger durations. Consider the parameter GW_REVAP as an ex-ample. Although it is the fifth most sensitive parameter when

Fig. 3. Sensitivity of stream flow volume to variousSWAT model parameters. Note that order of sensitivity changes withprecipitation duration.

del �adapted from Santhi et al. �2001��
AT mo

24 years of precipitation data is used, it is not even in the picturewhen 2 years of precipitation data are used. The order of sensi-tivity in Fig. 3 is important when calibrating the model to have alow MBE of SF. However, R2 and ENS are sensitive not only tovolume but also timing of flow. We did not perform sensitivityanalysis for the latter and relied on the model developers’ expe-riences as depicted in Fig. 2 in calibrating model parameters.

Simulations with Climate Stations as PrecipitationData Source

In calibrating the SWAT model we first relied on the two climatestations as the precipitation data source, as shown in Fig. 1�a�:

Table 2. Mass Balance Error �MBE�, Coefficient of Determination �R2�,Surface Runoff �SR� during Calibration Period. Climate Stations Are Use

Time scale

Stream flow

MBE�%� R2 ENS

MB�%

Monthly −3.8 0.85 0.83 −4.

Daily 0.74 0.74

Fig. 4. Observed stream flow and estimated surface runoff plotted ag04�. Local climate stations are used as the precipitation data source. T


Mount Pocono and Stroudsburg stations. Model parameters arecalibrated at the monthly time scale as described in Fig. 2. Fig. 4compares the observed SF and estimated SR with SWAT simu-lated counterparts for the calibration period 7/1/2002 to 5/31/2004. We actually started simulations from 7/1/1970 with mea-sured precipitation data as input. This corresponds to a warm-upperiod of 32 years, which is uncharacteristically long. The ideabehind using long warm-up periods is to minimize the effect ofinitial conditions such as antecedent moisture, initial groundwatertable height, etc., which are often not known. We decided on thislong period after experimenting with various warm-up periods.Daily simulations are performed as well, the results of which for

ash-Sutcliffe Efficiency �ENS� for Stream Flow �SF�, Base flow �BF� andrecipitation Data Source.

Base flow Surface runoff

R2 ENS

MBE�%� R2 ENS

0.30 0.08 −3.6 0.77 0.77

WAT simulated counterparts for the calibration period �7/1/02–5/31/panels show monthly results, and bottom panel shows daily results.

and Nd as P

E�

0

ainst Sop two

ECEMBER 2006

SF are also shown in Fig. 4. We had to adjust one more parameter,the surface runoff lag coefficient �surlag�, to improve the dailysimulation performances.

Both the monthly and daily SF simulation results match wellto the observed SF. SWAT also generated monthly SR values,comparable to SR estimated from observed stream flow. Thehighest monthly volume error is during the month of June 2003,where SWAT underestimates SR by 42%. Table 2 summarizesMBE, R2, and ENS at both the monthly and daily time scales forthe calibration period. The threshold efficiency measures set forcalibration are met at the monthly scale. SWAT slightly underes-timates the total volume of SF, SR, and BF, as all have negativeMBE. Although not required as a performance criteria, R2 and ENS

of BF are also computed and shown in Table 2 for completeness.The fact that the statistical measures R2 and ENS computed for SFat the daily time scale meet the criteria is surprising, as SWAT isrecommended for long-term simulations, at no less than amonthly time scale.

Although the results are promising, in the next section weappraise the use of radar estimated precipitation, which in the

Fig. 5. Comparison of NEXRAD estimated precipitation with

absence of gauge data is the only available precipitation data

JOURNAL OF HYDROL

source. Even if gauge measurements are available, some studiesindicate that radar data may still be preferable and provide bettermodel performances �e.g., Bedient et al. 2000; Jayakrishnan et al.2005�. We did not have hourly precipitation data for the Strouds-burg gauge station, and the daily precipitations are actually re-ported from 7:00 a.m. to 7:00 a.m., meaning that they do notreally represent the calendar day. Clearly this does not affectmonthly simulation results, unless there is a very significant stormthat happened to start at the end of a month and end at the begin-ning of the following day. Nevertheless, at the daily time scale,this nuance can play a more important role and the use ofNEXRAD in this regard may lead to improved results.

Simulations with NEXRAD

Before running the SWAT model with NEXRAD data, we wantedfirst to see how well NEXRAD estimates precipitation at the twoclimate stations. Fig. 5 compares the daily and monthly areallyaveraged NEXRAD estimated precipitations to measurements atthe gauges. All the values in the figure are in mm. Mount Pocono

urements from local gauge stations. Values shown are in mm.
meas
station falls in between two NEXRAD cells �Fig. 1�a��, �961,614�


and �961,615�, where the numbers in parentheses represent the Xand Y coordinates of the centroids of the NEXRAD cells in theHRAP coordinate system. Therefore, we compared Mount Po-cono point precipitation measurements to NEXRAD estimates atboth cells as well as to their arithmetic averages. The Stroudsburgstation falls inside the grid �966,614�, as seen in Fig. 1�a�. Esti-mated hourly precipitations were aggregated in these cells toobtain the daily and monthly precipitation estimates. For a real-istic comparison with the Stroudsburg data, we computed dailyvalues from 7:00 a.m. to 7:00 a.m. at the cell �966,614�. Fig. 5reveals that �961,615� represents Mount Pocono relatively betterthan both �961,614� and the average of the two, as it has a higherR2, the slope of its regression equation is closer to 1, and it has asmaller intercept. Further, �961,615� overestimates precipitationby only 5.7%, as compared to the 10.8% overestimation of�961,614� from 1/1/2002 to 4/30/2005. Over the time period from1/1/2002 to 2/28/2005, NEXRAD underestimates precipitationin the Stroudsburg station by 10.0%. In this watershed, com-parisons with gauge measurements indicate that NEXRAD tech-nology provides an alternative source of precipitation data. Notethat NEXRAD estimates are areal average precipitations of anapproximately 4�4 km2 grid, in contrast to point measurements�100 cm2� of gauge stations. Differences between the two arepartly attributed to variations of point measurements from arealaverage estimates of precipitations, and partly due to measure-ment errors associated with the instrument itself.

Fig. 6 depicts the annual average precipitation distributionwithin the Pocono Creek watershed, derived from 3.5 years ofavailable NEXRAD data. The precipitation contours are obtainedby interpolation, where precipitation data are assumed to be lo-cated at the centroids of the HRAP grid cells. The bull’s-eyeappearance in the figure is a result of this interpolation and theuse of the full black and white spectrum over the range of annualprecipitation values. Within the watershed boundary, average an-

Fig. 6. Average annual precipitation distribution within

nual precipitation varies from 1,503 to 1,687 mm. This corre-


sponds up to a 12% variation. A topographic map of the basin isalso shown in the same figure. Close inspection reveals that pre-cipitation increases from the watershed outlet in the northwestdirection, where it peaks at the highest point of the basin �wherethe bull’s eye is�, and from there on decreases to the north andnorthwest. It appears that storms moving inland from the AtlanticOcean in the northwest direction are dumping more rain whenthey hit mountain ranges and hillsides. It is obvious here thatprecipitation is not spatially uniform within the Pocono Creekwatershed because of its unique topographical features.

At the next stage, we calibrated the SWAT model usingNEXRAD as the precipitation data source. The only parameteradjusted is the CN. We reduced the previously gauge-driven, cali-brated CN values in all HRUs by 1.5. The volume of precipitationover the study watershed estimated with NEXRAD was greaterthan that estimated from rain gauges, and thus this adjustment inCN had to be made. We also experimented with other parameters,but they did not further improve the model performances. Fig. 7compares SWAT simulations to observed SF and estimated SR atthe monthly time scale and SF at the daily time scale. Overall themodel performs well at both time scales, and the results are com-parable to the gauge-driven simulations illustrated in Fig. 4. TheNEXRAD driven model performance statistics MBE, R2, and ENS

summarized in Table 3 compare well to their counterparts givenin Table 2. From the results obtained, and despite a relativelysmaller calibrated CN, it is reasonable to draw the conclusion thatNEXRAD estimated precipitation data is a good alternative togauge measured precipitation data. The benefit of using distrib-uted rainfall data was negligible at both time scales when flow atthe watershed outlet is concerned. On the other hand, noticeableimprovement is expected for stream flow estimates at the sub-watershed outlets with the use of distributed rainfall data over theavailable rain gauge data �see Fig. 6 for spatial distribution ofprecipitation in the Pocono Creek watershed�. These two obser-

cono Creek watershed with comparison to topography.
the Po
vations are based on the calibration period. In the next section, we

ECEMBER 2006

use an 11 month period split sample data set to further validatethe SWAT model with both gauge and NEXRAD data.

Model Validation

The period from 6/1/2004 to 4/30/2005 is chosen as the validationperiod. One storm event immediately draws attention in this pe-riod. On April 2 and 3, 2005, NEXRAD estimated averages of104 and 10 mm of rain over the watershed, respectively. On thesame days, the climate stations estimated about 115 and 10 mm ofrain. The April 2 event is truly an extreme event, with a returnperiod of approximately 10 years. Although there has not been

Fig. 7. Observed stream flow and estimated surface runoff�7/1/02–5/31/04� period. NEXRAD estimates are used as the precipitshows daily results.

Table 3. Mass Balance Error �MBE�, Coefficient of Determination �R2�,Surface Runoff �SR� during Calibration Period. NEXRAD Data Is Used

Time scale

Stream flow

MBE�%� R2 ENS

MB�%

Monthly −3.8 0.85 0.84 −4.

Daily 0.74 0.73

JOURNAL OF HYDROL

any significant amount of rain during the previous three days, theobserved stream flows on April 2 and 3 were 80 and 87 mm,respectively. Contrary to the total measured 125 mm of precipita-tion �conservative estimate based on rain gauges� on April 2 and3, the USGS stream gauge reported 167 mm of total flow on thesedates. Even if all the precipitation fallen became stream flow,there is still about 42 mm unaccounted for. Thus, we can con-clude that the measured stream flow �what is really measured isthe stage of the flow; flow discharges are estimated from stagemeasurements using rating curves� is probably overestimated. Anerror in the precipitation seems less likely, as NEXRAD estimates

ted against SWAT simulated counterparts for the calibrationata source. Top two panels show monthly results, and bottom panel

sh-Sutcliffe Efficiency �ENS� for Stream Flow �SF�, Base Flow �BF� andipitation Data Source.


R2 ENS

MBE�%� R2 ENS

0.31 0.05 −3.2 0.79 0.79

plotation d

and Naas Prec

E�

5


and gauge measurements are consistent. As a precautionary step,we downloaded stream flow data from the USGS web site as theybecame available, and as of 1/30/2006, data after 9/30/2004 wasstill listed as provisional, meaning no quality check had beenperformed yet. Fig. 8 shows stream flow data downloaded fromUSGS at two different times. It can clearly be seen that USGSmakes adjustments after making quality control of the data. Notethat the initially posted stream flow values are later reduced, es-pecially in the winter months.

Next, simulations are performed again starting from 7/1/1970to curtail the effect of initial conditions. Figs. 9 and 10 compareobserved values to model simulations conducted with gauge andNEXRAD data, respectively. In both figures the top two panelscompare monthly SF and SR, and the bottom figure is for dailySF results. From visual inspection one can conclude that modelsimulations match pretty well with observed SF and estimatedSR. Table 4 summarizes the model efficiencies at the monthly anddaily time scales, respectively. In the table, g denotes gauge inputand n indicates NEXRAD input precipitation. The subscript x in

Fig. 8. Comparison of stream flow data downloaded from the USG

Fig. 9. Observed stream flow and estimated surface runof�6/1/04–4/30/05� period. Local climate stations are used as the precipirespectively.


the table is used to refer to the case where the validation periodexcludes the month of April 2005. In all cases the model calibra-tion criteria are satisfied. Conversely, when NEXRAD is used asthe precipitation data source, the maximum MBE is 7.9%, whichis about half the error obtained when gauge measurements areused as input. Close inspection of Figs. 9 and 10 reveals thatSWAT significantly underestimates flow in the month of April2005, as expected, especially during the first few days of April. Asimilar situation of model simulation overestimating the observedflow happened during the calibration period on June 2003. How-ever, the calibration period was long enough to average out thiseffect �compare 23 months of calibration data versus 11 monthsof data for validation�.

We computed in Table 4 the model efficiencies when April2005 is excluded from the validation period, considering it asunreliable. Subscript x is used to symbolize this scenario. As canbe seen, MBE in SR decreased from 15.1 to 5.9% with the gaugeinput simulations. In SF and BF, it dropped to 5.5 from 13.5% andto 4.8 from 10.6%, respectively, all underestimations. MBE in SF

fferent times. Major adjustments are clearly seen in winter months.

ted against SWAT simulated counterparts for the validationdata source. Top and bottom panels denote monthly and daily results,

S at di

f plottation

ECEMBER 2006

and SR also got smaller in NEXRAD simulations after excludingApril 2005 data but increased in BF. The ENS values have im-proved as well in monthly outputs of SR and SF, and R2 improvedslightly overall. Daily statistics have mixed results. Removal ofthe April 2 event did not have a significant impact on R2 butimproved ENS with both gauge and NEXRAD simulations. It isexpected that, as more data become available, the high mass bal-ance error caused by the April 2005 event in the validation periodwill diminish due to averaging over a longer period.

It is interesting to note that, while at the monthly time scaleNEXRAD had better performance statistics, at the daily timescale, gauge-driven simulations had slightly better statistics. Sev-eral reasons may have contributed to this. First, calibration effortswere more focused to a monthly time scale as part of the projectgoal. Secondly, NEXRAD calibration was built on the calibratedparameters with gauge-driven data; therefore, there is a smallbias.

Fig. 10. Observed stream flow and estimated surface runof�6/1/04–4/30/05� period. NEXRAD estimates are used as the precipitrespectively.

Table 4. Mass Balance Error �MBE�, Coefficient of Determination �R2�,Surface Runoff �SR� during Validation Period. g and n Represent Gauge aCase Where April 2005 Is Excluded.

Time scalePrecipitation

source

Stream flow

MBE�%� R2 ENS

Monthly g −13.5 0.81 0.66

gx −5.5 0.83 0.79

n −5.6 0.89 0.75

nx 3.3 0.94 0.87

Daily g 0.70 0.64

gx 0.72 0.71

n 0.66 0.62

nx 0.69 0.69

JOURNAL OF HYDROL

Summary and Conclusions

In this paper we modeled the hydrology in Pocono Creek water-shed using a distributed parameter watershed model, and in par-ticular we explored the potential for using the NEXRAD as analternative source of precipitation data to rain gauge stations. Wecalibrated and validated the SWAT model for the Pocono Creekwatershed in eastern Pennsylvania, and the model will be used ina subsequent effort to investigate the impact of population growthand projected future land use changes in Monroe County on thesustainability of water resources in the watershed.

The model was first calibrated using precipitation data fromtwo gauge stations located outside the watershed boundary. Weevaluated the model performance by visually comparing modelgenerated SF and SR hydrographs to observed SF and SR hydro-graphs, which was estimated from SF by separating BF, as well asthrough computed model efficiency statistics, i.e., mass balance

tted against SWAT simulated counterparts for the validationata source. Top and bottom panels denote monthly and daily results,

sh-Sutcliffe Efficiency �ENS� for Stream Flow �SF�, Base Flow �BF� andXRAD Input Precipitations, Respectively. Subscript x Is Used to Denote


MBE�%� R2 ENS

MBE�%� R2 ENS

−10.6 0.13 −0.26 −15.1 0.83 0.73

−4.8 0.19 −0.12 −5.9 0.83 0.81

−1.2 0.06 −0.40 −7.9 0.84 0.77

4.7 0.09 −0.41 2.5 0.84 0.84

f ploation d

and Nand NE


error �MBE�, coefficient of determination �R2�, and Nash-Sutcliffe efficiency �ENS�. Results were very promising. Althoughcalibration targets were set at the monthly time scale, daily modeloutputs were also assessed against their observed counterparts.The R2 and ENS values computed at the daily time scale alsosatisfied the threshold values set for monthly results.

As an alternative data source, we appraised the use of radar-generated precipitation. The NEXRAD data obtained from theMARFC provided hourly precipitation estimates over approxi-mately 4�4 km2 grids. Measured precipitations at the two gaugestations were in close agreement with the NEXRAD estimatedprecipitations at the grids that contain these gauge stations.Twelve NEXRAD grid cells were determined to be within thePocono Creek watershed boundary. Close examination of the spa-tial distribution of average annual precipitationgenerated out of these cells disclosed that precipitation is notuniformly distributed within the watershed, with the highest pre-cipitation occurring near the highest elevation of the watershed.Average annual precipitation appears to have varied by as muchas 12% within the watershed.

The SWAT model was fed with the spatial-distributedNEXRAD precipitation data and recalibrated by reducing the av-erage curve numbers �CN� by a value of 1.5 that were obtainedfrom model calibration when gauge precipitation data was used asinput. The SF and SR hydrographs were in close agreement withmeasured hydrographs. Efficiency statistics MBE, R2, and ENS

were close to ones obtained with gauge-driven simulations.Hence, when model outputs at the basin outlet are considered, itwas concluded that the spatial representation of precipitation didnot contribute to model improvement in the model calibrationperiod. The true merits of using the spatially distributedNEXRAD precipitation data can be better understood if simula-tions are performed at the subwatershed scale. However, this re-quires flow measurements inside the watershed. Unfortunately,such flow measurements are not available for the Pocono Creekwatershed at this moment for further investigation. The modelwas then run for an additional 11 month period for validationpurposes, both with gauge and NEXRAD data. NEXRAD gener-ated smaller MBE. The values of R2 and ENS were higher at themonthly time scale with NEXRAD, whereas at the daily timescale gauge-driven simulations resulted in higher R2 and ENS. Ex-clusion of the extreme event in April 2005 significantly reducedthe mass balance error. It also improved R2 and ENS at themonthly scale, but yielded mixed results at the daily time scale.

This study showed that spatially distributed precipitation dataobtained through radar reflectivity measurements provide a viablealternative to rain gauge measurements. However, estimation ofprecipitation with the help of radar still needs improvement. Atpresent, data from rain gauges will continue to be relied upon tocorrect NEXRAD hourly digital precipitation for mean field bias.Future refinement of this technology may provide a cost-effectivealternative source of precipitation data and may eventually elimi-nate the need for the much more costly rain gauges, at least forlarge-scale watershed applications.

Acknowledgments

The U.S. Environmental Protection Agency, through its Office ofResearch and Development, funded the research described herethrough in-house efforts and in part by an appointment to thePostgraduate Research Program at the National Risk Management
Research Laboratory, administered by the Oak Ridge Institute for

Science and Education through an interagency agreement be-tween the U.S. Department of Energy and the U.S. EnvironmentalProtection Agency. It has not been subjected to Agency reviewand therefore does not necessarily reflect the views of the Agency,and no official endorsement should be inferred. We deeply appre-ciate David Kitzmiller from the National Weather Service �NWS�for all his help with the NEXRAD data and Karen Reavy of theDelaware River Basin Commission for her assistance with geo-graphic information system �GIS� data. We are also thankful toDr. Yu Zhang for his support in PERL scripting.

References

Arnold, J. G., and Allen, P. M. �1999�. “Automated methods for esti-mating base flow and ground water recharge from streamflowrecords.” J. Am. Water Resour. Assoc., 35�2�, 411–424.

Arnold, J. G., Allen, P. M., Muttiah, R., and Bernhardt, G. �1995�.“Automated base flow separation and recession analysis techniques.”Ground Water, 33�6�, 1010–1018.

Bedient, P. B., Anthony, H., Benavides, J. A., and Vieux, B. E. �2003�.“Radar-based flood warning system applied to Tropical Storm Alli-son.” J. Hydrol. Eng., 8�6�, 308–318.

Bedient, P. B., Hoblit, B. C., Gladwell, D. C., and Vieux, B. E. �2000�.“NEXRAD radar for flood prediction in Houston.” J. Hydrol. Eng.,5�3�, 269–277.

Carpenter, T. M., Georgakakos, K. P., and Sperfslagea, J. A. �2001�.“On the parametric and NEXRAD-radar sensitivities of a distributedhydrologic model suitable for operational use.” J. Hydrol., 253,169–193.

Di Luzio, M., and Arnold, J. G. �2004�. “Formulation of a hybrid cali-bration approach for a physically based distributed model withNEXRAD data input.” J. Hydrol., 298, 136–154.

Di Luzio, M., Srinivasan, R., Arnold, J. G., and Neitsch, S. L. �2002�.“ArcView interface for SWAT2000, user’s guide.” TWRI Rep.TR-193, Texas Water Resources Institute, College Station, Tex.

Eckhardt, K., and Arnold, J. G. �2001�. “Automated calibration of a dis-tributed catchment model.” J. Hydrol., 251, 103–109.

Fohrer, N., Möller, D., and Steiner, N. �2002�. “An interdisciplinarymodeling approach to evaluate the effects of land use change.”Phys. Chem. Earth, Part B, 27, 655–662.

Green, W. H., and Ampt, G. A. �1911�. “Studies on soil physics. 1: Theflow of air and water through soils.” J. Agric. Sci., 4, 11–24.

Harrison, D. L., Driscoll, S. J., and Kitchen, M. �2000�. “Improving pre-cipitation estimates from weather radar using quality control and cor-rection techniques.” Meteorol. Appl., 6, 135–144.

Jayakrishnan, R., Srinivasan, R., Santhi, C., and Arnold, J. G. �2005�.“Advances in the application of the SWAT model for water resourcesmanagement.” Hydrolog. Process., 19, 749–762.

Mukhopadhyay, B., Cornelius, J., and Zehner, W. �2003�. “Application ofkinematic wave theory for predicting flash hazards on coupled alluvialfan—Piedmont plain landforms.” Hydrolog. Process., 17, 839–868.

Muleta, M. K., and Nicklow, J. W. �2005�. “Sensitivity and uncertaintyanalysis coupled with automatic calibration for a distributed water-shed model.” J. Hydrol., 306, 127–145.

Nash, J. E., and Sutcliffe, J. V. �1970�. “River flow forecasting throughconceptual models. Part 1: A discussion of principles.” J. Hydrol., 10,282–290.

Neary, V. S., Habib, E., and Fleming, M. �2004�. “Hydrologic modelingwith NEXRAD precipitation in middle Tennessee.” J. Hydrol. Eng.,9�5�, 339–349.

Neitsch, S. L., Arnold, J. G., Kiniry, J. R., Williams, J. R., and King,K. W. �2002a�. “Soil and water assessment tool theoretical documen-tation, version 2000.” TWRI Rep. TR-191, Texas Water ResourcesInstitute, College Station, Tex.

Neitsch, S. L., Arnold, J. G., Kiniry, J. R., Williams, J. R., and King, K.
W. �2002b�. “Soil and water assessment tool user’s manual, version
ECEMBER 2006

2000.” TWRI Rep. TR-192, Texas Water Resources Institute, CollegeStation, Tex.

Pocono Creek Pilot Project. �2001�. “Geology, geomorphology,geohydrology, and surface water hydrology of the Pocono CreekBasin.” Draft Technical Rep., �http://www.state.ni.us/drbc/pocono_geology.PDF� �Jan. 30, 2006�.

Rallison, R. E., and Miller, N. �1981�. “Past, present, and future SCSrunoff procedure.” Rainfall runoff relationship, V. P. Singh, ed., WaterResources, Littleton, Colo., 353–364.

Reed, S., et al. �2004�. “Overall distributed model inter-comparisonproject results.” J. Hydrol., 298, 27–60.

Reed, S. M., and Maidment, D. R. �1999�. “Coordinate transformationsfor using NEXRAD data in GIS-based hydrologic modeling.” J. Hy-drol. Eng., 4�2�, 174–182.

Santhi, C., Arnold, J. G., Williams, J. R., Dugas, W. A., Srinivasan, R.,and Hauck, L. M. �2001�. “Validation of the SWAT model on a largeriver basin with point and nonpoint sources.” J. Am. Water Resour.Assoc., 37�5�, 1169–1188.

Smith, M. B., et al. �2004�. “The distributed model intercomparisonproject �DMIP�: Motivation and experiment design.” J. Hydrol., 298,
4–26.
JOURNAL OF HYDROL

Sophocleous, M., and Perkins, S. M. �2000�. “Methodology and applica-tion of combined watershed and ground-water models in Kansas.”J. Hydrol., 236, 185–201.

Sorooshian, S., Gupta, V. K., and Fulton, J. L. �1983�. “Evaluation ofmaximum likelihood parameter estimation techniques for conceptualrainfall-runoff models: Influence of calibration data variability andlength on model credibility.” Water Resour. Res., 19�1�, 251–259.

Tripathi, M. P., Panda, R. K., Raghuwanshi, N. S., and Singh, R. �2004�.“Hydrological modeling of a small watershed using generated rainfallin the soil and water assessment tool model.” Hydrolog. Process., 18,1811–1821.

U.S. Department of Agriculture �USDA� Soil Conservation Service.�1972�. “Section 4: Hydrology.” National engineering handbook,Washington, D.C.

Williams, J. R. �1969�. “Flood routing with variable travel time or vari-able storage coefficients.” Trans. ASAE, 12�1�, 100–103.

Xie, H., Zhou, X., Vivoni, E. R., Hendricks, J. M. H., and Small, E. E.�2005�. “GIS-based NEXRAD Stage III precipitation database: Auto-mated approaches for data processing and visualization.” Comput.
Geosci., 31, 65–76.

Hydrologic Modeling of an Eastern Pennsylvania …kalinla/papers/JHE2006.pdfHydrologic Modeling of...

Documents

Transcript of Hydrologic Modeling of an Eastern Pennsylvania …kalinla/papers/JHE2006.pdfHydrologic Modeling of...