RESEARCH ARTICLE10.1002/2014WR015589

Quantifying storage changes in regional Great Lakeswatersheds using a coupled subsurface-land surface processmodel and GRACE, MODIS productsJie Niu1,2, Chaopeng Shen3, Shu-Guang Li1, and Mantha S. Phanikumar1

1Department of Civil and Environmental Engineering, Michigan State University, East Lansing, Michigan, USA, 2EarthSciences Division, Lawrence Berkeley National Laboratory, Berkeley, California, USA, 3Department of Civil andEnvironmental Engineering, Pennsylvania State University, University Park, Pennsylvania, USA

Abstract As a direct measure of watershed resilience, watershed storage is important for understandingclimate change impacts on water resources. In this paper we quantify water budget components and stor-age changes for two of the largest watersheds in the State of Michigan, USA (the Grand River and the Sagi-naw Bay watersheds) using remotely sensed data and a process-based hydrologic model (PAWS) thatincludes detailed representations of subsurface and land surface processes. Model performance is evaluatedusing ground-based observations (streamflows, groundwater heads, soil moisture, and soil temperature) aswell as satellite-based estimates of evapotranspiration (Moderate-resolution Imaging Spectroradiometer,MODIS) and watershed storage changes (Gravity Recovery and Climate Experiment, GRACE). We use themodel to compute annual-average fluxes due to evapotranspiration, surface runoff, recharge and ground-water contribution to streams and analyze the impacts of land use and land cover (LULC) and soil types onannual hydrologic budgets using correlation analysis. Watershed storage changes based on GRACE dataand model results showed similar patterns. Storage was dominated by subsurface components and showedan increasing trend over the past decade. This work provides new estimates of water budgets and storagechanges in Great Lakes watersheds and the results are expected to aid in the analysis and interpretation ofthe current trend of declining lake levels, in understanding projected future impacts of climate change aswell as in identifying appropriate climate adaptation strategies.

1. Introduction

Storing water is one of the important hydrologic functions of a watershed [Black, 1997; Sivapalan, 2006].Watershed storage is a key variable used in watershed classification [Wagener et al., 2007] and is intimatelytied to surface and subsurface stormflows and runoff ratios in a watershed [Sawicz et al., 2011; Sayama et al.,2011]. Storage is also a key variable that controls the coevolution of hydrologic and biogeochemical statesof a watershed. Recent research highlights the link between watershed storage and nutrient levels [Seil-heimer et al., 2013] and between nutrient (specifically, nitrogen) levels, net primary production and hydro-logic fluxes such as evapotranspiration (ET) [Dickinson et al., 2002; Shen et al., 2013]. Understanding changesin watershed storage is important for climate change studies as storage is directly related to the resilienceof a watershed [Tague et al., 2008]; however, quantifying storage is difficult since many watershed modelsdo not explicitly simulate subsurface processes and thus do not adequately describe the past storage trajec-tory and the storage-discharge relations due to complex subsurface heterogeneity [Beven, 2006; Sayamaet al., 2011]. Recently there has been a surge of interest in the application of physics-based watershed mod-els that explicitly describe surface (e.g., St. Venant equations for channel flow) and subsurface (e.g., Richardsand/or Darcy equation-based) processes (e.g., see the descriptions of seven hydrologic models in Maxwellet al. [2014]). The application of physics-based watershed models to large river basins is a topic of increasinginterest. Recently gravity-based satellite measurements have been used to estimate water budgets and totalwater storage anomaly (TWSA) for continents [Wang et al., 2014] and large river basins such as the Missis-sippi [Rodell et al., 2007] and the Amazon [Pokhrel et al., 2013]. Long et al. [2014] and Rodell et al. [2004]describe uncertainties in water budget estimates of TWSA. High precision gravimeters have been used tostudy TWSA dynamics in response to climate variability and extremes or storage – baseflow relationships atdifferent catchment scales [Creutzfeldt et al., 2012; Creutzfeldt et al., 2013]. Combining gravity-based

Key Points:� Quantified water budgets and

storage changes in two Great Lakeswatersheds� Storage increased in both watersheds

over a 10 year period (2002–2012)� Application of a process-based

watershed model to regionalwatersheds

Correspondence to:J. Niu,jniu@lbl.gov

Citation:Niu, J., C. Shen, S.-G. Li, and M. S.Phanikumar (2014), Quantifyingstorage changes in regional GreatLakes watersheds using a coupledsubsurface-land surface process modeland GRACE, MODIS products, WaterResour. Res., 50, 7359–7377,doi:10.1002/2014WR015589.

Received 15 MAR 2014

Accepted 22 AUG 2014

Accepted article online 28 AUG 2014

Published online 19 SEP 2014

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7359

Water Resources Research

PUBLICATIONS

measurements with MODIS products for evapotranspiration and leaf area index, passive distributed temper-ature sensing [Steele-Dunne et al., 2010] and ground-based measurements for other key hydrologic variables(e.g., soil moisture, including the method based on cosmic-ray neutron probes [Zreda et al., 2008], stream-flows, groundwater heads and temperature) opens up the possibility to test physics-based watershed mod-els for their ability to describe major hydrologic fluxes and changes in watershed storage in regionalwatersheds; however, there are no such studies in the Great Lakes region to the best of our knowledge.

The Great Lakes account for 18% of the World’s and 90% of the United States’ freshwater supply [Sea GrantCollege Program, 1985]. The Great Lakes are an important source of fresh water, recreational resource andtransportation corridor for the Midwestern United States and Canada [Cherkauer and Sinha, 2010]. Estimates ofrunoff (related to watershed storage) and other components of the hydrologic cycle are important to under-stand and quantify lake level changes. According to previously reported research annual average and high tem-peratures in summer are on the rise, and the frequency of extreme rain events has been increasing [Christensenet al., 2007; Kling et al., 2003; Trapp et al., 2007; Wuebbles and Hayhoe, 2004]. It is important to understand howkey components of the hydrologic cycle are changing in response to climate change, relative to current condi-tions. However, except for a few major components of the water cycle, considerable variability exists in esti-mates of current water budgets due to differences in model conceptualizations of hydrologic processes (e.g.,lumped parameter versus process-based models). Regional groundwater flow is a major component of thewater budget globally and also in the Great Lakes region. However, not all watershed models explicitly describegroundwater processes; therefore, there is a need to quantify water budgets using models that include subsur-face processes. While a number of climate change-related projects are underway, there is a need to quantifycurrent water budgets, comparing estimates from different types of hydrologic models in order to assess andcompare inter-model variability in water budget estimates to projected future changes due to climate change.

The hydrologic water budget represents the balance between precipitation, evapotranspiration, surface run-off, and total water storage change, and includes all water pathways at or near the land surface, includingstreamflow, infiltration, recharge and exfiltration. There are several ways to estimate hydrologic budgets, e.g.,using field observations and data for different hydrologic components, or via real-time satellite products [Gaoet al., 2010]. Among these water budget components, streamflow can be observed by stream flow gauges.With dense weather station network and satellite-based estimates, precipitation can be quantified. The chal-lenge remains for estimating other spatially varying components of the hydrologic budget at large scales. Forexample, information on how the largest storage components within the watershed are changing has impor-tant implications for water resources management. In addition, due to the spatial heterogeneity and samplingerrors, the existing observational data alone may not be sufficient to provide accurate water budget informa-tion for research interests [Pan and Wood, 2006; Roads et al., 2003]. Spatially distributed hydrologic fluxes frommodel reanalysis have become more popular, e.g., NLDAS [Mitchell et al., 2004] and NLDAS-2 [Xia et al., 2012a,2012b]. Well-tested, process-based hydrologic models, which are based on conservation principles of mass,momentum and energy, may be well suited for providing detailed estimates of water budgets and provideinformation on how storage components are changing within the watershed.

Recently, process-based models became an important tool for studying the hydrologic water budget [Ber-bery et al., 2003; Luo et al., 2005; Roads et al., 2003; Shen et al., 2013; Shen and Phanikumar, 2010; Su and Let-tenmaier, 2009]. The PAWS (Process-based Adaptive Watershed Simulator) model used in this paper uses acoupling method to link surface and subsurface processes efficiently in order to facilitate the application ofprocess-based models to large watersheds. The model has been tested extensively using available analyticalsolutions, laboratory data [Shen and Phanikumar, 2010; Maxwell et al., 2014], streamflows, field and satellite-based observations for medium and large watersheds [Shen et al., 2013; Shen and Phanikumar, 2010; Shenet al., 2014; Riley and Shen, 2014].

The aim of this paper is to quantify storage changes and water budgets in two regional Great Lakes water-sheds using a physics-based watershed model and independent (satellite and ground-based) measure-ments of key hydrologic fluxes. Our working hypothesis is that if the model is able to describe observationsfor different domains (groundwater heads, soil moisture and stream flows) and major fluxes (e.g., ET), all ofwhich contribute to changes in storage, then model-based estimates of water storage changes can beexpected to be reliable. Specifically, we seek to answer questions such as: How is storage changing inregional Great Lakes watersheds? Can process-based hydrologic models describe the observed changes instorage? What hydrologic components are mainly responsible for the observed storage changes?

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7360

The paper is organizedas follows. After brieflydescribing the PAWSmodel and various datasources, we examinewater budgets and stor-age components for twoof the largest water-sheds in the State ofMichigan (the GrandRiver and the SaginawBay watersheds) over 12years (1995–2007). Simu-lated streamflows,groundwater heads, soilmoisture, soil tempera-ture, and ET are com-

pared to the observed data to demonstrate model performance. Simulated storage changes are thencompared with gravity-based satellite measurements of storage change (GRACE data reported as waterthickness anomalies). A statistical linear correlation analysis is also performed to examine the relation of thehydrologic budget components to the physical characteristics of the watersheds including land use andland cover, and soils type. A trend analysis of water thickness anomaly data is then reported and an expla-nation for the observed trends is provided.

2. Methodology

2.1. The PAWS ModelThe major hydrologic processes simulated by PAWS [Shen and Phanikumar, 2010] are depicted in Figure 1 andthe solution schemes are summarized in Table 1. The discretized differential equations for overland, soil waterand groundwater flows are solved using structured grids. The river network, which is extracted from NationalHydrography Dataset (NHD), is explicitly modeled [Shen et al., 2014]. A flexible strategy that computes theexchange fluxes between land grid and river cells allows river and land domains to be discretized independently.Two subdomains—the ponding water layer and the flow domain are used to separate water on the ground.Water can only move between cells (including river cells) in the flow domain, while infiltration and evaporationhappen in the ponding layer. Runoff from ponding layer to the flow domain is estimated based on Manning’sequation and is solved simultaneously with the soil-water flow. A layer of snow, which is quantified by the snow-cover fraction and Snow Water Equivalent (SWE), may also exist on the ground depending on the climatic condi-tion. A mosaic approach is used to model the subgrid heterogeneity within each land cell. Several generic plantfunctional types (PFT) [Thornton and Zimmermann, 2007] represent model classes and are used to reclassify theland use data. Rainfall and snowmelt are joined and add to the ponding layer after canopy interception. In PAWS,for each land cell there is one soil column modeled. The three-dimensional unsaturated vadose zone is concep-tualized as an array of one dimensional soil columns connected to the unconfined aquifer at the bottom. PAWSmakes the simplifying assumption that water moves only vertically in the unsaturated zone [Shen and Phaniku-mar, 2010, Shen et al., 2013]. Water and energy can be exchanged between the soil column and the unconfinedaquifer. Through a layer of aquitard, water can percolate further down to deeper aquifers and exchange with riverwater via riverbed materials. Water can also be directly extracted from the aquifer layers by pumping activities.For the confined aquifer which is represented as a mixture of different fractured bedrock formations as shown inFigure 1, the exchange exists primarily with the unconfined layer above. Another assumption is that the lateralflow is uniformly distributed along the saturated thickness of the unconfined and confined aquifers. The details ofthe model conceptualization and coupling between different layers can be found in Shen and Phanikumar [2010].

The coupling of CLM (version 4) to PAWS [Shen et al., 2013] provides detailed, process-based vegetationdynamics [Thornton and Zimmermann, 2007], energy cycle and carbon/nutrient cycling. The soil hydrology andriver routing routines in PAWS replace the corresponding procedures in CLM. Except due to snowmelt andfreeze-thaw phase change, soil moisture states are unchanged in CLM. The resulting ET fluxes, soil water/icecontent are returned to PAWS for the calculations of flow processes after completing CLM calculations.

Table 1. Processes and Governing Equations of the Hydrological Model PAWS

Process Governing Equations

Snowfall accumulation and melting SNICAR model [Toon et al., 1989] in CLM 4.0Canopy interception Bucket model in CLM, storage capacity related

to total Leaf Area Index (tLAI)Depression storage Specific depthRunoff Manning’s formula 1 Kinematic wave formulation 1

Coupled to Richards equationInfiltration/exfiltration Processes coupled to Richards equationOverland flow 2-D Diffusive wave equationOverland/channel exchange Weir formulationChannel network Dynamic wave or diffusive waveEvapotranspiration Penman Monteith 1 Root extractionSoil moisture Richards equationLateral groundwater flow Quasi-3-DRecharge/discharge

(Vadose zone/groundwater interaction)Noniterative coupling inside Richards equation

Stream/groundwater interaction Conductance/leakageVegetation processes CLM ver. 4.0

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7361

2.2. Description of SitesThe Grand River (GR) watershed is the second largest watershed in Michigan (Figure 2). It is located in themiddle of Michigan’s Lower Peninsula. The GR is the longest river in Michigan, which runs 420 km acrossseveral counties before emptying into Lake Michigan. The watershed drains an area of 14,431 km2 including18 counties and 158 townships. The basin mainly contains flat terrain and many swamps and lakes. The Sag-inaw Bay (SB) watershed is the largest drainage basin in Michigan (Figure 2), draining approximately 15% ofthe total land area in the state. The watershed includes a part of 22 counties and contains the largest con-tiguous freshwater coastal wetland system in the United States. It is located in the east central portion ofMichigan’s Lower Peninsula. The Saginaw River is a 35 km river formed by the confluence of tributaries atsoutheast of Saginaw. It flows northward into the Saginaw Bay of Lake Huron just northeast of Bay City. Thearea of the SB watershed is 22,556 km2 and over half of the land use in the region is agricultural. The landuse and land cover map for both watersheds is shown in Figure 3.

For both watersheds, we use a 1 km2 cell size for spatial (horizontal) discretization, which results in a gridsize of 1953170 for GR and 2013210 for SB. We use 20 vertical layers to discretize the vadose zone and tosolve the Richards equation and two layers for the fully saturated groundwater domain – the top layer inthe groundwater domain represents the unconfined aquifer while the bottom layer represents the confinedaquifer. The resolution for the river cells is 1 km. For temporal discretization, the base time step is 1 h forgroundwater and land surface processes, 10 min for the overland flow and channel flow, and adaptive timesteps ranging from 10 min to 1 h are applied for the unsaturated zone.

2.3. Data SourcesThe 30 m resolution National Elevation Dataset (NED) from U.S. Geological Survey (USGS) (http://ned.usgs.gov) was used for the Digital Elevation Model (DEM) input for both watersheds to calculate the slope andfor surface runoff routing computations. The daily based meteorological elements, precipitation, snowfall,maximum temperature, minimum temperature, wind speed, and relative humidity acquired from NationalClimatic Data Center (NCDC, stations listed in Table 2) of the National Oceanic and Atmospheric Administra-tion (NOAA) or Michigan Automated Weather Station Network (MAWN) [Enviro-weather, 2010] were used asweather data inputs for the model. In our study regions there are 14 weather stations available. The nearestneighbor interpolation scheme was used for spatial interpolation of data. For land use and land cover(LULC) data we used the Integrated Forest Monitoring Assessment and Prescription (IFMAP) data set thathas a 30 m resolution provided by the Michigan Department of Natural Resources [MDNR, 2010]. Using theland use classification percentile values reported in the literature [e.g., Jia et al., 2001; Lu and Weng, 2006],we transformed the land use classification into several different model classes, called Plant Functional Types

Run

off

Exchange

ED

Rain

P

Snow

Stream

EO

T

R

Eg

Groundwater

Snowmelt

Inf

Water table

Vadose Zone Impervious Cover

TSublimation

Exfiltration

“Red Beds”Grand River and Saginaw Formations,Parma sandstone, and Bayport Limestone

Michigan Formation

Figure 1. Definition sketch illustrating the bedrock geology and some of the key processes modeled in PAWS. T: transpiration, P: precipitation, EO: evaporation from overland flow/stream, Eg: evaporation from bare soil, ED: evaporation from depression storage, Inf: infiltration, R: recharge.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7362

(PFT). Since land use maps normally have finer resolution than model grid, we use three PFTs in a modelcell. The eight main MDNR land use class coverage percentiles for both watersheds are listed in Table 3.

Soil Survey Geographic (SSURGO) database from U.S. Department of Agriculture, Natural Resources Conser-vation Service (NRCS) was processed to provide soil water retention and unsaturated conductivity informa-tion, via the pedotransfer functions provided in Rosetta [Schaap et al., 2001]. For the channel flow network,we used National Hydrography Dataset (NHD) from USGS. The streamflow and groundwater head data forcomparison were observed by a number of USGS gauging stations (listed in Table 4 and shown in Figure 2)within the two watersheds. There are a few stations with soil moisture and soil temperature data availablefrom the Enviro-Weather Network (data available at http://www.enviroweather.msu.edu/homeMap.php).We compared soil moisture and soil temperature at 10 cm depth for Enviro-Weather station located at Beld-ing (Kropf farms, latitude 43.1125, longitude 285.3122, elevation 257 m) in the GR watershed.

The evapotranspiration data were downloaded from the MODIS Global Evapotranspiration Project (MOD16),which is a part of NASA/EOS project to estimate global terrestrial evapotranspiration from earth land surfaceby using satellite remote sensing data (http://www.ntsg.umt.edu/project/mod16). The MOD16 global ET datasets are regular 1 km2 land surface ET data sets for the 109.03 million km2 global vegetated land areas at 8day, monthly, and annual intervals. The data set covers the time period 2000–2012. The MOD16 ET data setsare estimated using an improved ET algorithm [Mu et al., 2011] over the previous one [Mu et al., 2007]. The ETalgorithm is based on the Penman-Monteith equation. GRACE equivalent water thickness product was down-loaded as monthly mass grids from the Jet Propulsion Laboratory (JPL) (http://grace.jpl.nasa.gov/data/

83°W84°W85°W86°W

45°N

44°N

43°N

42°N

0 20 40 km

GR watershedSB watershedCountiesNHD streamsUSGS Gauges

USGS Groundwater StationsNCDC Weather StationsMAWN Stations

LakeHuron

Sagina

w Bay

LakeMichigan

479

173

Elevation (m):

Figure 2. Map of Grand River (GR) and Saginaw Bay (SB) watersheds. Elevation is shown as the color gradient. National Hydrography Dataset(NHD) streams, U.S. Geological Survey (USGS) gauges, groundwater stations and National Climatic Data Center (NCDC) weather stations are shown.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7363

gracemonthlymassgridsland/). The data are based on the Release 5.0 (RL05) spherical harmonics from JPL,CSR (the Center for Space Research at the University of Texas) and GFZ (GeoForschungsZentrum in Potsdam)and a glacial isostatic adjustment correction was applied to the data based on the model of [Geruo et al.,2013]. The uncertainties in spherical harmonics from the latest release (RL05) of CSR have been reduced byabout 40% compared with the previous CSR release (RL04) [Long et al., 2013]. The distributed GRACE productuses a destriping filter and a 300 km wide Gaussian filter as well to minimize North-South stripes in themonthly maps. This data product provides time series for global river basins with seasonal and trend esti-mates over land area with a spatial sampling of 18318 in latitude and longitude (approximately 111 km at theEarth’s equator). As described earlier, a majority of the previous applications of GRACE data focused on conti-nents and large river basins such as the Amazon. This is consistent with the large footprint of GRACE, whichhas a resolution�200,000 km2. Although the GR and SB watersheds are considered large watersheds, theyare relatively small compared to the GRACE footprint (e.g., the GR watershed has an area of approximately14,000 km2). For relatively small watersheds the uncertainties in total water storage (TWS) anomalies can be

expected to be large. To address the scale issuewe first extracted GRACE data for a larger rectan-gular area that encloses our watershed and thenused area-weighted averaging to obtain the TWSanomaly for the watershed. For the GR water-shed, we used 1 km2 grid cells and a computa-tional grid with 195 3 170 cells; therefore thearea of the rectangular bounding box is33,150 km2 with 7 sampled and postprocessedGRACE grid cells covering the area. After com-puting the area-weighted average anomalies,we compute GRACE error estimates using thealgorithm reported by JPL (http://gracetellus.jpl.nasa.gov/data/gracemonthlymassgridsland/) [Landerer and Swenson, 2012; Swenson andWahr, 2006]. Since surface mass variations at

Figure 3. Land Use and Land Cover map for the Grand River (GR) and the Saginaw Bay (SB) watersheds.

Table 2. National Climatic Data Center (NCDC) Weather Stations

StationsNumber Stations Name Latitude Longitude

200146 Alma 43.38 284.67201818 Corunna 2NE 43.00 284.07202395 East Lansing 4 S 42.67 284.48202437 Eaton Rapids 42.52 284.65203333 Grand Rapids 42.88 285.52203429 Greenville 2 NNE 43.20 285.25203664 Hastings 42.65 285.28204078 Ionia 2 SSW 42.95 285.08204150 Jackson Reynolds FLD 42.27 284.47204320 Kent City 2 SW 43.20 285.77204641 Lansing Capital City AP 42.78 284.58204944 Lowell 42.93 285.33205712 Muskegon CO AP 43.17 286.23209006 Williamston 3NE 42.72 284.25

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7364

small spatial scales tend to be attenu-ated due to the low-pass filtering ofGRACE spherical harmonics, the scalingfactor should be applied to the originalgridded data to restore signal losses. Thescaling grid is a set of scaling coeffi-cients, one for each 1� bin of the landgrids, which can be accessed from thelink ftp://podaac-ftp.jpl.nasa.gov/allData/tellus/L3/land_mass/RL05/ascii/. To com-pute error estimates for the scaled val-

ues, two additional grids are provided in the same link, the measurement error and the leakage error thatis the residual error after filtering and rescaling. The error data of all 7 GRACE grid cells were used to cal-culate the variance and standard deviation which were used for further estimation of the confidenceintervals of the GRACE anomalies time series data. This approach, which was used by [Ferrant et al., 2014]to apply GRACE data to a small watershed (983 km2), allows us to take the large uncertainties into accountwhile evaluating model results against GRACE data.

2.4. Simulation DetailsThere are uncertainties associated with parameters such as riverbed conductance and the van Genuchtenparameters for the soils. There are 8 calibration parameters that are described in Shen et al. [2013]. Parame-ters that are spaially variable can be adjusted using ‘‘global multipliers’’ within the framework of a lineartransformation (that is, a positive multiplicative constant that multiplies the original variable plus an addi-tive constant, which can be zero). For example, a global multiplier applied to the hydraulic conductivity fielduniformly increases or decreases conductivity while honoring the geology (that is, retaining the natural vari-ability and heterogeneity in the original data). The parameter estimation, based on the differential evolutionalgorithm [Price et al., 2005] was done using the High Performance Computing Center (HPCC) resources atMichigan State University by minimizing the following objective function f ðxÞ, which represents model errorrelative to observed streamflows and groundwater heads:

f ðxÞ5XN

i51

wcifci1XM

k51

wgk fgk

XN

i51

wci1XM

k51

wgk51

fci512NASHi1RNASHi

2

� �

fgk512NASHk

NASH512

Xn

j51

Oj2Pj� �2

Xn

j51

Oj2Oj� �2

RNASH512

Xn

j51

ffiffiffiffiffiOj

p2

ffiffiffiffiPj

p� �2

Xn

j51

ffiffiffiffiffiOj

p2

ffiffiffiffiffiOj

p� �2

(1)

where x denotes the vector of parameters, fci and fgk denote the objective functions for individual streamgauging stations (c denotes channel, index i denotes the i-th stream gauge, i51,2. . .N) or stations for whichgroundwater head data are available (g denotes groundwater, index k denotes the k-th well, k, k51,2,. . .M),and wci and wgk are the weights given to a station. How to assign weights to different stream gauging sta-tions and groundwater wells is an important research question. In the past any additional information froma station is considered useful and therefore worthy of including in the objective function f ðxÞ. Recently, Fry

Table 3. Land Use and Land Cover (LULC) Coverage Percentages in theGrand River and Saginaw Bay Watersheds

GR (%) SB (%)

Needle leaf evergreen temperate tree 2.44 3.71Broadleaf deciduous temperate tree 19.64 25.48Broadleaf deciduous temperate shrub 0.63 1.91C3 non-arctic grass 5.98 8.39Crops 57.91 47.84Impervious 5.30 3.73Water 1.29 1.25Wetland 6.81 7.69

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7365

et al. [2013] showed that infor-mation from some gauging sta-tions can reduce model skilland provided evidence to chal-lenge this view. The factors thatinfluence the relative impor-tance of a gauging station aremany and a systematic analysisof these factors for our largewatersheds is beyond thescope of the present paper. Inthe absence of such informa-tion we have assigned equal

weights to all stations such that the sum of the weights is equal to unity. The Nash-Sutcliffe efficiency metric(NASH) and the square-root-transformed NASH metric (RNASH) are also defined above where Oj and Pj

denote observations and simulations at a given time respectively (j 5 1,2,. . .n). As described in [Shen andPhanikumar, 2010], NASH is biased toward the peaks in streamflow data; therefore use of the RNASH metricis expected to give more emphasis on groundwater processes such as the baseflow contribution to streams.For the SB watershed, we did not adjust any parameters. Instead all model parameters and multipliers forthe GR watershed were used for the SB watershed to assess model performance and to test whether param-eters from a geographically similar watershed will produce acceptable results. The model performance wasevaluated using the NASH, the absolute percent bias (APB), and the root mean squared error (RMSE)defined as below.

APB5

Xn

j51

Oj2Pj� �

3100

Xn

j51

Oj� �

RMSE5

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1n

Oj2Pj� �2

r(2)

2.5. Storage and Water Budget CalculationsTo estimate storage changes based on water balance components, we use the following equation for theTWS, which represents the mass (directly related to volume via density) of water in all hydrologic domainsincluding the subsurface component. The total water storage anomaly (TWSA) is computed by subtractingthe mean value over the simulation period (TWS) from the instantaneous value (TWS):

TWS5GW1SW5ðVDZ1GWU1GWCÞ1ðOVN1CS1SWE1CFÞ

TWSA5TWS2TWS(3)

Here GW represents the water stored in the subsurface that includes the unconfined (GWU) and confined(GWC) aquifers and water in the vadose zone (VDZ) as well; OVN represents the amount of water from over-land flow components (but not channel flow) and all surface water components including the pondingdomain, low land and depression storages; CS denotes canopy storage; SWE is snow water equivalent andCF represents the amount of water from channel flow. In equation (3), all terms have the units of volume[L3] or length [L] if divided by the drainage area. TWSA is the anomaly used for comparison with GRACEdata [Wang et al., 2014; Pokhrel et al., 2013]. Because of the subsurface coupling scheme employed in PAWS,the water in the vadose zone and the unconfined aquifer are computed together using the variable w asshown by the following equation:

w5 VDZ1GWUð Þ5ðE

EB

h dz

GW5w1GWC

(4)

where h is the soil moisture content, EB is elevation of top of the bedrock and E is the groundwater surfaceelevation.

Table 4. U.S. Geological Survey (USGS) Gauging Stations

Gauge Number Gauge Name Latitude Longitude

GR 04111000 Grand River near Eaton Rapids 42.5347 284.623104112500 Grand River at East Lansing 42.7272 284.748104113000 Grand River at Lansing 42.7505 284.555304114000 Grand River at Portland 42.8564 284.912204114498 Looking Glass River near Eagle 42.8281 284.759404116000 Grand River at Ionia 42.9720 285.069204117500 Thornapple River near Hastings 42.6159 285.236404119000 Grand River at Grand Rapids 42.9645 285.6764

SB 04147500 Flint River near Otisville 43.1111 283.519404155000 Pine River at Midland 43.5645 284.369204157000 Saginaw River at Saginaw 43.4128 283.963004148500 Flint River near Flint 43.0389 283.7717

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7366

3. Results and Discussion

3.1. Stream Flow and Groundwater Head ComparisonsFigure 4 shows 12 year streamflow comparison at eight different USGS gauging stations in the GR water-shed. The NASH values for all comparisons ranged from 0.38 to 0.64. The same parameter set estimated forthe GR watershed simulation was applied for the SB watershed simulation since major watershed-specificinput data including soils, geology, climate and topography have already been used for the SB watershed.The NASH were still above 0.37 for four USGS gauges in the SB watershed shown in Figure 5. We note that,in all cases, the base flow is generally described very well. However, there are mismatches in streamflowsincluding the underestimation of streamflow peaks during snowmelt periods due to measurement uncer-tainty of snowfall depths at stations, and some deficiencies of the snow energy balance routine of currentCLM model as it does not account for the effect of snow cover fraction in heat exchange, which wouldresult in errors in snow melt timing and we have found this to be a major suspect for some under-estimation of spring peaks. For example, for USGS 4116000, 4119000, and 4114000, the peaks in spring2007 and generally after that are not well resolved. Careful examination revealed that coverage of precipita-tion records for two important weather stations (Williamston and Eaton Rapids) was intermittent with signif-icant loss of data after 2005. The missing records are filled by borrowing records from neighboring stations,which resulted in a drop in model performance. The lack of weather stations with valid/reliable data recordsduring the simulation period also caused the reduction of NASH for the stream flow comparison for the SBwatershed. The plots of simulated versus observed groundwater head from Wellogic data set for each com-putation grid cell are shown in Figure 6a for the GR and Figure 6b for the SB watersheds. The overall R2 val-ues are over 0.9 showing a spatially good match. The discrepancies came from the coarse observation of

0

20

40

60

80

100 ObservedSimulated

NASH = 0.38APB = 41.28%RMSE = 8.58USGS 4111000

0

20

40

60

80

100 NASH = 0.48APB = 47.48%RMSE = 5.37USGS 4112500

0

50

100

150

200

250NASH = 0.49APB = 40.37%RMSE = 16.01USGS 4113000

0

100

200

300 NASH = 0.56APB = 36.18%RMSE = 17.23USGS 4114000

010

2030

40

5060

NASH = 0.49APB = 42.89%RMSE = 3.32USGS 4114498

0100

200300

400

500600

NASH = 0.64APB = 33.63%RMSE = 32.50USGS 4116000

12/07/95 09/02/98 05/29/01 02/23/04 11/19/060

40

80

120

160NASH = 0.60APB = 36.04%RMSE = 6.00USGS 4117500

0100

300

500

700

900NASH = 0.61APB = 29.43%RMSE = 53.56USGS 4119000

Date (mm/dd/yy) Date (mm/dd/yy)

Dis

char

ge (m

3 /s)

12/07/95 09/02/98 05/29/01 02/23/04 11/19/06

Figure 4. Simulated and observed streamflow comparison for the Grand River (GR) watershed at different U.S. Geological Survey (USGS)gauges. USGS gauge station ID and model performance metrics are shown for each subplot.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7367

the Wellogic data and the fact that no parameters were estimated may have resulted in the deviation forlarge head values in the SB watershed. The availability of transient groundwater head observations fromUSGS gauging station is limited in our study area. Thus we only have one groundwater head comparison inthe GR watershed (Holt, Michigan) shown in Figure 7. The comparison (R2 5 0.72) represents a good matchconsidering the fact that transient head comparisons are sensitive to local agricultural, municipal and indus-trial pumping which are difficult to quantify accurately.

3.2. Soil Moisture and Soil Temperature ComparisonsThe soil moisture and soil temperature comparisons at 10 cm depth are presented in Figures 8a and 8b,respectively. It should be noted that our simulation results represent an average over the 1 km2 grid cell,whereas the observed data represent a point measurement as data were collected using a Campbell Sci-entific CS616 water content reflectometer (WCR) with a sampling volume of a few cubic centimeters. Thesimulated soil moisture is generally lower compared to observations and this is particularly true duringwarmer periods and during winter months. Several researchers found that the response of WCRs such asthe CS616 is sensitive to both temperature and soil type and that a soil-specific (or an isothermal) calibra-tion alone cannot capture the effects of temperature and therefore can introduce significant errors intothe measurement. For example, Saravanathiiban [2014] found that if a temperature correction is notapplied, the measured volumetric water content (VWC) is overestimated by 40 to 50% for temperaturesbelow 10�C during the winter months and that it is overestimated by about 10% during the summermonths when temperature exceeds 23�C. For their soil, Benson and Wang [2006] found that the VWC canvary by 0.1 over the environmental temperature range of 5–35�C if a correction is not applied. They alsonoted that the use of the manufacturer’s equation will also result in overestimation of the VWC. This tem-perature effect can be clearly seen in our comparisons since a correction was not applied. For example,around 23 February 2004, while the simulated results show a dip in soil moisture, the observed valuesremained quite high in Figure 8a, consistent with the observations of earlier researchers [e.g., Benson andWang, 2006; Saravanathiiban, 2014]. In view of these differences, the comparisons are consideredacceptable.

3.3. Evapotranspiration Comparisons With MODIS DataModel ET result was both temporally and spatially averaged through the whole computational domain andeach month during 8 years (2000 to 2007). MODIS data were extracted for the same watershed area and

12/07/95 09/02/98 05/29/01 02/23/04 11/19/06

0

20

40

60

80

100NASH = 0.67APB = 30.04%RMSE = 5.14USGS 4155500

0

200

400

600

800

1000

1200NASH = 0.46APB = 41.60%RMSE = 98.14USGS 4157000

0

50

100

150

200

250NASH = 0.37APB = 48.37%RMSE = 9.72USGS 4147500

0

50

100

150

200

250

300 NASH = 0.43APB = 49.59%RMSE = 16.72USGS 4148500

12/07/95 09/02/98 05/29/01 02/23/04 11/19/06

Date (mm/dd/yy) Date (mm/dd/yy)

Dis

char

ge (m

3 /s)

obssim

Figure 5. Simulated and observed streamflow comparison for the Saginaw Bay (SB) watershed at different U.S. Geological Survey (USGS)gauges. USGS gauge station ID and quality measurement values are shown for each subplot.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7368

time period. The comparisons for bothwatersheds are presented in Figures9a and 9b in the form of time series.The seasonal variation is clear: ETreaches its maximum value duringsummer and a minimum value duringwinter as expected. The NASH valuesof the 8 year comparison for the GRand SB watersheds shown in Figures9a and 9b are 0.71 and 0.82 respec-tively, which represent a good matchbetween the observed and the simu-lated, except for the deviation at thehigher values in July and August. Thepossible reason for the deviation isattributed to the different ET algo-rithms and water stress formulationsused between CLM and MODIS esti-mates. PAWS1CLM uses a resistanceapproach to describe ET based on thetwo-big leaf model [Dai et al., 2004],while the MODIS product is based onthe Penman-Monteith formulation[Mu et al., 2011; Mu et al., 2007].

A comparison (observed versus simu-lated) of the spatial maps of 5 yearaverage ET for the GR and SB water-sheds is shown in Figures 10 and 11.The mean annual average ET valuesare 568 and 520 mm/yr, which is com-parable to 579 and 617 mm/yr of theMODIS evapotranspiration data. Thehighest ET values (about >800 mm/yr)occur due to evaporation from openwater bodies, while some of the low-est ET values (below 300 mm/yr) arefound for urban areas (there is noMODIS data available for urban areas;thus these areas are blanked out inthe MODIS ET map) and the places

with low sand percentage in soil. While there are similarities in the observed and simulated spatial patternsin ET (e.g., areas with open water), we also note significant differences in the patterns. The most significantdifference is noted in the northern part of the Grand River watershed where the clay percent of soil isextremely high in some grid cells. Sandy soils readily bring water to the surface producing high ET valuesand the opposite is generally true of clayey soils (low ET); although low ET values are expected for thesegrid cells based on soil type, such low values are absent in the MODIS data sets. The reason for this differ-ence is most likely related to the different ET algorithms used in PAWS1CLM and to process MODIS data.

3.4. Budget Analysis3.4.1. Linear Correlation Analysis of Annual Average ResultsTo better understand the factors influencing the hydrologic water budgets and to understand the reasonsfor the spatial variability in key budget components such as ET, a linear correlation analysis was performedon different water budget components against the explanatory variables including LULC and soil types. Theresults are summarized in Table 5. In the table, q is the pair-wise linear correlation coefficient, and p is the

160 180 200 220 240 260 280 300 320 340 360 380160

180

200

220

240

260

280

300

320

340

360

380

Observed Wellogic GW head (m)

Sim

ulat

ed G

W h

ead

(m)

160 180 200 220 240 260 280 300 320160

180

200

220

240

260

280

300

320

R2 = 0.99APB = 0.56%RMSE = 1.90

R2 = 0.97APB = 1.32%RMSE = 4.71

(a)

(b)

Figure 6. Simulated groundwater head comparison with WELLOGIC data set forGR (a) and SB (b) watersheds.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7369

probability for testing the null hypothesis of no correlation against the alternative hypothesis that there is asignificant correlation. Statistical significance is assessed at a p value of 0.05. In the GR watershed, ET valuesare positively correlated to forested areas strongly (q50:27), but negatively related to impervious land use,which represents urban areas (q520:35). ET is also strongly correlated to wetlands (q50:26). In the SBwatershed, ET is also strongly and positively correlated to wetlands (q50:38), followed by forested areas(q50:35). The different ET correlation coefficients in the two watersheds are due to differences in the areas

09/02/98 05/29/01 02/23/04 11/19/06264.5

265

265.5

266

266.5

267

267.5

Holt, Michigan. GroundwaterR2 = 0.72, APB = 0.07%, RMSE = 0.23

Wat

er ta

ble

elev

atio

n (m

)

Date (mm/dd/yy)

obssim

Figure 7. Simulated and observed transient groundwater heads at Holt, Michigan (Grand River watershed).

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4(a) 10cm soil moisture

Soil

Moi

stur

e (m

3 /m3 )

SimulatedObserved

10/11/02 02/23/04 07/07/05 11/19/06 04/02/08−10

−5

0

5

10

15

20

25

30

35

Soil

Tem

pera

ture

(°C

)

(b) 10cm soil temperature

Time (MM/DD/YY)

Figure 8. Comparison of observed and simulated (a) soil moisture and (b) soil temperature at Belding, Michigan (Grand River watershed).

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7370

of different LULC in two watersheds; for example, there is a larger fraction of forested areas in the SB water-shed than in the GR watershed (in Table 3, needleleaf evergreen temperate tree is 2.44% in GR and 3.72% inSB, broadleaf deciduous temperate tree is 19.64% in GR and 25.48% in SB). We also found that ET is very sen-sitive to soil types. Larger percentages of sand (q50:55 in GR and q50:63 in SB) and lower percentages ofclay (q520:59 in GR and q520:69 in SB) in the soil result in higher ET although some earlier work [Istan-bulluoglu et al., 2012; Wang et al., 2009] found a negative dependence between mean annual ET and basinsand coverage in some of the major Nebraska Sand Hills.

The correlation analysis results not only support our understanding of ET dependence on soils and land usebut also explain the 5 year average ET spatial maps very well. Correlation analysis results for other hydro-logic components are also given in the same table. We also note that the two major runoff generationmechanisms (Dingman, 2002) in the two watersheds—infiltration excess (or Hortonian runoff) and satura-tion excess (or Dunne runoff) are strongly correlated to the urban and forested land uses respectively. Satu-ration excess is negatively correlated with forested land use while infiltration excess is positively correlatedwith urban land use. For example, in the GR watershed, the correlation coefficient is 20.23 between for-ested land use and saturation excess. In the SB watershed, the corresponding value is 20.40. Similarly, thecorrelation coefficients between urban land use and infiltration excess are 0.39 for the GR watershed and

0

20

40

60

80

100

120MODISPAWS

01/15/00 05/29/01 10/11/02 02/23/04 07/07/05 11/19/06 04/02/08

(a) Grand River Watershed NASH = 0.71, APB = 25.76%, RMSE = 15.03

Date (mm/dd/yy)

ET (m

m/m

onth

)

0

20

40

60

80

100

120(b) Saginaw Bay WatershedNASH = 0.82, APB = 21.62%, RMSE = 12.52

Figure 9. Monthly simulated and MODIS ET data comparison for the Grand River (GR) (a) and the Saginaw Bay (SB) (b) watersheds for theperiod 2000–2007.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7371

0.34 for the SB watershed respectively. These results partially support the view that forest vegetation main-tains the soil water storage minimizing runoff, and that water is withdrawn from the soil to support ET. Sur-face runoff is also strongly correlated to soil types, negatively with soil sand percent and positively with claypercent. Infiltration is very sensitive to the land use types other than different soil type percentile coverage(small q values for soil sand and soil clay in GR and SB in Table 5). It is negatively correlated to wetlands andforested areas (q520:44 and 20:49 for wetlands in GR and SB respectively, q520:37 for forest in bothwatersheds) but strongly positively correlated to crops (q50:56 and 0.52), and no correlation was foundwith urban areas (q50:06 and 0.05).

3.4.2. Storage and PartitionAnnual averages of different water budget components for the GR and SB watersheds are presented inTable 6, where SUBM is sublimation; Qoc is surface runoff or overland flow; Dperc is groundwaterrecharge and the other terms are described earlier. The components shown in Table 6 represent themajor hydrologic components and, as in the case of runoff, represent different processes (infiltrationexcess, saturation excess, melting snow and so on) contributing to the same component, therefore theyare not expected to add up to a 100%. In our study, the sum of SUBM, Qoc, Qgc, Dperc and ET is about10% to 20% larger than P, which is reasonable because there may be other inputs/outputs or storagemechanisms in the watershed. We can calculate the ratio of each hydrologic component based on the

Figure 10. Spatial map of mean evapotranspiration for the Grand River watershed for the 5 year period (2003–2007) based on (a) MODISdata and (b) the PAWS model.

Figure 11. Spatial map of mean evapotranspiration for the Saginaw Bay watershed for the 5 year period (2003–2007) based on (a) MODISdata and (b) the PAWS model.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7372

amount of precipitation; for example, the annual average ET percentages are 68.9% and 65.8% for GR andSB respectively. These ET percentages are comparable to the values reported for three watersheds in Illi-nois by Arnold and Allen [1996].

A time series plot of the TWSA from GRACE data and the PAWS model anomalies computed using equation(3) is shown in Figure 12 for the GR watershed. The GRACE error estimates, computed using the algorithmprovided on the JPL website are used to estimate the 95% confidence intervals for the data and these areshown in the same plot using the light blue shaded area. As expected, TWS increases in cold seasons, whenthere is no plant growth (small ET), and declines as we proceed to spring and summer. Figure 12 clearlyillustrates the seasonal energy cycle, together with effects of individual precipitation events. The differencesof amplitude variation between model and data shown in the figure are attributed to the coarse resolutionof GRACE data and smaller spatial scale of the watersheds in this study, and the smoothing processes usedin the GRACE data that are necessary for reducing high-frequency noise in both high-degree and high-order spherical harmonics. In spite of these differences both the model result and the GRACE data show asimilar pattern for watershed storage over the 10 year period considered in this work. Figure 12 alsoincludes monthly storage changes due to key hydrologic fluxes computed using the model and shown as astacked column plot. It is clear that storage changes due to subsurface water components dominate the

TWSA signal with overland flow (OVN), can-opy storage (CS), snow water equivalent(SWE) and channel flows (CF) making rela-tively small contributions to the overall TWSA.In Figure 12, GW denotes the contributionfrom the entire subsurface domain. Althoughnot shown in Figure 12, the contribution fromw (vadose zone plus unconfined aquifer,defined in equation 4) dominates the contri-bution from GW (the subsurface domain) andthe overall TWSA signal with water in the con-fined aquifer contributing relatively little tothe overall change in storage.

Trend analysis results are sensitive to theperiod chosen for analysis and results based

Table 5. Spearman Correlation Coefficientsa

Watershed

Land Use/Land Cover Soils

Forest* Grass* Crops Urban Wetland Soils Sand Soils Clay

ET GR q 0.27 20.03 20.21 20.35 0.26 0.55 20.59p 0.00 0.00 0.00 0.00 0.00 0.00 0.00

SB q 0.35 20.11 20.28 20.34 0.38 0.63 20.69p 0.00 0.00 0.00 0.00 0.00 0.00 0.00

Inf GR q 20.37 0.03 0.56 0.06 20.44 20.04 0.09p 0.00 0.00 0.00 0.00 0.00 0.00 0.00

SB q 20.37 0.13 0.52 0.05 20.49 20.06 0.12p 0.00 0.00 0.00 0.00 0.00 0.00 0.00

SatE GR q 20.23 20.17 0.23 0.09 0.13 20.44 0.41p 0.00 0.00 0.00 0.00 0.00 0.00 0.00

SB q 20.40 20.12 0.40 0.35 20.13 20.57 0.55p 0.00 0.00 0.00 0.00 0.00 0.00 0.00

InfE GR q 20.10 0.05 20.02 0.39 20.13 20.57 0.58p 0.00 0.00 0.02 0.00 0.00 0.00 0.00

SB q 20.26 0.01 0.16 0.34 20.15 20.57 0.59p 0.00 0.09 0.00 0.00 0.00 0.00 0.00

Recharge GR q 20.17 0.00 0.23 0.09 20.26 20.12 0.18p 0.00 0.98 0.00 0.00 0.00 0.00 0.00

SB q 20.09 0.14 0.11 0.00 20.29 20.13 0.18p 0.00 0.00 0.00 0.81 0.00 0.00 0.00

aq is the Correlation Coefficient. p < 0:05 indicates statistical significance. ET: evapotranspiration, Inf: infiltration, SatE: saturationexcess, InfE: infiltration excess, GR: Grand River watershed, SB: Saginaw Bay watershed. forest*: broadleaf deciduous temperate tree;grass*: c3 non-arctic grass.

Table 6. Annual Budgetsa

GR GR (%P) SB SB (%P)

P 875.5 100.0 886.8 100.0SUBM 13.1 1.5 16.0 1.8InfE 47.3 5.4 57.7 6.5SatE 28.8 3.3 74.5 8.4Dperc 173.6 19.8 142.4 16.1Qoc 223.4 25.5 247.9 27.9Qgc 21.3 2.4 20.6 2.3Qout 243.2 27.8 267.8 30.2ET 603.1 68.9 583.2 65.8

aUnit: mm/yr. %P: percent of annual precipitation, P: precipitation,SUBM: sublimation, InfE: infiltration excess, SatE: saturationexcess, Dperc: groundwater recharge, Qoc: surface runoff, Qgc:groundwater outflow, GR: Grand River watershed, SB: Saginaw Baywatershed.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7373

on longer data sets maybe more reliable. To detect trends in TWSA for our period of study, we have used thenonparametric Mann-Kendall (MK) test [Helsel and Hirsch, 2002] and computed the Sen’s slope (s) for bothdata sets (GRACE data and model results). The Mann-Kendall test is used to test the null hypothesis H0 thatthere is no trend in the data. To establish a trend by rejecting the null hypothesis, the p value from the testmust be less than a significance level (typically 0.05). Test results based on 10 years (2002–2012) of GRACEdata indicated that storage changes have an increasing trend for both GR and SB watersheds (Mann-Kendallp < 0:05). The results of the trend analysis are shown in Figure 13 for the GR watershed and similar results areobtained for the SB watershed as well (figure not shown). The Sen’s slope for GRACE data for the GR water-shed is 10.0078 mm/d over the 10 year period and the slope based on model results is 10.0085 mm/d, whichconfirm the increasing trend over the period considered.

Wimbrow [2012] analyzed trends in precipitation and streamflow records for the entire State of Michiganusing the Mann-Kendall test and its variants. Using precipitation data from nearly 563 weather stations andstreamflow data from all available USGS gauging stations within the State, he presented spatial maps ofannual trends in precipitation and streamflow. His results showed increasing trends in both precipitationand streamflow for the two watersheds considered in this work and over a major portion of the lower pen-insula of Michigan in general. The increasing trends were noted for the time period of the present study

and over a much longer periodspanning several decades.These trends in precipitationand streamflow are expectedto be factors responsible forthe increasing trend in storagefor the two watersheds.Increasing streamflows resultin increased storage in reser-voirs. Increasing precipitationincreases soil moisture whilealso increasing recharge to thegroundwater domain [McCart-ney and Smakhtin, 2010].

Recent research shows thatstorage (or residence time) isdecreasing in the Great Lakesdue to declining lake levelswhich can be attributed to a

10/11/02 02/23/04 07/07/05 11/19/06 04/02/08 08/15/09 12/28/10 05/11/12

−60

−40

−20

0

20

40

60GRACE JPL EBGRACE JPLPAWS

GWOVNCSSWECF

Date (mm/dd/yy)

Tota

l Wat

er S

tora

ge A

nom

aly

(mm

)

R2 = 0.52APB = 58.32%RMSE = 13.73

Figure 12. Comparison of total water storage anomaly (TWSA) between gravity-based satellite measurements (Gravity Recovery andClimate Experiment, GRACE data) and PAWS model simulations. The blue shaded area denotes error bands for the GRACE data.

10/11/02 02/23/04 07/07/05 11/19/06 04/02/08 08/15/09 12/28/10 05/11/12−60

−40

−20

0

20

40

60

GRACE GR, p = 0.00

GRACE GR, s = 0.0078 ± 0.003PAWS GR, p = 0.00

PAWS GR s = 0.0085 ± 0.00063

Date (mm/dd/yy)

Tota

l Wat

er S

tora

ge A

nom

aly

(mm

)

Figure 13. Trend analysis of total water storage anomaly (TWSA) of GRACE data and PAWSmodel results for the Grand River watershed based on the Mann-Kendall test and Sen’sslope estimate.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7374

number of factors including (a) declining subsurface contribution to lakes (b) declining surface runoff and (c)increasing open lake evaporation or decreasing ice cover among other factors [Gronewold and Stow, 2014].Nguyen et al. [2014] found that residence time of water in the inner Saginaw Bay decreased by 15% over aperiod of roughly four decades due to falling lake levels. Our results suggest a link between increasing storageon land and decreasing storage in the Great Lakes such as Lake Huron.

4. Conclusions

A process-based hydrologic model has been applied to two of the largest watersheds in the State of Michi-gan and the water budgets and storage changes were analyzed. In addition to estimating the water budg-ets and examining their dependence on several key factors, this paper also demonstrated the application ofa process-based hydrologic model to large watersheds in the Great Lakes region. The model was able tosimulate different hydrologic components and states including surface runoff, channel flow, groundwater,ET, soil moisture, soil temperature and changes in storage. Vegetation growth dynamics were also simulatedby coupling PAWS with the land surface model CLM. The model results are in good agreement with obser-vations as discussed earlier. The linear correlation analysis suggests that the variations of different hydro-logic components are sensitive to LULC and soil types. The process-based model was able to predict themagnitude of key hydrologic components and the partitioning of precipitation as well as changes in water-shed storage. The annual averages of ET range from 600 to 700 mm, approximately 60–70% of the annualprecipitation, which is comparable to estimates from the MODIS data and other studies reported in the liter-ature. The GRACE equivalent water thickness anomaly and the anomaly computed from model results arein agreement within the uncertainties of the data. In response to the questions we set out to answer earlier,a trend analysis of anomaly data based on the nonparametric Mann-Kendall test indicated that storage isincreasing in both watersheds over the period of study. We also found that subsurface components ofwater (mainly water in the vadose zone and the unconfined aquifer) are responsible for the observedchanges in storage and that surface water components as well as water in the confined aquifer contributedlittle to the observed storage changes. Finally, well-tested, process-based hydrologic models are capable ofreproducing observed trends in storage change within the uncertainties of the data.

ReferencesArnold, J. G., and P. M. Allen (1996), Estimating hydrologic budgets for three Illinois watersheds, J. Hydrol., 176(1-4), 57–77, doi:10.1016/

0022-1694(95)02782-3.Benson, C. H., and X. Wang (2006), Temperature-compensating calibration procedure for water content reflectometers, in Proceedings of

TDR 2006, Paper ID 50, 16 p., Purdue Univ., West Lafayette, Indiana. [Available at https://engineering.purdue.edu/TDR/Papers, lastaccessed 5Aug 2014.]

Berbery, E. H., Y. Luo, K. E. Mitchell, and A. K. Betts (2003), Eta model estimated land surface processes and the hydrologic cycle of the Mis-sissippi basin, J. Geophys. Res., 108(D22), 8852, doi:10.1029/2002JD003192.

Beven, K. (2006), Searching for the Holy Grail of scientific hydrology: Qt5H S!; R!

� �A as closure, Hydrol. Earth Syst. Sci., 10(5), 609–618.

Black, P. E. (1997), Watershed functions, J. Am. Water Resour. Assoc., 33(1), 1–11, doi:10.1111/j.1752-1688.1997.tb04077.x.Cherkauer, K. A., and T. Sinha (2010), Hydrologic impacts of projected future climate change in the Lake Michigan region, J. Great Lakes

Res., 36, 33–50, doi:10.1016/j.jglr.2009.11.012.Christensen, J. H., et al. (2007), Regional climate projections, in Climate Change 2007: The Physical Science Basis. Contribution of Working

Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, edited by S. Solomon et al., pp. 887, Cam-bridge Univ. Press, Cambridge, U. K.

Creutzfeldt, B., P. A. Ferr�e, P. A. Troch, B. Merz, H. Wziontek, and A. G€untner (2012), Total water storage dynamics in response to climate var-iability and extremes—Inference from long-term terrestrial gravity measurement, J. Geophys. Res., 117, D08112, doi:10.1029/2011JD016472.

Creutzfeldt, B., P. Troch, P. A. Ferr�e, A. G€untner, T. Graeff, and B. Merz (2013), Storage-discharge relationships at different catchment scalesbased on local high-precision gravimetry, Hydrol. Processes, 28(3), 1465–1475, doi:10.1002/hyp.9689.

Dai, Y. J., R. E. Dickinson, and Y. P. Wang (2004), A two-big-leaf model for canopy temperature, photosynthesis, and stomatal conductance,J. Clim., 17(12), 2281–2299, doi:10.1175/1520-0442(2004)017<2281:ATMFCT>2.0.CO;2.

Dickinson, R. E., et al. (2002), Nitrogen controls on climate model evapotranspiration, J. Clim., 15(3), 278–295, doi:10.1175/1520-0442(2002)015<0278:NCOCME>2.0.CO;2.

Dingman, S. L. (2002), Physical Hydrology, 2nd ed., pp. 646, Waveland Press, Long Grove, Ill.Enviro-weather (2010), Enviro-weather Automated Weather Station Network (formerly known as MAWN), Michigan State University, East

Lansing, Mich. [Available at http://www.agweather.geo.msu.edu/mawn/.]Ferrant, S., Y. Caballero, J. Perrin, S. Gascoin, B. Dewandel, S. Aulong, F. Dazin, S. Ahmed, and J. C. Marechal (2014), Projected impacts of

climate change on farmers’ extraction of groundwater from crystalline aquifers in South India, Sci. Rep., 4, 3697, doi:10.1038/srep03697.

Fry, L. M., T. S. Hunter, M. S. Phanikumar, V. Fortin, and A. D. Gronewold (2013), Identifying streamgage networks for maximizing the effec-tiveness of regional water balance modeling, Water Resour. Res., 49, 2689–2700, doi:10.1002/wrcr.20233.

AcknowledgmentsThis research was funded by a NOAAgrant to the last author (award3002283555). We thank Han Qiu for hisassistance with data compilation,model runs, and postprocessing. Datasets used as model inputs or for modeltesting are owned by several agenciesincluding the USGS, USDA, NOAA,NASA/JPL, the Michigan Departmentof Natural Resources (MDNR) andMAWN (Michigan Automated WeatherNetwork or Enviro-Weather) anddetails of these sources (with web linkswhere available) are provided in thepaper. We acknowledge AGU’s datapolicy; however, we are not in aposition to share these publiclyavailable data sets, as we do not haveownership of the data.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7375

Gao, H., Q. Tang, C. R. Ferguson, E. F. Wood, and D. P. Lettenmaier (2010), Estimating the water budget of major US river basins via remotesensing, Int. J. Remote Sens., 31(14), 3955–3978, doi:10.1080/01431161.2010.483488.

Geruo, A., J. Wahr, and S. Zhong (2013), Computations of the viscoelastic response of a 3-D compressible Earth to surface loading: An appli-cation to Glacial Isostatic Adjustment in Antarctica and Canada, Geophys. J. Int., 192, 557–572, doi:10.1093/gji/ggs030.

Gronewold, A. D., and C. A. Stow (2014), Water loss from the Great Lakes, Science, 343(6175), 1084–1085, doi:10.1126/science.1249978.Helsel, D. R. and R. M. Hirsch (2002), Statistical methods in water resources, U.S. Geol. Surv. Tech. Water Resour. Invest., Book 4, Chap. 12,

pp. 323–352. [Available at http://water.usgs.gov/pubs/twri/twri4a3/.]Istanbulluoglu, E., T. Wang, O. M. Wright, and J. D. Lenters (2012), Interpretation of hydrologic trends from a water balance perspective:

The role of groundwater storage in the Budyko hypothesis, Water Resour. Res., 48, W00H16, doi:10.1029/2010WR010100.Jia, Y. W., G. H. Ni, Y. Kawahara, and T. Suetsugi (2001), Development of WEP model and its application to an urban watershed, Hydrol. Proc-

esses, 15(11), 2175–2194, doi:10.1002/hyp.275.Kling, G. W., et al. (2003), Confronting Climate Change in the Great Lakes Region: Impacts on our Communities and Ecosystems, 104 pp., Union

of Concerned Sci., Cambridge, Mass.Landerer, F. W., and S. C. Swenson (2012), Accuracy of scaled GRACE terrestrial water storage estimates, Water Resour. Res., 48, W04531,

doi:10.1029/2011WR011453.Long, D., B. R. Scanlon, L. Longuevergne, A.-Y. Sun, D. N. Fernando, and H. Save (2013), GRACE satellite monitoring of large depletion in

water storage in response to the 2011 drought in Texas, Geophys. Res. Lett., 40, 3395–3401, doi:10.1002/grl.50655.Long, D., L. Longuevergne, and B. Scanlon (2014), Uncertainty in evapotranspiration from land surface modeling, remote sensing, and

GRACE satellites, Water Resour. Res., 50, 1131–1151, doi:10.1002/2013WR014581.Lu, D., and Q. Weng (2006), Use of impervious surface in urban land-use classification, Remote Sens. Environ., 102(1–2), 146–160, doi:

10.1016/j.rse.2006.02.010.Luo, Y., E. H. Berbery, and K. E. Mitchell (2005), The operational eta model precipitation and surface hydrologic cycle of the Columbia and

Colorado basins, J. Hydrometeorol., 6(4), 341–370, doi:10.1175/JHM435.1.Maxwell, R. M., et al. (2014), Surface-subsurface model intercomparison: A first set of benchmark results to diagnose integrated hydrology

and feedbacks, Water Resour. Res., 1531–1549, doi:10.1002/2013WR013725.McCartney, M., and V. Smakhtin (2010), Water Storage in an Era of Climate Change: Addressing the Challenge of Increasing Rainfall Variability,

pp. 24, Int. Water Manage. Inst., Colombo, Sri Lanka. [Available at http://www.iwmi.cgiar.org/Publications/Blue_Papers/PDF/Blue_Paper_2010-final.pdf.]

MDNR (2010), 2001 IFMAP/GAP Lower peninsula land cover, Michigan Department of Natural Resources, Lansing, Mich. [Available athttp://www.mcgi.state.mi.us/mgdl/?rel5thext&action5thmname&cid55&cat5Land1Cover12001, last accessed 28 Nov 2009.]

Mitchell, K. E., et al. (2004), The multi-institution North American Land Data Assimilation System (NLDAS): Utilizing multiple GCIP productsand partners in a continental distributed hydrological modeling system, J. Geophys. Res., 109, D07S90, doi:10.1029/2003JD003823.

Mu, Q., F. A. Heinsch, M. Zhao, and S. W. Running (2007), Development of a global evapotranspiration algorithm based on MODIS andglobal meteorology data, Remote Sens. Environ., 111, 519–536, doi:10.1016/j.rse.2007.04.015.

Mu, Q., M. Zhao, and S. W. Running (2011), Improvements to a MODIS global terrestrial evapotranspiration algorithm, Remote Sens. Environ.,115(8), 1781–1800, doi:10.1016/j.rse.2011.02.019.

Nguyen, T. D., P. Thupaki, E. J. Anderson and M. S. Phanikumar (2014), Summer circulation and exchange in the Saginaw Bay—Lake HuronSystem, J. Geophys. Res. Oceans, 119, 2713–2734, doi:10.1002/2014JC009828.

Pan, M., and E. F. Wood (2006), Data assimilation for estimating the terrestrial water budget using a constrained ensemble Kalman filter,J. Hydrometeorol., 7(3), 534–547, doi:10.1175/JHM495.1.

Pokhrel, Y. N., Y. Fang, G. Miguez-Macho, P.-J. F. Yeh, and S. Han (2013), The role of groundwater in the Amazon wate cycle: 3. Influence onterrestrial water storage computations and comparison with GRACE, J. Geophys. Res. Atmos., 118, 3233–3244, doi:10.1002/jgrd.50335.

Price, K., R. Storn, and J. Lampinen (2005), Differential Evolution: A Practical Approach to Global Optimization, Springer, Berlin.Riley, W. J., and C. Shen (2014), Characterizing coarse-resolution watershed soil moisture heterogeneity using fine-scale simulations, Hydrol.

Earth Syst. Sci., 18(7), 2463–2483, doi:10.5194/hess-18-2463-2014.Roads, J., et al. (2003), GCIP water and energy budget synthesis (WEBS), J. Geophys. Res., 108(D16), 8609, doi:10.1029/2002JD002583.Rodell, M., J. S. Famiglietti, J. L. Chen, S. I. Seneviratne, P. Viterbo, S. Holl, and C. R. Wilson (2004), Basin scale estimates of evapotranspiration

using GRACE and other observations, Geophys. Res. Lett., 31, L20504, doi:10.1029/2004GL020873.Rodell, M., J. Chen, H. Kato, J. Famiglietti, J. Nigro, and C. Wilson (2007), Estimating groundwater storage changes in the Mississippi River

basin (USA) using GRACE, Hydrogeol. J, 15(1), 159–166, doi:10.1007/s10040-006-0103-7.Saravanathiiban, D. S. (2014), Preferential flow through earthen landfill covers: Field evaluation of root zone water quality model and labo-

ratory validation of lattice Boltzmann method, PhD dissertation, Dep. of Civ. Environ. Eng., Mich. State Univ., East Lansing.Sawicz, K., T. Wagener, M. Sivapalan, P. A. Troch, and G. Carrillo (2011), Catchment classification: Empirical analysis of hydrologic similarity

based on catchment function in the eastern USA, Hydrol. Earth Syst. Sci., 15(9), 2895–2911, doi:10.5194/hess-15–2895-2011.Sayama, T., J. J. McDonnell, A. Dhakal, and K. Sullivan (2011), How much water can a watershed store?, Hydrol. Processes, 25(25), 3899–3908,

doi:10.1002/hyp.8288.Schaap, M. G., F. J. Leij, and M.Th. van Genuchten (2001), ROSETTA: A computer program for estimating soil hydraulic parameters with hier-

archical pedotransfer functions, J. Hydrol., 251, 163–176.Seilheimer, T. S., P. L. Zimmerman, K. M. Stueve, and C. H. Perry (2013), Landscape-scale modeling of water quality in Lake Superior and

Lake Michigan watersheds: How useful are forest-based indicators?, J. Great Lakes Res., 39(2), 211–223, doi:10.1016/j.jglr.2013.03.012.Sea Grant College Program (1985), Bulletins E-1866-70, Cooperative Extension Service, Michigan State University, East Lansing, Mich.Shen, C., J. Niu, and M. S. Phanikumar (2013), Evaluating controls on coupled hydrologic and vegetation dynamics in a humid continental

climate watershed using a subsurface-land surface processes model, Water Resour. Res., 49, 1–21, doi:10.1002/wrcr.20189.Shen, C., J. Niu, and K. Fang (2014), Quantifying the effects of data integration algorithms on the outcomes of a subsurface–land surface

processes model, Environ. Model. Software, 59, 146–161, doi:10.1016/j.envsoft.2014.05.006.Shen, C. P., and M. S. Phanikumar (2010), A process-based, distributed hydrologic model based on a large-scale method for surface-

subsurface coupling, Adv. Water Resour., 33(12), 1524–1541, doi:10.1016/j.advwatres.2010.09.002.Sivapalan, M. (2006), Pattern, process and function: Elements of a unified theory of hydrology at the catchment scale, in Encyclopedia of

Hydrological Sciences, 193–219, John Wiley, Chichester, U. K., doi:10.1002/0470848944.hsa012.Steele-Dunne, S. C., M. M. Rutten, D. M. Krzeminska, M. Hausner, S. W. Tyler, J. Selker, T. A. Bogaard, and N. C. van de Giesen (2010), Feasibil-

ity of soil moisture estimation using passive distributed temperature sensing, Water Resour. Res., 46, W03534, doi:10.1029/2009WR008272.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7376

Su, F., and D. P. Lettenmaier (2009), Estimation of the surface water budget of the La Plata Basin, J. Hydrometeorol., 10(4), 981–998, doi:10.1175/2009JHM1100.1.

Swenson, S. C., and J. Wahr (2006), Post-processing removal of correlated errors in GRACE data, Geophys. Res. Lett., 33, L08402, doi:10.1029/2005GL025285.

Tague, C., G. Grant, M. Farrell, J. Choate, and A. Jefferson (2008), Deep groundwater mediates streamflow response to climate warming inthe Oregon Cascades, Clim. Change, 86(1-2), 189–210, doi:10.1007/s10584-007-9294-8.

Thornton, P. E., and N. E. Zimmermann (2007), An improved canopy integration scheme for a land surface model with prognostic canopystructure, J. Clim., 20(15), 3902–3923, doi:10.1175/JCLI4222.1.

Toon, O. B., C. P. Mckay, T. P. Ackerman, and K. Santhanam (1989), Rapid calculation of radiative heating rates and photodissociation ratesin inhomogeneous multiple-scattering atmospheres, J. Geophys. Res., 94(D13), 16,287–16,301, doi:10.1029/JD094iD13p16287.

Trapp, R. J., N. S. Diffenbaugh, H. E. Brooks, M. E. Baldwin, E. D. Robinson, and J. S. Pal (2007), Changes in severe thunderstorm environmentfrequency during the 21st century caused by anthropogenically enhanced global radiative forcing, Proc. Natl. Acad. Sci. U. S. A., 104(50),19,719–19,723, doi:10.1073/pnas.0705494104.

Wagener, T., M. Sivapalan, P. Troch, and R. Woods (2007), Catchment classification and hydrologic similarity, Geogr. Compass, 1(4), 901–931,doi:10.5194/hess-15-2895-2011.

Wang, H., H. Guan, H. A. Gutierrez-Jurado, C. T. Simmons (2014), Examination of water budget using satellite products over Australia,J. Hydrol., 511, 546–554.

Wang, T., E. Istanbulluoglu, J. Lenters, and D. Scott (2009), On the role of groundwater and soil texture in the regional water balance: Aninvestigation of the Nebraska Sand Hills, USA, Water Resour. Res., 45, W10413, doi:10.1029/2009WR007733.

Wimbrow, G. H. (2012), A statistical analysis of streamflow trends for the State of Michigan, MS thesis, Dep. of Civ. Environ. Eng., Mich. StateUniv., East Lansing.

Wuebbles, D., and K. Hayhoe (2004), Climate change projections for the United States Midwest, Mitigation Adaptation Strategies GlobalChange, 9(4), 335–363, doi:10.1023/B:MITI.0000038843.73424.de.

Xia, Y. L., et al. (2012a), Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation Sys-tem project phase 2 (NLDAS-2): 1. Intercomparison and application of model products, J. Geophys. Res., 117, D03109, doi:10.1029/2011JD016048.

Xia, Y. L., et al. (2012b), Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation Sys-tem project phase 2 (NLDAS-2): 2. Validation of model-simulated streamflow, J. Geophys. Res., 117, D03110, doi: 10.1029/2011JD016051.

Zreda, M., D. Desilets, T. P. A. Ferr�e, and R. L. Scott (2008), Measuring soil moisture content non-invasively at intermediate spatial scaleusing cosmic-ray neutrons, Geophys. Res. Lett., 35, L21402, doi:10.1029/2008GL035665.

Water Resources Research 10.1002/2014WR015589

NIU ET AL. VC 2014. American Geophysical Union. All Rights Reserved. 7377