Spatiotemporal Evaluation of Simulated Evapotranspiration and Streamflow Over Texas Using the Wrf-Hydro-Rapid Modeling Framework1

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

Vol. 54, No. 1 AMERICAN WATER RESOURCES ASSOCIATION February 2018

SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK1

Peirong Lin, Mohammad Adnan Rajib , Zong-Liang Yang, Marcelo Somos-Valenzuela , Venkatesh Merwade, David R. Maidment, Yan Wang, and Li Chen2

ABSTRACT: This study assesses a large-scale hydrologic modeling framework (WRF-Hydro-RAPID) in terms of its high-resolution simulation of evapotranspiration (ET) and streamflow over Texas (drainage area: 464,135 km2). The reference observations used include eight-day ET data from MODIS and FLUXNET, and daily river discharge data from 271 U.S. Geological Survey gauges located across a climate gradient. A recursive digital filter is applied to decompose the river discharge into surface runoff and base flow for comparison with the model counterparts. While the routing component of the model is pre-calibrated, the land component is uncalibrated. Results show the model performance for ET and runoff is aridity-dependent. ET is better predicted in a wet year than in a dry year. Streamflow is better predicted in wet regions with the highest efficiency ~0.7. In comparison, streamflow is most poorly predicted in dry regions with a large positive bias. Modeled ET bias is more strongly correlated with the base flow bias than surface runoff bias. These results complement previous evaluations by incorporating more spatial details. They also help identify potential processes for future model improvements. Indeed, improving the dry region streamflow simulation would require synergistic enhancements of ET, soil moisture and groundwater parameterizations in the current model configuration. Our assessments are important preliminary steps towards accurate large-scale hydrologic forecasts.

(KEY TERMS: evapotranspiration; streamﬂow; surface runoff; base ﬂow; MODIS; WRF-Hydro; Noah-MP; RAPID.)

Lin, Peirong, Mohammad Adnan Rajib, Zong-Liang Yang, Marcelo Somos-Valenzuela, Venkatesh Merwade, David R. Maidment, Yan Wang, and Li Chen, 2018. Spatiotemporal Evaluation of Simulated Evapotranspiration and Streamﬂow over Texas Using the WRF-Hydro-RAPID Modeling Framework. Journal of the American Water Resources Association (JAWRA) 54(1): 40-54. https://doi.org/10.1111/1752-1688.12585

INTRODUCTION in transformation of hydrologic forecasting approaches from statistical and stochastic methods based on recorded observations (McEnery et al., The ability and accuracy of weather prediction 2005) to deterministic and probabilistic forecasts have greatly improved over the past decade, resulting based on numerical weather prediction (NWP) tools

1Paper No. JAWRA-16-0033-P of the Journal of the American Water Resources Association (JAWRA). Received January 27, 2016; accepted August 9, 2017. © 2017 American Water Resources Association. Discussions are open until six months from issue publication. 2Graduate Research Assistant (Lin) and Professor (Yang), Jackson School of Geosciences, University of Texas at Austin, 2225 Speedway, Stop C1160, Austin, Texas 78712; Graduate Research Assistant (Rajib) and Associate Professor (Merwade), Lyles School of Civil Engineering, Purdue University, West Lafayette, Indiana 47907; Professor (Yang), Key Laboratory of Regional Climate-Environment Research for Temper- ate East Asia of Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029, China; Associate Professor (Somos-Valen- zuela), Department of Agricultural and Forestry Sciences, University of La Frontera, Temuco 4780000, Chile; Professor (Maidment), Center for Research in Water Resources, University of Texas at Austin, Austin, Texas 78712; and Graduate Student (Wang) and Assistant Professor (Chen), College of Hydrology and Water Resources, Hohai University, Nanjing 210000, China (E-Mail/Yang: [email protected]).

JAWRA 40 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

(Cloke and Pappenberger, 2009). Numerous studies hydrologic simulations. While global- and continen- have highlighted recent efforts in developing such tal-scale evaluations of Noah-MP have been docu- forecasting systems for Europe (Thielen et al., 2009), mented (e.g., Yang et al., 2011; Cai et al., 2014a), Africa (Thiemig et al., 2010), South America (Paiva these studies implemented the model at a resolution et al., 2011), and the entire globe (Pappenberger usually coarser than 0.125° and used limited gauge et al., 2012; Alfieri et al., 2013; Winsemius et al., locations at major rivers and basin outlets as a refer- 2013). ence for model evaluation. David et al. (2013) evalu- Likewise, a continental-scale high-resolution ated RAPID’s performance using runoff outputs from hydrologic forecasting system for the United States multiple LSMs, with a focus on the routing parame- (U.S.) has been developed recently (Maidment, 2016), ter optimization, but they did not establish linkages which literally transforms a real-time weather map between streamflow prediction errors with biases in to a river forecast map, using an earth system model- LSM-simulated water fluxes. Therefore, it is impor- ing framework. This hydrologic forecasting system tant to assess the model performance at finer spa- uses the community WRF-Hydro model (Gochis et al., tiotemporal scales than previous studies and 2015), an architectural framework which leverages establish the relationships between the errors in the the U.S. operational weather forecasting, the state-of- modeled streamflow and the biases in runoff compo- the-art land surface models (LSMs), and the geo- nents. referenced vector river network — the National This study uses WRF-Hydro-RAPID over Texas in Hydrography Dataset Plus (NHDPlus), providing the offline mode such that Noah-MP is run in a streamflow forecasts at 2.67 million river reaches stand-alone mode (i.e., uncoupled to WRF) to simu- operationally across the contiguous U.S. Comparing late land surface hydrology (Gochis et al., 2015). Due to previous gauge-based forecasting points at ~4,000 to emerging interests in using satellite-based evapo- nationally (McEnery et al., 2005), this new forecast- transpiration (ET) data to calibrate hydrologic mod- ing system represents a factor of 741 increase in spa- els (e.g., Kunnath-Poovakka et al., 2016), this study tial coverage, signaling a substantial shift from also evaluates the modeled ET against the Moderate watershed hydrology to continental hydrology. Resolution Imaging Spectroradiometer (MODIS) ET The National Flood Interoperability Experiment to understand the potential usefulness of the remo- (NFIE), a research collaboration among academia, tely sensed ET product in improving streamflow pre- federal agencies, and private partners, was proposed dictions. to translate the unprecedented information on river While the current operational U.S. National Water flow forecasts to inundation maps, which will eventu- Model (NWM) has evolved from the NFIE initiative, ally inform local emergency response planning (Maid- it has several advanced features compared to the ment, 2016). Since the onset of NFIE, continuous model configurations presented in this study. Specifi- efforts have been made in a wide range of aspects, cally, NWM has a streamflow data assimilation capa- including the development of baseline geospatial bility, a more physically based hydrologic routing hydrologic data for the continental flood forecasting scheme, a feature allowing simple level-pool routing system (Viger et al., 2016), the river hydraulics and for over 1,200 major reservoirs, and tools for model inundation mapping capability at high resolution calibrations (http://water.noaa.gov/about/nwm). (Follum et al., 2016), and the development of web The aim of this study is not to obtain the best- platforms for warning delivery to emergency respon- achievable prediction skills as required in an opera- ders and the general public (Swain, 2015). While tional setting. Instead, it is to establish a better these efforts are related mostly to the communication understanding of the model behaviors across a cli- and the translation of hydrologic simulations, their mate gradient with a focus on the interplay between potential strengths and weaknesses in terms of water biases in ET and individual runoff components. Based balances have not been well evaluated. Such evalua- upon the default Noah-MP configurations, the results tion is essential for ensuring the robustness of the presented here should be viewed as a reference for hydrologic modeling framework across the intended the baseline model performance, with a hope to bene- spatiotemporal scales. fit future model users who work on specific water- As the demonstration version of the continental- sheds in Texas and similar regions and to guide scale river modeling system, the WRF-Hydro-RAPID future model improvement activities. framework was used at the 2015 NFIE Summer The model and data used in this study are first Institute (Maidment, 2016). Within this framework, described, followed by the methodology. Next, results the refined Noah LSM with multi-parameterization and discussions on assessing the modeled ET, daily options (hereafter Noah-MP) (Niu et al., 2011) and streamflow, and the individual runoff components are the Routing Application for Parallel computatIon of presented. The last section summarizes the major Discharge (RAPID) (David et al., 2011) were used for conclusions of this study.

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 41 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN

MODEL AND DATA kept uncalibrated with its default parameter values as suggested by Niu et al. (2011) and Cai et al. (2014a). The time steps of Noah-MP and RAPID are 3-h and 15- Model Description and Experimental Configurations min, respectively; the model outputs are averaged to daily values from 2008 to 2011. WRF-Hydro (Gochis et al., 2015) is the hydrological extension of the Weather Research and Forecasting model (WRF), a next-generation NWP tool for research Observational Data and operational purposes, which has a large worldwide user community mainly in atmospheric science (http:// The 15-min river discharge data are obtained from www.wrf-model.org/index.php). As a model coupling the U.S. Geological Survey (USGS) gauge stations in architecture for easier linkage between hydrology sim- the Texas Hydrologic Region 12 using the ulations with atmospheric models, WRF-Hydro incor- HydroDesktop batch downloading tool (Ames et al., porates horizontal routing processes and water 2012). In total, the Region 12 contains 731 USGS resource management and provides several LSM gauges, but only 271 of them have continuous time options and hydrologic routing options to serve various series covering the entire study period (2008-2011). hydrometeorological applications. Among the several Since filling out missing values would artificially routing options, WRF-Hydro supports the alternatives introduce additional uncertainties, we only use these of grid-based and reach-based schemes (Gochis et al., 271 gauge observations in the model evaluation. 2015), one of which is the RAPID model. RAPID We use the remote sensing-derived eight-day actual directly resolves streamflow computations on the ET data from the MODIS MOD 16 product (Mu et al., NHDPlus river network in a vector-based environment 2011). Despite the uncertainties commonly seen in (David et al., 2011). Compared to grid-based routing remotely-sensed estimates, the MOD16 ET is used as a models, vector-based models have the advantages of reference to compare with the modeled ET because it is higher computational efficiency, and more effective arguably one of the best available large-scale ET data- hazard communication and visualization due to its geo- sets with a reasonably high spatiotemporal resolution graphic information system (GIS) capabilities (Lehner and well-documented uncertainty assessments. Mu and Grill, 2013; Yamazaki et al., 2013), and thus WRF- et al. (2013) found that the MODIS ET product has Hydro-RAPID is a hybrid framework consisting of both ~0.3 mm/day of average mean absolute error with grid- and vector-based modeling units. WRF-Hydro respect to 46 FLUXNET site estimates across North encapsulates several state-of-the-art LSMs including and South America. An independent site-specific Noah-MP (Niu et al., 2011); Noah-MP is used in this assessment showed that the root-mean-square-error of study because it was adopted by both the NFIE and the the MOD16 product ranged from 26 to 32 mm per operational NWM to simulate land surface hydrology. month (Velpuri et al.,2013).Inthisstudy,MODIS The model is configured at ~5km9 5 km grid resolu- eight-day total ET data at ~1-km grid spacing are geo- tion (269 9 314 grid cells in total) and driven by 3- referenced, spatially re-scaled, and aggregated using an hourly atmospheric forcing from the North American automatic data processor tool (Rajib et al., 2016) to Land Data Assimilation (NLDAS-2). The simulation match the ~5-km model grids over Texas (three MODIS period is from January 1, 2008 to December 31, 2011. tiles covering the main portions of Texas are used). As Noah-MP provides multiple physical parameterization an additional reference, the in situ ET measurement options within one model, from which the runoff from a FLUXNET station (Freeman Ranch-Woodland, parameterization scheme with a simple 1-dimensional Texas, 29.9495°N, 97.9962°W) is also used to compare groundwater model (Niu et al., 2007) is used. Cai et al. with both MODIS ET and model-simulated ET. This (2014a) found that the groundwater model in Noah-MP FLUXNET station is the only one of the four available requires 34 years for spinning up for the Mississippi FLUXNET sites within our study domain that contains River Basin. Texas is drier than the Mississippi as a sufficient time-series data for the simulation period. All whole, so we conduct a 100-year spin-up process for our the above-mentioned observational datasets are used domain by driving the model with the 2008 atmo- only for model evaluation, but not in any form of param- spheric forcing for 100 times. The gridded total runoff eter estimation/calibration purposes. is passed to RAPID, using a catchment centroid-based grid-to-vector coupling interface (Lin et al., 2015) to simulate horizontal river transport. RAPID uses a Study Domain matrix-based Muskingum method to solve horizontal routing, and the Muskingum parameters k (travel Figure 1 shows the study domain, the major river time) and x (weighting factor) are obtained from an basins, and the distribution of climatological aridity optimal set derived by David et al. (2013). Noah-MP is index (AI) of the region. The LSM domain covers the

JAWRA 42 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

streamflow data into surface runoff and base flow components. This technique relies on the notion that high frequency variability of the streamflow is primarily caused by direct runoff. Hence, base flow can be identified by low-pass filtering of the streamflow hydrograph. Within its semi-automated structure, the filter algorithm takes the streamflow hydrograph, an exponential base flow recession constant, a, and the maximum value of base flow index (BFImax, long-term ratio of base flow to total streamflow) as the user inputs. While a can be determined by recession analyses (e.g., Tallaksen, 1995; Sujono et al., 2004), BFImax is a nonmeasureable quantity and a fitting constant for the algorithm (Eckhardt, 2005). In this study, daily streamflow hydrographs from all the 271 USGS gauges are filtered to obtain corresponding daily time series of surface runoff and base flow. In a separate setting before inserting into the FIGURE 1. Geographic Location, Aridity, and Major River Basins filter algorithm, each of the streamflow time series of the Study Domain. P/PET is defined as the ratio of annual precipitation to annual potential evapotranspiration based on covering their respective maximum range of available 36-year climatology. data until 2014 is analyzed to identify the dates showing zero streamflow. Catchments having no entire state of Texas to represent a strong west-east streamflow at the outlet for more than 50% of the precipitation gradient, while the river routing domain days within a year and consistently for each year of is set up for the Hydrologic Region 12, which drains their entire record are characterized as ephemeral. 2 an area of 464,135 km with 68,143 NHDPlus river As BFImax cannot be determined prior to separation, reaches excluding the Rio Grande, the Canadian, and two typical values are used over the entire study the Red River Basins based on the NHDPlus domain domain depending on the perennial (BFImax = 0.8) discretization. Defined as the ratio of annual precipi- and ephemeral (BFImax = 0.5) nature of the catch- tation to annual potential evapotranspiration (PET), ments (e.g., Eckhardt, 2005; Partington et al., 2012). AI is calculated, using the 36-year average NLDAS-2 Since it is inefficient to conduct numerical recession monthly forcing data. A high (low) AI value indicates analysis for calculating a for each of the streamflow a humid (arid) climate. According to the climate types hydrographs, a constant value of 0.96 is adopted from defined by the United Nations Environmental Pro- a multi-model comparison conducted by Eckhardt gramme, Texas has an arid climate (AI < 0.2) in the (2008) on 65 North American watersheds. To tenta- west, a semiarid climate (0.2 ≤ AI ≤ 0.5) in the cen- tively assess the sensitivity of these chosen values, tral, and a subhumid to humid climate (AI > 0.5) in BFImax and a are varied one at a time by 0.20 and the east. In this paper, we mainly refer to the dry 0.03 for four randomly chosen gauge stations, which (wet) region of Texas as regions with an AI ≤ 0.3 (vs. produces a maximum of 4% and 9.5% difference in AI > 0.3) to reflect the west-east climatic contrast. the average annual surface runoff, respectively (not The state has a large precipitation gradient with the shown here). Such small differences in response to annual rainfall ranging from 200 to 1,400 mm, and it the variations in BFImax and a indicate the represen- has high vulnerability to extreme events. The Gulf tativeness of these chosen values, although a shows a Coast is threatened by frequent hurricane-induced relatively stronger influence on the filter results com- flooding, while central/western Texas experiences pared to BFImax (Eckhardt, 2012). The overall perfor- recurrent flash droughts and floods. mance of the filter is compared against a recent study by the USGS and the Texas Water Develop- ment Board (Baldys and Schalla, 2016). Table 1 shows that the average BFI calculated, using Eck- METHODOLOGY hardt’s filter for five representative USGS gauge stations in the Brazos River Basin is only 8% higher than that reported by Baldys and Schalla (2016), Runoff Separation using the local-minimum method of Hydrograph Separation and Analysis (HYSEP) program. A semi-empirical two-parameter recursive digital The same filter is also applied to the model simu- filter (Eckhardt, 2005, 2008) is used to partition the lated streamflow at locations of the 271 USGS

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 43 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN

TABLE 1. Comparison of the Calculated Eckhardt’s BFI with the HYSEP-Derived BFI by Baldys and Schalla (2016). covðQ ; Q Þ CC ¼ obs mod ð1Þ robsrmod Calculated BFI Reference BFI USGS (Eckhardt’s Recursive (Baldys and Schalla, Station ID Digital Filter): 2008-2011 2016): 1966-2009 P N ðQi Qi Þ2 NSE ¼ 1 Pi¼i mod obs ð2Þ 08082000 0.2 0.18 N ð i Þ2 08080500 0.2 0.16 i¼i Qobs Qobs 08084000 0.35 0.32 08085500 0.24 0.24 qPffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 08095000 0.4 0.32 N ð i i Þ2= Average 0.27 0.24 i¼i Qmod Qobs N NRMSE ¼ ð3Þ Q Note: BFI, base flow index; HYSEP, Hydrograph Separation and obs Analysis program with local minimum method; USGS, U.S. Geolo- gical Survey. P N i i ¼ ðQ Q Þ gauges. Since the same filter is applied to separate NBIAS ¼ i iP mod obs ð4Þ N i both the modeled and gauge-observed streamflow, the i¼i Qobs uncertainty induced by the filter can be considered as minimal in the evaluation. RESULTS AND DISCUSSION

Skill Metrics for Daily Streamflow Evaluation Evapotranspiration Four skill metrics are used to evaluate the model performance in simulating daily streamflow, which Figure 2 shows the ET difference between the includes Pearson correlation coefficient (CC), Nash- model and the MODIS ET averaged across the four Sutcliffe efficiency (NSE), normalized root-mean- seasons of a wet year (2008) and a dry year (2011). square-error (NRMSE), and normalized bias Regardless of wet/dry years or seasons, it is seen (NBIAS). In Equations (1-4), Qobs and Qmod denote that the model consistently underpredicts ET over observed and model-simulated streamflow, respec- the wet southeastern Texas and overpredicts over the tively; N is the number of days. While these skill drier northwestern Texas and the Mexican Plateau metrics are mathematically related (Gupta et al., region when compared against MODIS ET data. The 2009), each of them is interpreted with a slightly dif- magnitude of the difference represents 0-50% under- ferent focus to facilitate a more comprehensive prediction in the wet region and more than 100% understanding on the model performance. CC (Equa- overprediction in the arid region, respectively. Differ- tion 1) measures the proportion of the total variance ence between modeled and MODIS ET is the largest of the observed data that can be explained by the in the summer (JJA) and smallest in the winter model (Legates and McCabe, 1999), which cannot be (DJF); in DJF the difference is generally less than used to assess model biases. NSE (Equation 2) and 0.3 mm/day, which is comparable to the documented RMSE, as the two most commonly used criteria in uncertainty in the MODIS ET estimates (e.g., Mu hydrological evaluations, measure both the variabil- et al., 2013). By comparing the spatially averaged ity of time series and the magnitude of errors (Gupta monthly MODIS ET with four NLDAS-2 LSMs, Long et al., 2009). NSE normalizes the squared model et al. (2014) observed a more dampened seasonal error, using the variance of the observed data, and cycle in remote sensing-derived ET (hereafter RS-ET) thus a value of zero suggests the model is only as products. This was explained by the limited respon- good as the mean observed data, while NSE = 1 sug- siveness to precipitation in RS-ET products mostly gests perfect model simulation and negative values due to a lack of soil wetness constraints, which is suggest large model water balance errors (Gupta and explicitly accounted for in the LSM-ET estimates. Kling, 2011). Without scaling, RMSE tends to have Compared to the generic finding on the “model over- larger values for main streams with higher dis- estimation” problem in JJA by Long et al. (2014), Fig- charge, and therefore, to provide an objective assess- ure 2 detects spatially explicit details on model ment of the model error across spatial scales without behavior showing both overprediction and underpre- mixing up with large/small streams, NRMSE (Equa- diction tendency of the model during the same sea- tion 3) is used. NBIAS (Equation 4) is calculated to son. Such a spatial pattern of the ET over- and identify both the magnitude and the sign of errors in underpredictions across Texas can also be seen in the mean streamflow. three other NLDAS-2 LSMs (Cai et al., 2014b),

JAWRA 44 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

FIGURE 2. Difference in the Modeled and the MODIS ET Estimates during a Wet Year (2008) and a Dry Year (2011). Seasonal mean differences are shown for spring (MAM), summer (JJA), autumn (SON), and winter (DJF). suggesting possible model structural or parameteriza- encompasses the land heterogeneity including soil, tion problems in simulating ET over contrasting cli- vegetation, and geological conditions. Therefore, as matic regions. detailed below, it is necessary to clarify several land Although both the MODIS ET and the LSM ET surface physical parameters and resistance terms use the Penman-Monteith (PM) approach, the main that are directly used in the PM approach, which differences between the two estimates can be attribu- would help shed light on the distinct spatial pattern ted to three factors: (1) the driving meteorology and of the ET flux difference in our case. energy data, or the forcing data difference, (2) the In the current version of Noah-MP, soil surface treatment of vegetation (i.e., whether the leaf area resistance (rs), a term describing the restriction for index (LAI) is prescribed or predicted; if prescribed, water vapor to diffuse from soil surface into the air, whether it is based on long-term climatology or satel- uses a parameterization proposed by Sakaguchi and lite data for specific years), and (3) the treatment of Zeng (2009) (hereafter SZ2009) being updated from soil wetness constraints. The spatial difference in ET the previous parameterization of Sellers et al. (1992). is negatively correlated (p < 0.05) with aridity (Fig- In SZ2009, rs is defined to account for soil wetness ure 3), an integrated climatic factor which implicitly and texture factors as shown in Equations (5-7).

L r ¼ ð5Þ s D

h 2þ3b ¼ h2 r ð Þ D D0 sat 1 6 hsat

exp½ð1 h =h Þw1 L ¼ d 1 sat ; ð7Þ 1 e 1

where D represents vapor diffusivity within the soil 2 FIGURE 3. Scatter Plot of the ET Difference (m /s), which is a function of soil texture, saturation (model observation) vs. the Aridity Index (annual P/PET). Each soil moisture hsat, and residual water content hr; and dot in the plot represents one grid cell. L represents dry layer thickness (m), which is a

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 45 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN

function of top layer soil wetness h1/hsat. There are North Texas to San Antonio in the south, where the two parameters (b and w) and three constants (D0, input soil texture is primarily sand. Among the 12 d1, and e) in the SZ2009 parameterization: b is the soil categories used in the model, sand supposedly soil texture parameter (Clapp and Hornberger, 1978); holds the least available water to sustain ET, and w is an empirical parameter controlling the exponen- thus the model predicts a smaller ET over the region tial relationship between L and h1/hsat. D0 is the compared to the MODIS ET, an estimate which does molecular diffusion coefficient of water vapor in the not explicitly consider the soil texture control on ET 5 2 air (2.2 9 10 m /s); d1 is the thickness of topsoil fluxes. layer (0.1 m); and e is 2.718. SZ2009 improves the In Figure 4, we plot the CC between the model previous soil evaporation parameterization by better and the MODIS eight-day ET time series for 2008 constraining soil wetness (Sakaguchi and Zeng, and 2011. At each ~5-km grid cell, CC is calculated 2009), yet the parameter controlling how L varies using 48 data samples along the whole year to indi- with the top layer soil wetness (i.e., w in Equation 7) cate the agreement in temporal variability between is set to 5 globally based on limited empirical field the model and the MODIS ET estimates. The two experiments. It is hence likely that the large wet/dry estimates have a better temporal agreement in a climate gradient may not be well-represented, using normal to wet year (2008) than an extreme drought a constant value for w in spatially distributed simula- year (2011), as indicated by more grid cells with a tions, given the sensitivity of rs to w has been docu- relatively higher CC in Figure 4a than Figure 4b. mented (Sakaguchi and Zeng, 2009) and that of ET to 2011 was an extreme drought year for Texas, and a w is observed in our preliminary investigation (not large extent of Texas observed reductions in green- shown here). By testing 139 hard-coded parameters ness (Sun et al., 2015). In our model configuration, in Noah-MP, a recent study (Cuntz et al., 2016) also however, a static monthly LAI based on the MODIS concluded that w is the most sensitive yet overlooked LAI climatology is used (the default Noah-MP LAI parameter for ET and runoff simulations. Therefore, option). This may result in incorrect greenness rep- it is suggested for future studies to adjust or derive resentations in the model, and consequently, erro- spatially distributed values for parameter based on neous hydrologic simulations. Noah-MP also RS-ET data, or to revisit the soil surface resistance provides a dynamic vegetation (DV) option, where parameterization in Noah-MP by developing more vegetation-related variables are calculated based on appropriate process-based schemes (e.g., Zeng et al., the climate, soil, and vegetation conditions at each 2011; Zhang et al., 2015) in order to dampen the model time step, albeit with its uncertainties and model’s oversensitivity to w (Cuntz et al., 2016). limitations. The results thus suggest the need for These could potentially improve the overall ET simu- utilizing the DV option in real-time hydrologic fore- lation, especially in western Texas where the direct casting cases and carefully evaluating the coupled soil evaporation is a major component due to the rela- vegetation-hydrologic processes, which could improve tively sparse vegetation on land surface. For the wet the temporal water fluxes simulation especially dur- southeastern Texas where canopy transpiration plays ing extreme wet or dry conditions. an important role, treatment of stomatal resistance Figure 5 shows the time series of model-simulated (rstomatal) (Ball et al., 1987), LAI (which defines struc- ET with corresponding MODIS estimates in 2008 and tural parameter for the vegetation greenness), and 2011 at four different locations. One location is a mechanisms such as the root water uptake (Teuling FLUXNET site (Freeman Ranch, Woodland) in which et al., 2006) could also lead to ET biases. Due to a case we assume that the in situ measurement (point lack of observations for these individual components, data) is representative of the average hydrologic con- however, this study falls short in concluding which dition in the particular 5-km grid cell where the transpiration estimate (model or MODIS) is more tower is located. The results are consistent with Fig- reliable. Accordingly, a more comprehensive under- ure 2 as the model and MODIS estimates have larger standing on the individual ET components (i.e., soil discrepancies in JJA than in other seasons. Besides, evaporation and plant transpiration) should benefit the model seems to overpredict ET fluxes in the dry from finer resolution ET measurements based on region compared to MODIS. At Locations A and B, much advanced algorithms in near future to better model and MODIS ET are in good agreement in both inform hydrologic modeling. years; at Location C, MODIS ET is able to show a Prominent ET differences between model and drop in ET fluxes in a drought year while the model MODIS are also found over two noticeable blue bands still predicts falsely high ET fluxes in that year. in the northeast-southwest direction of the southeast- From all locations, it is apparent that modeled ET ern Texas (Figure 2), where model predicts a much tends to show more responsiveness to high ET values smaller ET flux than the MODIS. This is a region temporally, which is consistent with Long et al. that roughly runs 480 km from the Red River in (2014). At the FLUXNET site, MODIS ET shows a

JAWRA 46 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

FIGURE 4. Correlation Coefﬁcient between the MODIS and the Modeled Eight-Day ET Estimates over (a) 2008 and (b) 2011. The statistics are computed using 48 data samples. Locations for the time series in Figure 5 are denoted in blue boxes and circles. relatively better match with the in situ measure- temporal variability of the in situ measurements bet- ments in terms of mean ET values, especially for ter, especially in capturing the higher ET peaks dur- 2008 where the mean MODIS ET is ~0.3 mm/day clo- ing both 2008 and 2011. Note that the site has an ser to the in situ measurements than modeled ET. aridity of ~0.4 and the land cover type is woody On the other hand, the model tends to imitate the savannah whereas the model’s dominant land cover

FIGURE 5. Time Series Comparison between the MODIS and the Modeled ET Estimates at Four Locations Shown in Figure 4. The in situ measurements are from the Freeman Ranch, Woodland FLUXNET site. The ET estimates from MODIS, model, and FLUXNET are shown in black, red, and green, respectively. X-axis is time series every eight days, and y-axis is ET in mm/day.

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 47 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN type is cropland/grassland mosaic. The closer match simulated streamflow against hundreds of USGS between MODIS ET and site observations in mean gauge observations (Figure 6). values but not in temporal variability may justify the Four skill metrics for the daily streamflow simula- use of MODIS ET as a spatial reference dataset to tion are spatially mapped in Figure 6. It is seen that adjust model simulations, although its applicability central and northeastern Texas has relatively higher as a temporal reference especially in regions of CC and NSE values (Figures 6a and 6b), suggesting diverse aridity should be comprehensively investi- reasonably good model performances over the sub- gated in the future. humid to humid regions of Texas, especially in the upper Trinity, the lower Colorado, and the San Jac- into River Basins where the NSE generally exceeds Streamflow 0.5. The best model performance is seen at the Onion Creek at Highway 183 (USGS gauge # 08159000), Figure 6 presents the model performance in simu- with the daily NSE equaling 0.66 based on the uncal- lating daily streamflow at 271 USGS gauges based on ibrated modeling results. With a drainage basin of a priori parameter set of the default WRF-Hydro- 831.4 km2, this gauge was one of the most severely RAPID configuration as described in the Model and flooded areas in central Texas during the 2013 Hal- Data section. Complementary to previous studies loween flood, and the results clearly highlight the evaluating Noah-MP at the outlets of major river model’s potential in issuing effective hydrologic fore- basins (Yang et al., 2011; Cai et al., 2014a), our eval- casts for this particular gauge (or the adjacent uation includes all the available gauges to provide a regions in general). However, large biases are seen in more comprehensive perspective of the model perfor- the western part and the Nueces River Basin, as dis- mance that has more practical values to the end played by negative NSE (Figure 6b) and large users of hydrologic forecasts. Since the routing com- NRMSE and NBIAS values (Figures 6c and 6d) over ponent of this modeling framework performs stream- these regions. Calibrating the LSM component flow computations at the level of individual NHDPlus (Noah-MP) might have produced a much better fit river reaches, each streamflow prediction is associ- with the observed streamflow hydrographs, but would ated with one exact geographic location and a unique impart an unknown degree of undetectable bias in identification, making it easier to evaluate the the surface runoff and base flow fluxes because of

FIGURE 6. Model Performance for Daily Streamﬂow Simulations from 2008 to 2011 (1,461 data samples). Four skill metrics are used, including (a) CC, (b) NSE, (c) NRMSE, and (d) NBIAS at 271 gauges across the Texas Hydrologic Region 12.

JAWRA 48 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

FIGURE 7. Hydrographs at Four USGS Gauge Stations (a–d); the Statistics Calculated Using the Daily Streamflow Are Shown. Parts (e) and (f) show the spatial pattern of the NSE calculated using the daily and monthly streamflow, respectively. Red dots are locations for the four stations shown on the left. parameter equifinality (Beven and Freer, 2001). assessing monthly/annual streamflow simulation Thus, it is the goal of this study to objectively docu- skills by the NLDAS-2 LSMs in 961 small watersheds ment the performance of the default WRF-Hydro- across the U.S. Similar to our findings, Cai et al. RAPID modeling framework, instead of conducting (2014a, b) and Xia et al. (2012) indicated relatively model calibration. larger biases in simulated streamflow, especially Figure 7 shows daily streamflow hydrographs at across the drier Great Plains. However, all the previ- four USGS gauges, along with their performance met- ous studies included a very limited portion of north/ rics against respective observed data. Figures 7a-7c northwestern Texas, not to mention their coarser represent three gauges which have relatively good timescale. The distinct spatial pattern of streamflow model performances with the NSEs equal to 0.66, bias as shown in Figures 6 and 7 resonates and com- 0.56, and 0.46, respectively. This is an indicator of plements the general findings from previous studies the model being quite effective in these areas with its regarding a potential weakness of the LSM in captur- uncalibrated configuration. On the contrary, Fig- ing semiarid to arid region processes, suggesting the ure 7d represents a gauge located in the drier west- need for future model improvements or applying bias ern Texas showing a poor NSE, which is mainly due corrections over these regions to fulfill operational to the overpredicted base flow and the flashiness in goals. surface runoff in the model simulation. In that part of Texas, the NSE values based on monthly averaged streamflow are not better than the daily NSE values Surface Runoff and Base Flow Components (Figures 7e and 7f), suggesting a persistent problem in the LSM to generate the correct amount of total To further understand the error sources in the runoff, which is not an issue likely to be resolved river flow predictions, Figure 8 provides a large-scale from better physical representations of the horizontal evaluation of the LSM-simulated surface runoff and flow routing processes. This is in line with the previ- base flow components. It is clearly seen that surface ous studies (Xia et al., 2012; Cai et al., 2014b) runoff is underpredicted in the humid eastern Texas

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 49 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN

FIGURE 8. Difference between Model and Observations Averaged from 2008 to 2011. Part (a) is for surface runoff (RUNSF) and (b) is for base ﬂow (RUNSB). Parts (c) and (d) show the percentage difference [(model obs.)/obs. 9 100%] for RUNSF and RUNSB, respectively. Warm (cold) color indicates the model underpredicts (overpredicts) ﬂow as compared to the observation.

(Figure 8a, warm colors), whereas base flow is over- the general base flow overprediction issue, some predicted by the model (Figure 8b, cold colors). The slight underpredictions are also seen over the three magnitude of difference between model and observa- wettest regions, including the San Jacinto River tion is bounded by 1 mm/day. For the arid to semi- Basin, the upper Trinity River Basin, and the Guada- arid western Texas, both surface runoff and base flow lupe and San Antonio River Basins (generally within are overpredicted by the model, with a positive bias 0to40%). less than +0.1 mm/day. However, given the dry clima- To understand the possible interplay between ET tology of this region, a bias of +0.1 mm/day bias is and two individual runoff components, we plot the quite significant as it represents an error percentage bias in modeled surface runoff (Figure 9a) and base of more than +100% with respect to the observed data flow (Figure 9b) against the bias in ET averaged over (see Figures 8c and 8d). Overall, large mass-balance 2008 to 2011. The term “bias” is used here under the errors are evident in the arid west and the relatively assumption that the USGS runoff and the MODIS dry Nueces River Basin in the southwest. This fur- ET data provide the best available observational esti- ther explains the negative NSE values (Figure 6b) mates, which may not be true due to data uncertain- over the drier parts of Texas, where the mass-balance ties. However, the rationale behind conducting this error dominates the level of model accuracy despite a analysis is due to the emerging interests in using generally reasonable capture of the temporal variabil- earth observations to calibrate hydrologic models at ity in streamflow (Figure 6a, CC ranging from 0.2 to large spatial scales, where RS-ET products are con- 0.6 in western Texas). In the humid eastern Texas, sidered as the best-available “observation” to improve the reasonable to good model performances are attrib- streamflow simulations (Cai et al., 2014a; Kunnath- uted to the relatively small percentage differences Poovakka et al., 2016). It is thus important to analyze (within 50%). The only exception is seen over a how these biases in the mutually dependent hydro- northeast-southwest band in eastern Texas (Fig- logic processes can be related with the theoretically ure 6b, negative NSE values). Within this band, the expected model performances, before remotely sensed low NSE values correspond well with the large base data should be introduced in model calibration. flow overprediction (>+100%) problem, suggesting In Figure 9, the ET bias and runoff bias are calcu- possible model improvements that are largely depen- lated as the drainage basin averages to represent dent on the base flow simulation in this area. Despite physically-consistent quantities. Each dot represents

JAWRA 50 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

FIGURE 9. (a) Scatter Plot of ET Bias (calculated using model minus MODIS ET) vs. Surface Runoff Bias (calculated using separated model results minus separated USGS surface runoff) Averaged from 2008 to 2011. Part (b) is a similar plot but for base ﬂow.

181 out of the 271 drainage areas of the USGS moisture and groundwater storages. Theoretically, gauges, whose area is larger than 625 km2 (5 9 5 when model-predicted storages have large positive LSM grid cells) to filter out some random noises that (negative) biases, it is likely that both of the outgoing are present in small watersheds. The color scheme fluxes (i.e., runoff and ET) can be underpredicted represents the AI at the 181 gauge locations, ranging (overpredicted) to close water balance in long-term from 0.2 to 0.6. Ideally, errors in ET and runoff (i.e., averages. Figure 9 shows that more than half of the two outgoing fluxes from the land surface) are dry region gauges (with overpredicted ET, surface expected to show a complementary negative relation- runoff, and base flow) and a few wet region gauges ship to close the water balance. Due to data uncer- (with underpredicted ET and surface runoff) are sub- tainties and the limited number of years used, it is ject to this problem. Therefore, at these locations, one not surprising that such a perfect relationship is less might not expect improved model performances by obvious in Figure 9, similar to what is being docu- simply calibrating the model using ET observations. mented in a recent study evaluating multiple LSMs Instead, more hydrologic observations (both in situ (Zhang et al., 2016). However, Figure 9 still displays and RS-based) on soil moisture and water table depth an interesting pattern where the base flow bias, com- (WTD) are needed to further improve the streamflow pared to surface runoff bias, is more distinguishably prediction. In our current model configuration, the influenced by the bias in ET (Figure 9b, a significant Noah-MP runoff parameterization scheme is tightly negative correlation). This is reasonable because base coupled to a simple 1D groundwater model with an flow is more directly influenced by soil moisture and unconfined aquifer (Niu et al., 2007), in which runoff groundwater storages in which the ET partitioning is parameterized as an exponential decay function of plays a significant role. Accordingly, model calibra- the model-simulated WTD. Under this parameterization against MODIS ET may improve base flow simu- tion scheme, Cai et al. (2014a) reported a reasonably lation, as also implied by Kunnath-Poovakka et al. simulated spatial pattern of WTD in their study (2016). In comparison, the bias in surface runoff domain, yet it was noted that the magnitude of WTD seems to be less sensitive to the bias in ET (Fig- was simulated too shallow. In an attempted analysis ure 9a). This is because generation of surface runoff to compare modeled WTD with the Texas Water is also closely related to rainfall intensity and land Development Board well measurements (not shown), surface heterogeneity in addition to being influenced we found that the simulated WTD is indeed too shal- by soil wetness conditions. Although arguably ET can low especially in the arid western Texas, which is still affect surface runoff by controlling the antece- likely related to the underrepresented subsurface dent soil moisture profile at the onset of a rainfall properties. This may help explain the overpredicted event, this cannot be ascertained from the current ET, surface runoff, and base flow in most of dry study due to the highly complex nonlinearity in the region gauges according to the Noah-MP runoff rainfall-runoff processes and data uncertainties. parameterization. However, this study is limited in Future studies need to be conducted with various ET proving the hypothesis, which warrants further products to further understand this problem. investigation in this direction. Besides, this study is Figure 9 also displays some unexpected patterns based on a four-year model simulation and one source where both the runoff and ET fluxes can be over-pre- of the RS ET product. More general and robust con- dicted (or underpredicted) in the same direction. clusions may benefit from multiple long-term simula- Such patterns can be ascribed to data uncertainties, tions and quantitative analyses with different data or it may also reflect model biases in the soil products in future studies.

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 51 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN

CONCLUSIONS the ongoing improvements of the NWM, a continental-scale high resolution flood forecasting framework operationally running for the U.S. (http://water.noaa. Uncertainty in streamflow prediction from a large- gov/about/nwm). Given that our uncalibrated model- scale hydrologic forecasting system mainly originates ing results under a much simpler model configuration from meteorological forcing, as well as the represen- than that adopted by the NWM still produce promis- tations of land surface and river routing processes. ing results for a range of gauge locations, and consid- This study is an attempt to evaluate the spatiotempo- ering there are continuous efforts by the NWM ral performance of an offline WRF-Hydro-RAPID development team that focuses on reducing the modeling experiment by focusing on its ET and uncertainty in streamflow prediction from all error streamflow simulations over Texas. First, by compar- sources (e.g., from atmospheric forcing to human ing model simulated ET and MODIS ET, we found influences), it is expected that the operational hydro- that, spatially, the modeled ET performance is arid- logic prediction skills of the nation will be enhanced ity-dependent; temporally, the ET fluctuations are in the future. Our study, by assessing the large-scale better captured in a normal-to-wet year than in an fine-resolution ET and streamflow simulations over extreme-drought year. Second, by comparing model- Texas, is one important initial step towards accurate simulated streamflow against 271 gauge observations, hydrologic forecasts. a more complete picture of the model performance is provided to complement previous studies that mainly focused on major rivers and basin outlets. In Texas, ACKNOWLEDGMENTS streamflow is better predicted in the wet region than in the dry region, with the highest daily NSE at ~0.7 This work was funded by the National Natural Science Founda- based on uncalibrated results. In comparison, pres- tion of China grant 41375088, the NSF Coupled Natural and Human Systems Program award 1518541, the Cynthia and George ence of the large positive runoff bias over the dry Mitchell Family Foundation, the Texas Water Research Network, region indicates a persistent problem in simulating and Microsoft Research. The authors are grateful to the CUAHSI the mass-balance components by the model. Third, by funding support for the Summer Institute (June 1 to July 18, 2015) applying a recursive digital filter to decompose runoff held at the National Water Center, Alabama, U.S. We thank David components, the results provide a large-scale evalua- Gochis (National Center for Atmospheric Research), Jim Nelson (Brigham Young University), and Fernando Salas (NOAA National tion of the surface and base flow simulations by Water Center) for providing project advice as NFIE advisors and Noah-MP, the LSM in the WRF-Hydro framework. course coordinators. David Arctur (University of Texas at Austin) We found that surface runoff is underpredicted while is thanked for English editing. We would also like to thank the base flow is over-predicted in the wet region. In the Editor-in-Chief Jim Wigington, Associate Editor David Tarboton dry region, both surface runoff and base flow are lar- (Utah State University), and three anonymous reviewers for their constructive comments. gely over-predicted, which results in the poor model performance there. Finally, by linking the errors (the differences calculated as the model estimates minus the reference estimates) between ET and the two LITERATURE CITED runoff components, we found that base flow errors Alfieri, L., P. Burek, E. Dutra, B. Krzeminski, D. Muraro, J. Thie- have a stronger correlation with ET errors than in len, and F. Pappenberger, 2013. GloFAS-Global Ensemble the case for surface runoff. This has implications for Streamflow Forecasting and Flood Early Warning. Hydrology large-scale model calibration studies that use RS-ET and Earth System Sciences 17:1161-1175. https://doi.org/10. data to improve streamflow predictions. At most of 5194/hess-17-1161-2013. the dry region gauges where runoff and ET errors do Ames, D.P., J.S. Horsburgh, Y. Cao, J. Kadlec, T. Whiteaker, and D. Valentine, 2012. HydroDesktop: Web Services-Based Soft- not show complementary relationships, calibrating ware for Hydrologic Data Discovery, Download, Visualization, ET may not yield improved streamflow predictions. and Analysis. Environmental Modelling & Software 37:146-156. In these situations, possible solutions to model https://doi.org/10.1016/j.envsoft.2012.03.013. improvements focusing on better physical parameteri- Baldys III, S. and F.E. Schalla, 2016. Base Flow (1966-2009) and zations of soil moisture and/or groundwater might Streamflow Gain and Loss (2010) of the Brazos River from the New Mexico-Texas State Line to Waco, Texas. U.S. Geological render more useful solutions through future endeav- Survey Scientific Investigations Report 2011-5224, 53 pp. http:// ors. pubs.usgs.gov/sir/2011/5224/report/SIR2011-5224.pdf. Although this study was conducted under the Ball, J.T., I.E. Woodrow, and J.A. Berry, 1987. A Model Predicting scope of the 2015 NFIE, our goal spanned beyond Stomatal Conductance and Its Contribution to the Control of flood prediction with a view to provide a holistic Photosynthesis under Different Environmental Conditions. In: Progress in Photosynthesis Research, J. Biggins (Editor). assessment on the model’s overall hydrologic perfor- Springer, Dordrecht, The Netherlands, pp. 221-224. https://doi. mance. Accordingly, the findings would complement org/10.1007/978-94-017-0519-6_48.

JAWRA 52 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION SPATIOTEMPORAL EVALUATION OF SIMULATED EVAPOTRANSPIRATION AND STREAMFLOW OVER TEXAS USING THE WRF-HYDRO-RAPID MODELING FRAMEWORK

Beven, K. and J. Freer, 2001. Equifinality, Data Assimilation, and Sensed Evapotranspiration and Soil Moisture for Streamflow Uncertainty Estimation in Mechanistic Modelling of Complex Prediction. Journal of Hydrology 535:509-524. ISSN 0022-1694. Environmental Systems Using the GLUE Methodology. Journal https://doi.org/10.1016/j.jhydrol.2016.02.018. of Hydrology 249(1-4):11-29. ISSN 0022-1694, https://doi.org/10. Legates, D.R. and G.J. McCabe, Jr., 1999. Evaluating the Use of 1016/s0022-1694(01)00421-8. “Goodness-of-Fit” Measures in Hydrologic and Hydroclimatic Cai, X., Z.-L. Yang, C.H. David, G.-Y. Niu, and M. Rodell, 2014a. Model Validation. Water Resources Research 35(1):233-241. Hydrological Evaluation of the Noah-MP Land Surface https://doi.org/10.1029/1998WR900018. Model for the Mississippi River Basin. Journal of Geophysical Lehner, B. and G. Grill, 2013. Global River Hydrography and Net- Research-Atmospheres 119:23-38. https://doi.org/10.1002/2013 work Routing: Baseline Data and New Approaches to Study the JD020792. World’s Large River Systems. Hydrological Processes 27:2171- Cai, X., Z.-L. Yang, Y. Xia, M. Huang, H. Wei, L.R. Leung, and 2186. https://doi.org/10.1002/hyp.9740. M.B. Ek, 2014b. Assessment of Simulated Water Balance from Lin, P., Z.-L. Yang, X. Cai, and C.H. David, 2015. Development and Noah, Noah-MP, CLM, and VIC over CONUS Using the NLDAS Evaluation of a Physically-Based Lake Level Model for Water Test Bed. Journal of Geophysical Research-Atmospheres Resource Management: A Case Study for Lake Buchanan, Tex- 119:13751-13770. https://doi.org/10.1002/2014JD022113. as. Journal of Hydrology: Regional Studies, 4B, 661-674. ISSN Clapp, R.B. and G.M. Hornberger, 1978. Empirical Equations for 2214–5818. https://doi.org/10.1016/j.ejrh.2015.08.005. Some Soil Hydraulic Properties. Water Resources Research 14 Long, D., L. Longuevergne, and B.R. Scanlon, 2014. Uncertainty in (4):601-604. https://doi.org/10.1029/WR014i004p00601. Evapotranspiration from Land Surface Modeling, Remote Sens- Cloke, H.L. and F. Pappenberger, 2009. Ensemble Flood Forecast- ing, and GRACE Satellites. Water Resources Research 50:1131- ing: A Review. Journal of Hydrology 375:613-626. https://doi.org/ 1151. https://doi.org/10.1002/2013WR014581. 10.1016/j.jhydrol.2009.06.005. Maidment, D.R., 2016. Conceptual Framework for the National Cuntz, M., J. Mai, L. Samaniego, M. Clark, V. Wulfmeyer, O. Flood Interoperability Experiment. Journal of the American Branch, S. Attinger, and S. Thober, 2016. The Impact of Stan- Water Resources Association 53(2):245-257. https://doi.org/10. dard and Hard-Coded Parameters on the Hydrologic Fluxes in 1111/1752-1688.12474. the Noah-MP Land Surface Model. Journal of Geophysical McEnery, J., J. Ingram, Q. Duan, T. Adams, and L. Anderson, Research-Atmospheres 121:10676-10700. https://doi.org/10.1002/ 2005. NOAA’s Advanced Hydrologic Prediction Service: Building 2016JD025097. Pathways for Better Science in Water Forecasting. Bulletin of David, C.H., D.R. Maidment, G.-Y. Niu, Z.-L. Yang, F. Habets, and the American Meteorological Society 86:375-385. https://doi.org/ V. Eijkhout, 2011. River Network Routing on the NHDPlus 10.1175/BAMS-86-3-375. Dataset. Journal of Hydrometeor 12(5):913-934. https://doi.org/ Mu, Q., M. Zhao, and S.W. Running, 2011. Improvements to a 10.1175/2011JHM1345.1. MODIS Global Terrestrial Evapotranspiration Algorithm. David, C.H., Z.-L. Yang, and S. Hong, 2013. Regional-Scale River Remote Sensing of Environment 115(8):1781-1800. https://doi. Flow Modeling Using Off-the-Shelf Runoff Products, Thousands org/10.1016/j.rse.2011.02.019, accessed September 2017. of Mapped Rivers and Hundreds of Stream Flow Gauges. Envi- Mu, Q., M. Zhao, and S.W. Running, 2013. Algorithm Theoretical ronmental Modelling and Software 42:116-132. https://doi.org/ Basis Document: MODIS Global Terrestrial Evapotranspiration 10.1016/j.envsoft.2012.12.011. (ET) Product (NASA MOD16A2/A3) Collection 5, National Ocea- Eckhardt, K., 2005. How to Construct Recursive Digital Filters for nic and Atmospheric Administration. http://www.ntsg.umt.edu/ Baseflow Separation. Hydrological Processes 19:507-515. sites/ntsg.umt.edu/files/MOD16_ATBD.pdf. https://doi.org/10.1002/hyp.5675. Niu, G.-Y., Z.-L. Yang, R.E. Dickinson, L.E. Guldn, and H. Su, Eckhardt, K., 2008. A Comparison of Baseflow Indices, Which Were 2007. Development of a Simple Groundwater Model for Use in Calculated with Seven Different Baseflow Separation Methods. Climate Models and Evaluation with Gravity Recovery and Cli- Journal of Hydrology 352(1–2):168-173. https://doi.org/10.1016/j. mate Experiment Data. Journal of Geophysical Research, 112: jhydrol.2008.01.005. D07103. https://doi.org/10.1029/2006jd007522. Eckhardt, K., 2012. Technical Note: Analytical Sensitivity Analysis Niu, G.-Y., Z.-L. Yang, K.E. Mitchell, F. Chen, M.B. Ek, M. Bar- of a Two Parameter Recursive Digital Baseflow Separation Fil- lage, A. Kumar, K. Manning, D. Niyogi, E. Rosero, M. Tewari, ter. Hydrology and Earth System Sciences 16:451-455. https://d and Y. Xia 2011. The Community Noah Land Surface Model oi.org/10.5194/hess-16-451-2012. with Multiparameterization Options (Noah-MP): 1. Model Follum, M.L., A.A. Tavakoly, J.D. Niemann, and A.D. Snow, 2016. Description and Evaluation with Local-Scale Measurements. AutoRAPID: A Model for Prompt Streamflow Estimation and Journal of Geophysical Research 116:D12109. https://doi.org/10. Flood Inundation Mapping over Regional to Continental 1029/2010JD015139. Extents. Journal of the American Water Resources Association Paiva, R.C.D., W. Collischonn, and C.E.M. Tucci, 2011. Large Scale 53(2):280-299. https://doi.org/10.1111/1752-1688.12476. Hydrologic and Hydrodynamic Modeling Using Limited Data Gochis, D.J., W. Yu, and D.N. Yates, 2015. The WRF-Hydro Model and a GIS Based Approach. Journal of Hydrology 406(3–4):170- Technical Description and User’s Guide, Version 3.0. NCAR 181. ISSN 0022-1694. https://doi.org/10.1016/j.jhydrol.2011.06. Technical Document, 120 p. http://www.ral.ucar.edu/projects/ 007. wrf_hydro/, accessed September 2017. Pappenberger, F., E. Dutra, F. Wetterhall, and H.L. Cloke, 2012. Gupta, H.V. and H. Kling, 2011. On Typical Range, Sensitivity, Deriving Global Flood Hazard Maps of Fluvial Floods through a and Normalization of Mean Squared Error and Nash-Sutcliffe Physical Model Cascade. Hydrology and Earth System Sciences Efficiency Type Metrics. Water Resources Research 47:W10601. 16:4143-4156. https://doi.org/10.5194/hess-16-4143-2012. https://doi.org/10.1029/2011WR010962. Partington, D., P. Brunner, C.T. Simmons, A.D. Werner, R. Ther- Gupta, H.V., H. Kling, K.K. Yilmaz, and G.F. Martinez, 2009. rien, H.R. Maier, and G.C. Dandy, 2012. Evaluation of Outputs Decomposition of the Mean Squared Error and NSE Perfor- from Automated Baseflow Separation Methods Against Simu- mance Criteria: Implications for Improving Hydrological Model- lated Baseflow from a Physically Based, Surface Water-Ground- ling. Journal of Hydrology 377(1–2):80-91. https://doi.org/10. water Flow Model. Journal of Hydrology 458-459:28-39. https://d 1016/j.jhydrol.2009.08.003. oi.org/10.1016/j.jhydrol.2012.06.029. Kunnath-Poovakka, A., D. Ryu, L.J. Renzullo, and B. George, 2016. Rajib, M.A., V. Merwade, and Z. Yu, 2016. Multi-Objective Calibra- The Efficacy of Calibrating Hydrologic Model Using Remotely tion of a Hydrologic Model Using Spatially Distributed Remotely

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 53 JAWRA LIN,RAJIB,YANG,SOMOS-VALENZUELA,MERWADE,MAIDMENT,WANG, AND CHEN

Sensed/In-Situ Soil Moisture. Journal of Hydrology 536:192-207. Point and Gridded FLUXNET and Water Balance ET. Remote https://doi.org/10.1016/j.jhydrol.2016.02.037. Sensing of Environment 139:35-49. https://doi.org/10.1016/j.rse. Sakaguchi, K. and X. Zeng, 2009. Effects of Soil Wetness, Plant Lit- 2013.07.013. ter, and Under-Canopy Atmospheric Stability on Ground Evapo- Viger, R.J., A. Rea, J.D. Simley, and K.M. Hanson, 2016. ration in the Community Land Model (CLM3.5). Journal of NHDPlusHR: A National Geospatial Framework for Surface- Geophysical Research 114:D01107. https://doi.org/10.1029/ Water Information. Journal of the American Water Resources 2008JD010834. Association 52(4):901-905. https://doi.org/10.1111/1752-1688. Sellers, P.J., M.D. Heiser, and F.G. Hall, 1992. Relations between 12429. Surface Conductance and Spectral Vegetation Indices at Inter- Winsemius, H.C., L.P.H. Van Beek, B. Jongman, P.J. Ward, and A. mediate (100 m2 to 15 km2) Length Scales. Journal of Geophysi- Bouwman, 2013. A Framework for Global River Flood Risk cal Research 97(D17):19033-19059. https://doi.org/10.1029/ Assessments. Hydrology and Earth System Sciences 17:1871- 92JD01096. 1892. https://doi.org/10.5194/hess-17-1871-2013. Sujono, J., S. Shikasho, and K. Hiramatsu, 2004. A Comparison of Xia, Y., K. Mitchell, M. Ek, J. Sheffield, B. Cosgrove, E. Wood, L. Techniques for Hydrograph Recession Analysis. Hydrological Luo, C. Alonge, H. Wei, J. Meng, B. Livneh, D. Lettenmaier, V. Processes 18:403-413. https://doi.org/10.1002/hyp.1247. Koren, Q. Duan, K. Mo, Y. Fan, and D. Mocko, 2012. Continen- Sun, Y., R. Fu, R. Dickinson, J. Joiner, C. Frankenberg, L. Gu, Y. tal-Scale Water and Energy Flux Analysis and Validation for Xia, and N. Fernando, 2015. Drought Onset Mechanisms North American Land Data Assimilation System Project Phase Revealed by Satellite Solar-Induced Chlorophyll Fluorescence: 2 (NLDAS-2): 2. Validation of Model-Simulated Streamflow. Insights from Two Contrasting Extreme Events. Journal of Geo- Journal of Geophysical Research 117:D03110. https://doi.org/10. physical Research: Biogeosciences 120:2427-2440. https://doi.org/ 1029/2011JD016051. 10.1002/2015JG003150. Yamazaki, D., G.A.M. de Almeida, and P.D. Bates, 2013. Improving Swain, N.R., 2015. Tethys Platform: A Development and Hosting Computational Efficiency in Global River Models by Implement- Platform for Water Resources Web Apps. All Theses and Disser- ing the Local Inertial Flow Equation and a Vector-Based River tations, Paper 5832. http://scholarsarchive.byu.edu/etd/5832, Network Map. Water Resources Research 49:7221-7235. accessed September 2017. https://doi.org/10.1002/wrcr.20552. Tallaksen, L.M., 1995. A Review of Baseflow Recession Analysis. Yang, Z.-L., G.-Y. Niu, K.E. Mitchell, F. Chen, M.B. Ek, M. Bar- Journal of Hydrology 165(1–4):349-370. ISSN 0022-1694. lage, L. Longuevergne, K. Manning, D. Niyogi, M. Tewari, and https://doi.org/10.1016/0022-1694(94)02540-r. Y. Xia, 2011. The Community Noah Land Surface Model with Teuling, A.J., R. Uijlenhoet, F. Hupet, and P.A. Troch, 2006. Impact Multiparameterization Options (Noah-MP): 2. Evaluation over of Plant Water Uptake Strategy on Soil Moisture and Evapo- Global River Basins. Journal of Geophysical Research 116: transpiration Dynamics during Drydown. Geophysical Research D12110. https://doi.org/10.1029/2010JD015140. Letters 33:L03401. https://doi.org/10.1029/2005GL025019. Zeng, Y., Z. Su, L. Wan, and J. Wen, 2011. A Simulation Analysis Thielen, J., J. Bartholmes, M.-H. Ramos, and A. de Roo, 2009. The of the Advective Effect on Evaporation Using a Two-Phase Heat European Flood Alert System-Part 1: Concept and Development. and Mass Flow Model. Water Resources Research 47:W10529. Hydrology and Earth System Sciences 13:125-140. https://doi. https://doi.org/10.1029/2011WR010701. org/10.5194/hess-13-125-2009. Zhang, C., L. Li, and D. Lockington, 2015. A Physically Based Sur- Thiemig, V., F. Pappenberger, J. Thielen, H. Gadain, A. de Roo, K. face Resistance Model for Evaporation from Bare Soils. Water Bodis, M. Del Medico, and F. Muthusi, 2010. Ensemble Flood Resources Research 51(2):1084-1111. https://doi.org/10.1002/ Forecasting in Africa: A Feasibility Study in the Juba-Shabelle 2014WR015490. River Basin. Atmospheric Science Letters 11:123-131. https://d Zhang, Y., H. Zheng, F. Chiew, J. Pena-Arancibia,~ and X. Zhou, oi.org/10.1002/asl.266. 2016. Evaluating Regional and Global Hydrological Models Velpuri, N.M., G.B. Senay, R.K. Singh, S. Bohms, and J.P. Verdin, against Streamflow 10 and Evapotranspiration Measurements. 2013. A Comprehensive Evaluation of Two MODIS Evapotran- Journal of Hydrometeor 17(3):995-1010. https://doi.org/10.1175/ spiration Products over the Conterminous United States: Using JHM-D-15-0107.1.

JAWRA 54 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION