Integrating Climate Forecasts with the Soil and Water Assessment Tool (SWAT) for High-Resolution Hydrologic Simulations and Forecasts in the Southeastern U.S.

Abstract

This study provides high-resolution modeling of daily water budget components at Hydrologic Unit Code (HUC)-12 resolution for 50 watersheds of the South Atlantic Gulf (SAG) region in the southeastern U.S. (SEUS) by implementing the Soil and Water Assessment Tool (SWAT) model in the form of a near real-time, semi-automated framework. A near real-time hydrologic simulation framework is implemented with a lead time of nine months (March–December 2017) by integrating the calibrated SWAT model with National Centers for Environmental Prediction coupled forecast system model version 2 (CFSv2) weather data to forecast daily water balance components. The modeling exercise is conducted as a precursor for various future hydrologic studies (retrospective or forecasting) for the region by providing a calibrated hydrological dataset at high spatial (HUC-12) and temporal (1-day) resolution. The models are calibrated (January 2003–December 2010) and validated (January 2011–December 2013) for each watershed using the observed streamflow data from 50 United States Geological Survey (USGS) gauging stations. The water balance analysis for the region shows that the implemented models satisfactorily represent the hydrology of the region across different sub-regions (Appalachian highlands, plains, and coastal wetlands) and seasons. While CFSv2-driven SWAT models are able to provide reasonable performance in near real-time and can be used for decision making in the region, caution is advised for using model outputs as the streamflow forecasts display significant deviation from observed streamflow for all watersheds for lead times greater than a month.

1. Introduction

Global food security has come under severe threat, especially with the increasing human population and associated demand for freshwater resources and food production [1,2,3]. There are concerns over the increasing vulnerability of the southeastern U.S. (SEUS) to water-related conflicts due to continued growth in industrial, agricultural, metropolitan, and recreational water demand [4,5,6]. It is important to note that the SEUS is expected to have the largest absolute increase in population compared to any other region in the U.S. by 2030 [7,8]. Since water is a limited resource, it is likely that conflict for water sharing and use will become more commonplace and severe in the coming years.

Effective water resources planning has been very important for policymakers and engineers alike for this region, especially in the wake of several droughts in the region within a span of less than a decade. In the last fifteen years, the SEUS region has experienced an array of severe droughts, including severe and/or extreme drought conditions in early 2017 that impacted around half of the study region. These droughts have not only caused enormous strain on the agriculture and farms activities in the region but have also heightened instances of wildfires in the region, especially in the Appalachian region [9]. As the frequency of future hydrologic extremes is expected to increase, a reliable hydrologic forecast can be useful for the efficient planning of available water resources [10]. Previously, Lu et al. [11] developed an empirical model to estimate long-term annual actual evapotranspiration for forested watersheds in the SEUS. Limaye et al. [12] developed a macroscale hydrologic model for regional climate assessment studies in the SEUS. However, these studies do not address the watershed-scale variability in the region, which plays a crucial role in informed water resources management in the region.

Soil moisture and evapotranspiration (ET) are two important components of the hydrologic cycle and play an important role in the water, energy, and carbon cycle. ET plays an important role in linking the water, energy, and carbon cycles, and hence is crucial for the climate system and the hydrological cycle [13]. Accurate soil moisture and ET estimation is essential for delivering (or improving the quality of) reliable meteorological, hydrological, and environmental forecasts, such as precipitation [14], discharge estimates and prediction [15,16], identifying flash-flood prone areas and flood forecasting [17,18,19], the identification and characterization of droughts [20,21,22], and understanding soil–vegetation–atmosphere interactions [23,24] and boundary layer fluxes for improved water resource management [25,26]. However, both soil moisture and ET exhibit significant spatiotemporal variability across different scales owing to a range of processes, such as precipitation, and energy and water fluxes [27,28]. Factors including orography (on all scales), vegetation, and soil texture [29] also influence ET diurnally. There are no long-term observation data for soil moisture, unlike other hydrological variables, such as precipitation and streamflow, which are frequently recorded and reported. Additionally, large-scale remote sensing of ET using mechanistic or empirical models with the remotely sensed land surface and atmospheric properties [30,31,32,33] is popular; however, comparisons of the large number of remotely sensed ET datasets have indicated significant uncertainties associated with these products, with relative errors of 15–30% [33,34,35].

Currently, a combination of climate forecasts or reanalysis data with hydrological models is used to simulate the impacts of land use and climate change and other anthropogenic activities on the hydrological cycle [36,37]. These models provide a physical representation of the spatially variable hydrological processes [38] and key hydrological components such as ET and soil moisture. Previous studies have integrated hydrological models with a climate forecast dataset to forecast the state of hydrologic variables at seasonal/sub-seasonal time steps [39,40,41,42]. Utilizing information from seasonal climate forecasts and long-term climate predictions has proven to be of great interest to water planners and decision-makers to improve preparedness towards, and mitigation of, impacts due to climatic extremes [43,44]. Dynamical seasonal forecasting systems that are based on coupled atmosphere–ocean–land general circulation models (CGCMs) have been widely used for hydrologic forecasting in recent years [45,46]. Seasonal climate forecasts are useful in improving farmers’ ability to mitigate adversities linked to climate extremes, particularly in rain-fed systems [41]. Farmers can use such forecasts to improve their crop diversification strategies and to decide which crops to plant [47,48]. Skillful prediction of the field-scale soil moisture is important to decide crop type, resource use, and crop insurance and contract renewals [44]. Application of the climate forecast system reforecast and reanalysis (CFSRR) products in hydrologic monitoring and prediction, seasonal soil moisture estimation, and understanding land-atmosphere interactions [49,50,51,52,53,54] has also been extensively explored. Very recently, Zhang et al. [55] integrated CFSv2 [56] with the Variable Infiltration Capacity (VIC) model and satellite data to predict drought in southwestern China and reported that CFSv2-coupled hydrological models only provide reliable soil moisture drought predictions for lead times of 1 month. Recently, Kang and Sridhar [57] demonstrated the use of a CFSv2 weather forecast dataset in conjunction with VIC and SWAT models for improving drought prediction over the U.S.

However, most of these studies do not address the local landscape processes and biophysical heterogeneity to understand the watershed-scale hydrological responses to hydrological extremes, such as high/extreme precipitation or droughts, at local scales. Hydrological models implementable in a near real-time, semi-automated framework at the sub-watershed scale of 12-digit Hydrologic Unit Code (HUC-12) resolution are largely unavailable for the SAG region, despite growing water-related conflicts and growing demand for water for both agricultural and industrial applications. Recent studies by Kang and Sridhar [58] and Sehgal and Sridhar [59] highlight the importance of sub-watershed scale variability in hydrological processes and provide the impetus for high-resolution modeling with a focus on seasonal and sub-seasonal forecasting for the SAG region.

This study implements the SWAT model for 50 watersheds in the SEUS, calibrates and validates the models using the observed streamflow data, and provides retrospective hydrological simulations for the SAG region from 1982–2013 at a daily time step. Furthermore, the SWAT model is integrated with climate drivers from CFSv2 models to provide near real-time simulation and 9-month lead forecasts of hydrological variables from January 2014 to March 2017, and from April 2017 through to December 2017, respectively. The choice of SWAT model for simulating watershed hydrology is carried out by carefully evaluating the advantages of using semi-distributed hydrological models over lumped or fully distributed models, and the ease of implementation of the SWAT model compared to other similar, process-based models. While complete understanding of catchment hydrology may not be achievable using fully-distributed models, primarily due to the lack of quality dataset at a high spatial and temporal resolution, the lumped models are often criticized for not being capable of providing a reliable representation of the spatial variability in the land uses and hydrological processes [60]. Therefore, a semi-distributed model, such as SWAT, provides a tradeoff between fully-distributed and lumped models.

The objectives of this study are to provide high-resolution (HUC-12, daily), validated hydrological datasets for future application to the study region by implementing the following steps:

(a)

Develop the SWAT model to simulate various hydrologic variables at a daily time step for the South Atlantic-Gulf (SAG) region of the southeastern U.S. (SEUS) at HUC-12 (12-digit Hydrologic Unit Code) resolution.

(b)

Two-fold validation of the dataset using observed streamflow and in-situ soil moisture dataset.

(c)

Develop a near real-time forecasting framework with a forecast window of 9 months to forecast the water balance components at the sub-watershed scale using the calibrated SWAT model initialized with climate drivers from CFSv2 at a daily time step.

The purpose of this study is to develop models to represents the sub-watershed scale processes at a high resolution (HUC-12), which can be used under a dynamic setting to develop water resources scenarios for the routine assessment of water availability with relative ease over sub-continental scales.

2. SWAT Model

Model Background

The Soil and Water Assessment Tool (SWAT) is used in this study to simulate the hydrologic cycle at a sub-watershed scale. SWAT operates in a continuous-time mode, and is a process-based, semi-distributed, comprehensive water balance model [61,62,63]. Simulations from the SWAT model incorporates weather, hydrology, sedimentation, soil temperature, plant growth, nutrients, pesticides, and land management [64]. The model uses precipitation, radiation, and temperature, along with elevation, soil, and land cover/land use data as inputs for simulating surface and sub-surface hydrological processes. The unique combinations of soil type and land use (and slope if desired) are lumped as a hydrologic response unit (HRU) that dictates the physical response of the watershed to climate drivers. Thus, each sub-watershed is divided into several HRUs and the land component of the hydrologic cycle is modeled for each HRU. Surface runoff is estimated in the model using the Soil Conservation Service Curve Number (SCS-CN) method [65].

A two-layer structure represents the aquifer system in SWAT, where the first layer represents the shallow aquifer and can interact with the surface water or add to the groundwater flow, whereas the water in the deeper layer is considered to be out of the system. These fluxes are calculated as a lumped average value for the entire HRU. The surface runoff is routed to the outlet to the watershed as streamflow. SWAT uses a temperature dependent equation like the Penman-Monteith equation for estimating the potential evapotranspiration, while plant growth is modeled using a modified Environmental Policy Integrated Climate (EPIC) model to estimate actual plant transpiration [66]. For more details on the SWAT model, readers are further referred to the SWAT literature database for literature on SWAT for various hydrological modeling applications [67].

3. Study Area and Model Dataset

3.1. Study Area

This study evaluates the hydrology of the South Atlantic-Gulf (SAG) region in the southeastern U.S., as shown in Figure 1. The geographic area of the SAG region spans an area of 724,326 square kilometers and encompasses Florida (FL) and South Carolina (SC), and parts of Georgia (GA), Alabama (AL), Louisiana (LA), Mississippi (MI), North Carolina (NC), Tennessee (TN), and Virginia (VA) [68]. The western part of the SAG region is bounded by Appalachia and is predominantly forested with low development and urbanization, high elevation, and silty soil. The central and southern parts of the SEUS are dominated by woody wetlands, open spaces, and cultivated crops. The southern tip of Florida is dotted with water bodies and wetlands, with some areas of cultivated crops and developed spaces. The soil in the coastal plains has a high percentage of sand and low percentages of clay and silt (Figure 2).

The SAG region is identified using a two-digit number “03” in the HUC system. This region is divided into 18 basins, each distinguished using a four-digit HUC identifier ranging from 0301–0318. These HUC-4 basins are further divided into 50 watersheds, and the SWAT model is used to model the hydrology for each watershed at a HUC-12 resolution. HUC-12 is the 12-digit identifier for the sub-watersheds. Out of a total of 7453 HUC-12 sub-watersheds provided by the National Hydrography Database plus (NHD+) in the study area, the SWAT model delineated 7391 sub-watersheds for the entire SAG region, thus ensuring high spatial resolution of the simulations.

Table 1. CFSv2 monthly ensembles with respective initial day of the month for four members initialized at 00Z, 06Z, 12Z, and 18Z [69].

Initialization Month	Total Members	Initialization Days
January	28	1, 6, 11, 16, 21, 26, 31
February	20	5, 10, 15, 20, 25
March	24	2, 7, 12, 17, 22, 27
April	24	1, 6, 11, 16, 21, 26
May	28	1, 6, 11, 16, 21, 26, 31
June	24	5, 10, 15, 20, 25, 30
July	24	5, 10, 15, 20, 25, 30
August	24	4, 9, 14, 19, 24, 29
September	24	3, 8, 13, 18, 23, 28
October	24	3, 8, 13, 18, 23, 28
November	24	2, 7, 12, 17, 22, 27
December	24	2, 7, 12, 17, 22, 27

provides the list of 18 basins in the SAG region; each is referred to by its HUC-4 identifier and the name of the sub-watersheds with the outlet matching with the U.S. Geological Survey (USGS) streamflow-gauging stations. Note that the authors provide the watershed numbers only for ease of reference and the identification of these watersheds in the study, and they are not to be confused with the watersheds at HUC-6, which are larger than these watersheds in most cases.

3.2. Model Dataset

The basic drivers for the SWAT model were the U.S. Geological Survey (USGS)-derived Digital Elevation Model (DEM), State Soil Geographic (STATSGO) soil data, and land use/land cover data from the National Land Cover Data (NLCD) 2011 [70]. Weather data (precipitation and temperature) is obtained from the National Centers for Environmental Prediction (NCEP) Climate Forecast System Reanalysis (CFSR) data [56,71,72]. The reanalysis data is available for a period of 36 years from 1979 to 2013 and can be accessed through the global weather database for SWAT (URL: https://globalweather.tamu.edu). Studies have previously evaluated the performance of the reanalysis product and have reported satisfactory performance in watershed-scale modeling using the dataset [72,73,74,75]. One USGS station for each watershed was chosen for calibrating and validating the streamflow generated using the SWAT model (listed in Table 1). The land use, soil, and elevation data used in the models are shown in Figure 2.

The CFSR is extended as an operational, real-time product that forecasts the atmospheric variables with a lead time of nine months and was initialized multiple times a day, and on different days of the month (

Table 2. List of the 18 basins in the South Atlantic-Gulf region identified with a four-digit Hydrologic Unit Code and a six-digit identification number (not to be confused with HUC-6) for the constituent watersheds along with their respective outlet USGS stations (reproduced with permission from Sehgal and Sridhar [59]).

Basin	USGS Station	Watershed	River/Basin	Basin	USGS Station	Watershed	River/Basin
0301	02066000	030101	Roanoke	0309	02292900	030902	Caloosahatchee
	02075500	030102	Dan	0310	02296750	031001	Peace
	02047000	030103	Nottoway		02301500	031002	Alafia
0302	02084000	030201	Tar		02313000	031003	Withlacoochee
	02091814	030202	Neuse	0311	02323500	031101	Suwannee
0303	02134500	030301	Lumber		02317500	031102	Alapaha
	02106500	030302	Black	0312	02330150	031201	Ochlockonee
	02108000	030303	NE Cape fear	0313	02338000	031301	Chattahoochee
0304	02129000	030401	Pee Dee		02350512	031302	Flint
	02134500	030402	Lumber		02358700	031303	Apalachicola
	02132000	030403	Lynches	0314	02369600	031401	Yellow
0305	02147020	030501	Catawba		02366500	031402	Choctawhatchee
	02169500	030502	Congaree		02374250	031403	Conecuh
	02175000	030503	Edisto	0315	02397000	031501	Coosa (Rome)
0306	02197000	030601	Savannah		02407000	031502	Coosa (Childersburg)
	02198000	030602	Barrier Creek		02428400	031503	Alabama
	02202500	030603	Ogeechee	0316	02448500	031601	Noxubee
	02203000	030604	Canoochee		02466030	031602	Black warrior
0307	02223500	030701	Oconee		02469761	031603	Tombigbee
	02215000	030702	Ocmulgee	0317	02478500	031701	Chickasawhay
	02225500	030703	Ohoopee		02474500	031702	Tallahala
	02228000	030704	Satilla		02479300	031703	Red Cr.
0308	02234000	030801	St. Johns (Geneva)	0318	02482550	031801	Pearl (Carthage)
	02244040	030802	St. Johns (Buffalo bluff)		02488500	031802	Pearl (Monticello)
0309	02270500	030901	Arbuckle Cr.		02489500	031803	Pearl (Bogalusa)

). Precipitation and temperature time series were obtained from CFSv2 data from 1 April 2011, through to 19 December 2017 by using the first instance of the CFSv2 model dataset from 1 April 2011, through to 12 March 2017, and a 9-month forecast data was then obtained from the most recent initialization (12 March 2017 in this study). Figure 3 provides a schematic of data assimilation for the study. The red ellipse highlights the model run for each initialization, which was used for SWAT model implementation, with a total number of K initializations from 1 April through to 12 March 2017.

4. Methods

4.1. Modeling Framework

The SWAT model was calibrated from 2003 through to 2010 with a three-year warm-up period (2000 through to 2002) and was validated from 2011 through to 2013. The parameterization of the models was carried out in ArcSWAT 2010 (https://swat.tamu.edu/software/arcswat/). The sequential uncertainty fitting-2 (SUFI-2) algorithm [76,77,78] was employed to calibrate the model parameters using the SWAT-CUP (calibration/uncertainty or sensitivity program) interface. The SUFI-2 algorithm was used based on its ability to provide the best fitness while accounting for the high degree of uncertainty between the simulated and observed dataset [79]. The parallel processing routine [80] available with SUFI-2 was used while simulating 1000 iterations for each watershed. Once the calibrated parameters were obtained from SWAT-CUP, a MATLAB-based script was written to update SWAT input-output files and run the SWAT model with calibrated parameters for the validation period and re-run the SWAT model for the period of January 1982 through to December 2013 with a three-year warm-up period (1979 through to 1981) to obtain the long-term hydrologic variables for this study.

4.2. Calibration of SWAT Model Using the Sequential Uncertainty Fitting (SUFI)-2 Algorithm

A set of 24 parameters were chosen for calibration to address various hydrologic components of the watershed, such as surface runoff (curve number, soil and plant evaporation, surface runoff and Manning’s coefficient, and available soil water capacity, etc.), baseflow (groundwater “revap”, aquifer–soil interaction, depth of water in the shallow aquifer, time for water leaving the root zone to reach the shallow aquifer, and deep aquifer percolation), storm water (channel hydraulic conductivity, stormflow lag time, etc.), snow (snowmelt rate, snow temperature, etc.), etc. Nash Sutcliffe Efficiency (NSE) [81] was chosen as the performance criteria for the models. The choice and effect of these parameters are very well documented in the SWAT literature [62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87], and the selected parameters are consistent with the aforementioned studies. The same set of selected 24 parameters were calibrated for each watershed as a measure to achieve reasonable performance for watershed-scale streamflow simulation and a broad representation of the geographical and hydrological variability of the region. These parameters were carefully selected to address both wetland/riparian watershed hydrology [65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89] and inland basin characteristics, with special regard to the heterogeneity in elevation, vegetation, and soils of the study region.

A list and a brief description of the 24 parameters selected for this study is provided in

Table 3. Selected parameters for calibrating the SWAT Model using Sequential Uncertainty Fitting 2 (SUFI-2) algorithm for the daily discharge dataset from 1 January 2003 to 31 December 2010. The minimum and maximum parameter range for the SUFI-2 optimization is provided in the table. The calibrated values are shown in Figure 4.

Class	Name	Description	Method Used	Min.	Max.
HRU	ESCO	Soil evaporation compensation factor (unitless)	Replace (v)	0	1
	EPCO	Plant evaporation compensation factor (unitless)	Relative (r)	−0.2	0.2
	HRU_SLP	Average slope steepness (mm/mm)	Relative (r)	−0.2	0.2
	OV_N	Manning’s “n” value for overland flow	Replace (v)	0.01	30
	SLSUBBSN	Average slope length (m)	Relative (r)	−0.2	0.2
MGT	CN2	Initial SCS curve number for moisture condition II (unitless)	Relative (r)	−0.2	0.2
	BIOMIX	Biological mixing efficiency (unitless)	Relative (r)	−0.2	0.2
SOL	SOL_AWC	Available water capacity of the soil layer (mm⁄mm)	Relative (r)	−0.2	0.2
	SOL_BD	Moist bulk density (mg/m3 or g/cm3)	Replace (v)	0.9	2.5
BSN	SMTMP	Snowmelt base temperature (°C)	Relative (r)	−0.2	0.2
	SMFMN	Melt factor for snow on Dec 21 (mm H2O/°C·day)	Relative (r)	−0.2	0.2
	SMFMX	Melt factor for snow on June 21 (mm H2O/°C·day)	Relative (r)	−0.2	0.2
	MSK_CO1	Calibration coefficient used to control the impact of storage time constant (Km) for normal flow when Km is calculated for the reach	Relative (r)	−0.2	0.2
	MSK_CO2	Same as above, but for low flows	Relative (r)	−0.2	0.2
	SURLAG	Surface runoff lag coefficient	Replace (v)	0.05	24
GW	GWQMN	Threshold depth of water in the shallow aquifer required for return flow to occur (mm)	Replace (v)	0	5000
	GW_REVAP	Groundwater revaporation coefficient (unitless)	Replace (v)	0.02	0.2
	GW_DELAY	Groundwater delay (days)	Replace (v)	0	500
	ALPHA_BF	Base flow alpha factor (days)	Replace (v)	0	1
	RCHRG_DP	Groundwater recharge to deep aquifer (fraction)	Relative (r)	−0.2	0.2
	REVAPMN	Threshold depth of water in the shallow aquifer for revaporation to occur (mm)	Relative (r)	−0.2	0.2
RTE	CH_N2	Manning coefficient for channel (unitless)	Replace (v)	−0.01	0.3
	CH_K2	Effective channel hydraulic conductivity (mm⁄h)	Replace (v)	−0.01	500
	ALPHA_BNK	Baseflow alpha factor for storage (days)	Replace (v)	0	1

. In order to provide realistic estimates of the parameter space for each watershed, the initial parameters were fine-tuned before implementing calibration in the interest of computational ease and time. The minimum and maximum values provided in Table 3 provided the parameter range used for the search of the optimum value of the respective parameter in calibration. Since the total number of watersheds used in this study made it difficult to provide the calibrated values in the form of a table, Figure 4 provides the spatial map of the calibrated values of the parameters used in this study.

To highlight the influence of the land use characteristics on the estimation of the model parameters, a comparison is made between some of the selected model parameters, namely CN2, BIOMIX, HRU_SLP, and ESCO (for brevity), with the percentage of land use under five classes, namely developed land, forest, cultivated, grass/shrub/hay, and wetlands, at a watershed scale in Figure 5. It can be observed that there was a strong influence of land use characteristics on the model parameters. A higher HRU SLP corresponded to the higher percentage of forest areas, as much of the Appalachian mountain region is heavily forested. A higher value of BIOMIX corresponded to the higher percentage of forest or grasslands and a low amount of development in watersheds. High ESCO values corresponded to forested watersheds and a higher percentage of wetland areas, especially in the coastal watersheds. CN2 was impacted by a combination of factors, including a higher percentage of developed or cultivated areas or a higher percentage of wetlands. In some cases, CN2 also corresponded to a high forest cover percentage. However, this can be explained by the mountainous topography of these watersheds. It was observed that while the values of these parameters varied for each watershed (Figure 4), the overall performance of the models for streamflow simulation during calibration and validation periods at all 50 outlets was satisfactory with the given set of model parameters. The strong influence of land use, topography, soil type, and other watershed scale characteristics could be seen on the spatial variability of these parameters. For instance, predominantly forested areas showed higher values for BIOMIX due to the higher redistribution of soil constituents from biota (for example earthworms, etc.) compared to wetlands or developed watersheds; HRU_SLP values were higher for mountainous regions compared to plains and coastal areas; SLSUBBSN was found to be higher for coastal watersheds due to the low gradient. OV_N showed variability based on the surface characteristics of the overland flow and varied from basin to basin. Other important parameters, such as CN2, SOL AWC, ALPHA_BF, and EPCO, show significant spatial variability. A combination of these factors determined the response of a watershed to precipitation and the resultant discharge and storage. For example, low ALPHA_BF numbers signified a slow response to changing moisture conditions in the baseflow under no recharge to the watershed, whereas higher CN2 numbers signified higher runoff for a given precipitation amount. Interaction with groundwater (GW_DELAY, GW_QMN, and GW_REVAP) and runoff play a major role in the partitioning of precipitation in the watershed into runoff, soil, and groundwater storage. A slow surface-aquifer interaction, coupled with a high curve number, led to high values of discharge.

4.3. Performance Evaluation of the SWAT Model

The SWAT-simulated streamflow is compared with the observed discharge data from the selected USGS gauging stations for the purpose of performance assessment. Three statistical performance indices, namely the coefficient of determination (R²), the normalized root mean square error (NRMSE), and Willmott’s Index (WI) [90,91], were used for quantitative assessment of the results. The NRMSE were normalized values of the RMSE to a range of [0, 1] to facilitate comparison of the model performance with other watersheds. For all selected indices, values closer to 1 indicate better performance. Latest studies have proposed spatial-pattern-oriented model calibration and performance indices (based on simulating the spatiotemporal variability of the dataset) [92,93] based on satellite observations and/or advanced measurement and monitoring capabilities in the study region [94]. However, a lack of such a dataset at the spatial and temporal resolution of this study restricted the validation of the dataset to sub-watershed scale validation of observed streamflow from USGS, and validation of the seasonal and sub-seasonal variability of the simulated soil moisture using in-situ stations.

4.4. Hydrologic Simulations Using the CFSv2-Integrated Calibrated SWAT Model

The calibrated SWAT model was integrated with CFSv2-derived variables from April 2011 through to December 2017 with a warm-up period of 3 years, thus providing simulation outputs from January 2014 through to December 2017. Seamless daily simulation outputs were obtained from January 1982 through to December 2017 by combining the retrospective model simulations from the CFSR-integrated SWAT model and the CFSv2-integrated SWAT model. SWAT-CFSv2 model outputs (soil water, evapotranspiration, potential evapotranspiration, surface runoff, watershed precipitation, etc.) for the period of April 2011 through to March 2017 were obtained in near real-time mode since the input meteorological data was taken from the first CFSv2 model runs, whereas SWAT-CFSv2 output data from March through to December 2017 were the forecasted model outputs.

4.5. Water Balance Analysis

This study follows Sridhar and Hubbard [95] and Singh et al. [96] in studying the water balance of the study region. The fundamental water balance equation can be expressed as:

$$\frac{\partial S}{\partial t}=PREC+I-ET-R-D$$

(1)

where S is the soil water in the root zone (mm); t is the time (day); PREC is precipitation (mm/day); I is the irrigation (mm/day); ET is the actual evapotranspiration (mm/day); R is runoff (mm/day); and D is the drainage below the root zone (mm/day).

The daily estimates of water balance components for each watershed were added to the monthly totals and subsequently used to estimate the monthly water balance components for each watershed. The change in soil water storage $∆SW$ was calculated by subtracting the storage of the previous month from the current month’s (t) storage.

$$∆SW=S{W}_t-S{W}_{t-1}$$

(2)

The actual ET was computed for each month using the change in precipitation and soil water storage

$$\begin{gathered} \mathrm{ET}=\mathrm{PREC}-∆SW,\mathrm{if}∆SW\le 0 \\ \mathrm{ET}=\mathrm{PREC}+∆SW,\mathrm{if}∆SW>0 \\ \mathrm{ET}=\mathrm{PET},\hspace{1em}\mathrm{if}∆SW>0 \end{gathered}$$

(3)

The difference between potential ET (PET) and actual ET, was used to calculate the water deficit. Surplus water was estimated by subtracting $ET$ from the sum of the soil storage and precipitation.

$$\begin{gathered} \mathrm{Deficit}=\mathrm{PET}-\mathrm{ET} \\ \mathrm{Surplus}=\mathrm{PREC}-(\mathrm{ET}-∆SW) \end{gathered}$$

(4)

5. Results and Discussion

5.1. Streamflow Validation

A spatial representation of the SWAT models in simulating stream discharge is provided in Figure 6. It can be seen that while the performance of the SWAT models can be deemed satisfactory for most inland gauging stations, the performance of the models was weak in the coastal marshlands, especially in Florida. It is to be noted that at these gauging locations, the observed streamflow discharge was heavily influenced by tides. These tidal activities added a secondary oscillatory component in the streamflow records. These tidal cycles corresponded to a period of 24.84 h and successive averaging in daily computations of the recorded data by USGS led to a secondary oscillatory component in the data with a time period of about two weeks. These oscillations, which were unrelated to the watershed hydrology, were often missed by the SWAT models, leading to a high degree of discrepancy between the observed and the simulated streamflow. Furthermore, extensive basin management and regulation practices could also lead to poor simulations by the SWAT models as water management practices were not explicitly included in the SWAT model setup in this study. Additionally, it was observed that the performance of the SWAT model was better for larger, downstream watersheds due to reduced intermittency and sustained flow regimes. A combination of these factors caused the deviation of SWAT-simulated values from the observed dataset. For a vast majority of watersheds (approximately 30 inland watersheds), the models showed R² = [0.45, 0.68], NRMSE = [0.07, 0.1], and WI = [0.45, 0.68] for the calibration and validation datasets at daily time step, respectively. The performance of the models for the rest of the SAG region was assessed to be satisfactory based on the findings of several other studies conducted in various parts of the SEUS, such as Virginia [58], the coastal plain of southwestern Georgia [97], low-gradient watersheds in South Carolina [98], etc.

5.2. Validation of Soil Moisture Data

Given the importance of long-term high-resolution soil moisture (SM) simulations for the study region, a comparison was made between the SWAT simulated SM with nine in-situ stations located within the study area. These soil moisture stations belong to the U.S. Climate Reference network (USCRN), which provide observations for five standard depths, i.e., 5 cm, 10 cm, 20 cm, 50 cm, and 100 cm. Weekly averaged values of the observed SM from USCRN stations were compared with the corresponding values of the area averaged, sub-watershed scale SM from the total rooting depth of the SWAT models after suitable conversion to volumetric units. Figure 1 provides the geographic locations of the USCRN stations selected for this study. The SM time series for five soil sensor depths for one of the USCRN stations used in the study area are provided in the supplementary material. In order to capture the general dynamics of the SM across the total rooting depth, this study used in-situ soil moisture data from three sensor depths, i.e., 10 cm, 20 cm, and 50 cm, for simultaneous comparison with area-averaged total rooting depth soil moisture simulations from the SWAT models for the respective sub-watershed. For ease of comparison, a suitable conversion was applied to the SWAT simulated soil moisture to obtain the volumetric SM content using the total rooting depth of the respective sub-watershed.

Figure 7 provides a comparison of the in-situ and simulated SM values for three sensor depths. The days with the missing in-situ data were ignored in the analysis. Strong seasonality was evident from the plots, where SM values could be observed to decline in the summer months and recovered in the winter months.

Table 4. Performance evaluation of the sub-watershed (HUC-12) scale SWAT-simulated soil moisture (in volumetric units) using in-situ soil moisture observations for three sensor depths (10 cm, 20 cm, and 50 cm) at nine locations.

	Correlation			RMSE (m3/m3)
Station location (Lat, Lon)	10 cm	20 cm	50 cm	10 cm	20 cm	50 cm
Durham, NC (35.97, −79.09)	0.658	0.681	0.749	0.227	0.224	0.372
Blackville, SC (33.36, −81.33)	0.324	0.345	0.279	0.131	0.125	0.020
McClellanville, SC (33.15, −79.36)	0.550	0.543	0.502	0.080	0.017	0.076
Watkinsville, GA (33.78, −83.39)	0.581	0.602	0.643	0.043	0.012	0.078
Sebring, FL (27.15, −81.37)	0.424	0.436	0.533	0.081	0.029	0.100
Newton (W), GA (31.31, −84.47)	0.552	0.564	0.481	0.029	0.017	0.033
Newton (SW), GA (31.19, −84.45)	0.579	0.594	0.628	0.051	0.007	0.098
Selma, AL (32.33, −86.98)	0.726	0.673	0.339	0.290	0.029	0.045
Newton, MS (32.34, −89.07)	0.582	0.348	0.265	0.107	0.099	0.095

provides the comparison between the SWAT simulated SM with the in-situ measurements using correlation and RMSE as performance indices. It can be observed that the performance of the SM simulation was satisfactorily high except for the coastal watersheds (in Florida and another in South Carolina) for reasons highlighted in Section 5.1. While for most locations, the general inter-seasonal dynamics was captured by the SWAT simulated SM, there could be a systematic bias in the simulations or errors in capturing the extreme high and low values. Despite relatively high values of RMSE in some stations due to systematic bias in the simulations, high values of correlation between the in-situ and simulated SM provided the opportunity to use the simulated SM in applications where the general trend of the dataset and deviation from the retrospective values were of importance. This made SM simulations from SWAT useful for retrospective drought analysis and/or real-time monitoring, as used by Sehgal and Sridhar [59].

However, there were two important considerations in interpreting the comparison of model and in-situ SM data: (i) difference in the scale of the SWAT simulations and in-situ observations, and (ii) difficulty in determination of the representative sensor depth for capturing the rooting depth soil moisture dynamics. While point-scale measurements typically have a support scale of a few centimeters, the SWAT models provided a lumped estimate of SM simulations at the HRU scale. This difference in scale between the in-situ and SWAT simulated SM may have added systematic bias in the two datasets. Furthermore, moisture distribution within a soil profile is a non-linear and complex process with a strong influence of a combination of factors like local topography, vegetation, soil composition, soil hydraulic properties, etc. This makes it difficult to assign a representative soil depth to capture the root-zone soil moisture dynamics using the in-situ measurements.

5.3. Near Real-Time Daily Simulations Using the CFSv2-Integrated SWAT Model

To evaluate the performance of the CFSv2-integrated SWAT model using near real-time meteorological data from CFSv2 model outputs, the time series plots of the observed and simulated discharges for the Roanoke, Pearl, Satilla, and Coosa Rivers are shown in Figure 8. The R² values of the simulation are provided in the figure. It can be seen that the performance of the models for these watersheds was similar to the calibration and validation results, as discussed in the previous sections. The 9-month forecasts of simulated discharge were compared with the observed discharge values for the four watersheds, which are provided in Figure 9. It can be observed from Figure 9 that while the near real-time simulation using SWAT-CFSv2 models provided a realistic estimate of the discharge, the simulations using the forecasted meteorological dataset may not be considered to be reliable. The observed and forecasted river discharge values could be observed to be within a close range only up to a lead of approximately 1 month, beyond which the simulated and observed discharge show a high degree of disagreement. This issue could be countered by updating the SWAT model with every new initialization of CFSv2 (or a few times each month) in order to update the model state with the latest forcing dataset for reliable forecasts of hydrological variables.

5.4. Water Budget Analysis for the SAG Region

Figure 10 shows the monthly water budget for the 50 watersheds in the study summarized in the form of a box plot to capture the mean, median, and variability within individual components. Most watersheds in the study region received precipitation during the summer (June–August). With high temperatures and available moisture, precipitation in the summer rose to its peak value. Soil water remained relatively consistent in the region, with a minor dip during summer months. High temperatures during April through to August led to higher PET compared to ET, thus causing water-deficit conditions. Water-surplus conditions were attained with incoming precipitation in summer months and the subsequent fall in temperature in late spring and fall seasons. Figure 11 provides water budget plots for two select watersheds, namely the Roanoke and Pearl River watersheds. Differences in the water budget for the two watersheds is evident from Figure 11 and was primarily driven by the climatology of these watersheds. The Roanoke watershed received a total monthly precipitation of approximately 100 mm throughout the year. Increasing temperatures in the summer led to a fall in soil water storage. However, low temperatures and reduced plant transpiration in the late fall and winters helped bring soil water back to values above 150 mm. The Pearl River falls received the highest rainfall in summer months (June–August) and was often affected by the tropical cyclones. Soil water for the watershed consistently remained higher than 150 mm. High summer temperatures and moisture availability fueled high evapotranspiration in the watershed, especially during summers.

5.5. Hydrological Forecasts for South Atlantic-Gulf Region Using the CFSv2-Integrated SWAT Model

The effect of the drought in the region can be observed in Figure 12 for the year 2016, where a high PET and low soil moisture and precipitation led to water deficit conditions and hence drought in the SEUS. Dry conditions affected large parts of the study region, predominantly Alabama, Georgia, Florida, and many parts of North and South Carolina, leading to severe drought conditions in the affected regions. A similar condition was expected to be observed in July and August 2017, with relief due to forecasted precipitation in the region in the later part of the year. Spatial maps of the forecasted values of PREC, PET, ET, and soil water storage (SW) are provided in Figure 12. The spatial variability in the hydrologic variables is evident from Figure 12, which is strongly influenced by seasonality and hydroclimatology. Summer temperatures led to high PET and ET and low soil moisture in the region from May through to August. High spatial variability can be seen in the case of precipitation, where parts of Louisiana, Alabama, and Georgia were expected to receive higher precipitation during April and May, whereas many parts of Florida and the Carolinas receive precipitation during July through November. For interested readers, the observed total precipitation for the study region is provided from April–December 2017 in Figure S8.1 of the Supplementary Material.

5.6. Uncertainty in Model Outputs

Data, model, and parametric uncertainty are three main causes of errors in hydrological models. Model uncertainties can be a result of measurement errors associated with the model inputs (DEM, land use classification), from the model structure and its assumptions and simplification, and from approximations in determining parameters [99,100]. Parameter uncertainty, among other sources, is highly likely because of lack of availability of the measured parameters (e.g., soil hydraulic properties and basin characteristics, etc.). Calibrating models for the best parameters may also have issues such as equifinality [101] (different parameter sets resulting in similar model results), due to interactions and interrelationships between the parameters. This is specifically true for heavily managed watersheds in some parts of the study region where the model parameters adjust for anthropogenic influence on the watershed without any explicit inclusion of management practices in the design or implementation of the hydrologic model. The issue of equifinality and parametric uncertainty in the SWAT model is well discussed in the literature [102,103,104].

The performance of the SWAT-CFSv2 simulations also depend heavily on the forecasting ability of the CFSv2 weather data for the region of analysis. Several other studies have highlighted the issue of uncertainty in CFSv2 models and their influence on the simulation of various hydrologic models. Zhang et al. [55] tested the efficiency of VIC models initialized with satellite-aided monitoring (MONIT), CFSv2, and ensemble streamflow prediction (ESP), and evaluated the performance of the soil moisture forecast skill and observed that the uncertainty in the soil moisture forecast is mostly controlled by initial conditions in the first month and that uncertainties in the CFSv2 climate forecasts have the largest contribution to SM forecast errors at longer lead times. Furthermore, seasonality hydroclimatological variations have a significant impact on the modeling performance of CFSv2 models [105,106,107]. In other words, the hydrological predictions using SWAT-CFSv2 may be more skillful in some seasons compared to others due to the combined influence of teleconnections with large-scale climatic drivers, which may lead to a more predictable influence on the water budget of the study area (hurricanes and high precipitation or increased temperatures and reduced precipitation, etc.).

6. Conclusions

This study provides an important platform for generating high-resolution hydrologic simulations for the southeastern U.S. at high spatial (HUC-12) and temporal (daily) resolution, which can be of importance to water planners and policymakers of the region. The models were calibrated using the SUFI-2 algorithm module in SWAT-CUP for the period of January 2001–December 2010 and validated through to December 2013 at a daily time step using observed streamflow datasets from USGS.

From the calibrated model, we generated the water budget components for 50 watersheds in the South Atlantic-Gulf region of the southeastern US from January 1982–December 2013 using the CFSR meteorological dataset. In addition, we provided a framework for near real-time forecasting of hydrologic variables in the study area by integrating the SWAT model with meteorological drivers from CFSv2 climate forecast models with a lead time of 9 months. The CFSv2-integrated SWAT model estimated soil moisture and other hydrologic outputs corresponding to the near real-time CFSv2 model simulation output from January 2014 through to March 2017 and forecasted values for 9 months through to December 2017.

The results have been carefully evaluated by several quantitative and subjective approaches. Observed streamflow and in-situ soil moisture datasets were used to provide a statistical estimation of the accuracy of the model simulations, whereas the influence of land use on calibrated parameters, the complementary relationship between ET and PET [108,109], and the water balance of the study region were studied in order to gain confidence in the model simulations. The results indicated that the models were effective in capturing the hydrologic variability in the implemented watershed in terms of high accuracy in simulating streamflow discharge for 50 outlet points and a satisfactory correlation with the in-situ soil moisture dataset from nine stations across the region for the total rooting depth soil profile. The near real-time simulations using the CFSv2-integrated SWAT model were found to be satisfactory. However, the hydrological forecasts made using the forecasted meteorological dataset from CFSv2 needs improvement for longer lead time periods. The hydrological forecasts from the CFSv2-integrated SWAT model were found to be reasonable for a lead time of approximately 1 month, beyond which the uncertainty in the models render the hydrological forecast unreliable. However, regularly updating the integrated SWAT-CFSv2 models on a weekly/bi-weekly basis could help solve this shortcoming. Insights into the watershed-scale hydrological processes over the vast regions of the SEUS are useful for water planners and policymakers of the region, which is increasingly prone to water scarcity in the wake of high (and growing) agricultural and industrial water demand. As a follow-on study, the simulated dataset will be subsequently evaluated for varied applications such as drought monitoring and forecasting, flood/forest fire risk assessments, and other applications related to agricultural and water resources.