Hydrological models are developed and used to support policies at the watershed and regional scales (Brauman et al., 2022;Aryal et al., 2022;Mankar et al., 2020). These models are designed using a variety of mathematical equations (e.g., continuity equation, Darcy's law, and Manning's equation) to simulate the complex interactions of hydrological processes (Grimaldi et al., 2021;Herman et al., 2020;Ahmed et al., 2020;Sharma et al., 2022;Yin et al., 2024). Policymakers use model results to evaluate the effectiveness of proposed policies, improve effectiveness of policies, and mitigate unintended side effects when implemented (Maviza and Ahmed, 2021;Mishra et al., 2023;Nguyen et al., 2022b,a;Pandi et al., 2021;Prabha and Tapas, 2020;Tapas et al., 2022a;Tran et al., 2021a,b). For example, the Tar-Pamlico Nutrient Strategy utilized model results to set a 30 % nitrogen reduction goal, incorporating policies such as nutrient management, buffer protection, and nutrient trading (NCDEQ, 2018). These models serve as valuable tools for showing how water quality and quantity may change within a watershed due to alterations in land cover and management practices (Brauman et al., 2022). However, the accuracy of these multi-parameterized models has been questioned for informed decision-making (Grimaldi et al., 2019;Grimaldi et al., 2021).
In recent decades, the rapid growth in computational power has enabled the development of more complex models that use a higher number of parameters to accurately represent water movement in a watershed (Adeyeri et al., 2020;Tapas et al., 2022c). However, there are often limited numbers of observations available to assist with model calibration (Do et al., 2024b), raising challenges for the accuracy of model predictions (Gupta and Govindaraju, 2019). This mismatch between observed data and multi-parameter simulations increases modeling uncertainty (Beven and Freer, 2001;Wellen et al., 2015). Another type of uncertainty in environmental system models comes from input datasets (Bosshard et al., 2013). These datasets include rainfall, land cover, elevation, and soil information (Chaubey et al., 2005;Cho et al., 2009). Uncertainty in the rainfall dataset comes from low spatiotemporal resolution and assumptions made to average rainfall over the spatial scale used in the model (Fraga et al., 2019;Junqueira et al., 2022).
In recent years, many studies have examined the use of satellite-based precipitation products (SPPs) as an alternative to in-situ data. For example, Sharifi et al. (2019), Tran et al. (2023b), and Le et al. (2020) evaluated the performance of common satellite-based precipitation products (SPPs) in Austria and Vietnam, respectively. Tran et al. (2023b) and Le et al. (2020), ( 2023) conducted experiments demonstrating that the Global Precipitation Measurement (GPM) Integrated Multi-satellite Retrievals for GPM (IMERG) (Hou et al., 2014) provided the best model performance for flow calibration. However, Le et al. (2023) mentioned that SPPs tend to underestimate the total actual amount of rainfall and have issues with rainfall estimates in coastal and arid regions. Thus, one of the main objectives of this study is to evaluate some commonly used precipitation datasets, namely the Fifth Generation of European ReAnalysis (ERA-5) (Muñoz-sabater et al., 2021), IMERG (Hou et al., 2014) and the Gridded Surface Meteorological (gridMET) (Abatzoglou, 2013) to optimize flow simulations in a coastal region.
The Soil and Water Assessment Tool (SWAT), developed by the United States Department of Agriculture (USDA) and Texas A&M AgriLife, is a comprehensive hydrological model designed to predict the impacts of land use, land management, and climate change on sediment transport (Betrie et al., 2011;Murumkar et al., 2024), water quality (Oeurng et al., 2016), and nutrient cycling (Kansara et al., 2021) in river basins. This model has been widely used in previous studies throughout the world: Mekong River Basin (Mohammed et al., 2018b(Mohammed et al., , 2018a;;Mondal et al., 2022), Vietnam River Basins (Le et al., 2020;Nguyen et al., 2024Nguyen et al., , 2022aNguyen et al., , 2022d;;Tran et al., 2023b;Nguyen et al., 2024), Nepal (Kumar et al., 2017), and United States (Jha et al., 2006;Sehgal et al., 2018;Tapas et al., 2022b,c;Tapas, 2024a;Lakshmi, 2024a). In recent years, the new version of SWAT, known as SWAT plus (SWAT+), has been released with new features regarding spatial process interactions and visualization (Bieger et al., 2017;Tran et al., 2023a;Wu et al., 2023;Tapas, 2024a). This version allows users to better model the transport and retention of nutrients and sediment in the watershed.
Hydrological models depend on simplifications of natural processes, making it extremely unlikely to obtain accurate simulations with initial model setup (Gupta and Govindaraju, 2019;Vieux, 2001). To address this issue and improve the correlation between simulated and observed flow, hydrological modelers perform calibration. Although manual calibration by adjusting individual parameters is possible, the high number of parameter combinations makes this approach impractical (Efstratiadis and Koutsoyiannis, 2010;Sunmin, 2021). Consequently, extensive study has been devoted to developing more accurate hydrological models using automatic optimization techniques (Getirana, 2010;Yen et al., 2019). Arsenault et al. ( 2014) compared ten stochastic calibration methods and determined that Dynamically Dimensioned Search (DDS) (Tolson and Shoemaker, 2007) outperforms most other techniques, especially with increased model complexity. The Generalized Likelihood Uncertainty Estimation (GLUE) is a widely utilized method among hydrological modelers due to its simplicity (Blasone et al., 2008). Latin Hypercube Sampling (LHS) (Stein, 1987) is another popular method in uncertainty analysis, as it prevents repeated sampling (Yue et al., 2023). However, there is still a lack of comprehensive understanding regarding the application of these calibration methods when using SPPs in the context of enhancing river flow simulations at the basin scale.
The performance of hydrological models can be assessed using visual interpretations or statistical indices (Jain and Sudheer, 2008;Wealands et al., 2005;Lakshmi, 2024b). Visual analysis of simulated streamflow can give useful insights into model behavior (Wealands et al., 2005) but needs expert analysis, and is impractical from an automatic calibration point of view (Jain and Sudheer, 2008). However, performance indices do not require expert opinions and can be optimized using automated calibration techniques (Idrissou et al., 2020;Jain and Sudheer, 2008;Ritter & Munoz-Carpena, 2013). Several performance indices are available to measure the agreement between the observed data and simulated data. Nash Sutcliffe Efficiency (NSE; Nash and Sutcliffe, 1970) is the most widely used performance index for many hydrological models (Tran et al., 2023a;Nguyen et al., 2023aNguyen et al., , 2023b)); however, it is advisable to assess hydrological models using additional performance indices along with NSE (Jain and Sudheer, 2008;Ritter & Munoz-Carpena, 2013;Nguyen et al., 2023c). Other popular performance indices with hydrological modelers are Kling Gupta Efficiency (KGE) (Gupta et al., 2009) and coefficient of determination (R 2 ) (Adeyeri et al., 2020;Idrissou et al., 2020;Jain and Sudheer, 2008). NSE, KGE, and R 2 are popular as they can be compared with their ideal value of one. All these indices have shortcomings and strengths. For example, NSE does not distinguish between different kinds of errors (Jain and Sudheer, 2008), KGE corrects bias factors from NSE (Gupta et al., 2009), and R 2 is not able to measure non-linear relationships among different parameters (Barrett, 1974;Waseem et al., 2017).
The optimal flow prediction and parameter set can vary based on the rainfall dataset and autocalibration techniques. However, there are few studies that provide a comprehensive examination of the relationship between these factors. This study aims to examine the combined effects of these factors on SWAT+ flow simulations over a 17-year period. We performed this analysis on two major watersheds of North Carolina, the Tar-Pamlico River Basin, and the Cape Fear River Basin. We used three different sources of rainfall data (ERA-5, gridMET, GPM IMERG), three automatic calibration methods (DDS, LHS, and GLUE), and three evaluation metrics (NSE, KGE, R 2 ).
The overarching objective of this study is to evaluate the effectiveness of different combinations of rainfall datasets and optimization techniques in improving flow simulations while assessing the variability in performance indices and parameter sensitivity across two distinct watersheds. The objectives of this study are to (1) explore which combinations of three rainfall datasets and three optimization techniques yield better results for flow simulations, (2) identify how performance indices vary based on the chosen rainfall dataset and optimization technique, and (3) compare parameter sensitivity and uncertainty analysis for the two watersheds and three rainfall datasets.
This study enhances our understanding of hydrological processes in the Cape Fear and Tar-Pamlico watersheds, offering insights that are applicable to watersheds worldwide with similar characteristics. By identifying the optimal combinations of rainfall datasets, calibration methods, and performance indices, our findings provide a robust framework for accurate hydrological predictions. This approach can be adapted to diverse regions globally, including those with varying land uses, complex rainfall patterns, and coastal influences. These insights give hydrological modelers a better framework to simulate river flows, leading to improved water management and smarter decision-making worldwide.
This study was conducted in the Cape Fear watershed and the Tar-Pamlico watershed in North Carolina (Fig. 1). Cape Fear is the largest river basin in the state (NCDEQ, 2005), while the Tar-Pamlico is the fourth-largest (NCDEQ, 2014). These watersheds are two of only four river basins located entirely within North Carolina (Fig. 1). The Cape Fear basin discharges into the Atlantic Ocean, while the Tar-Pamlico River flows into the Pamlico Sound. Both watersheds host diverse ecosystems with a variety of habitats (NCDEQ, 2005(NCDEQ, , 2014)). The Tar-Pamlico watershed contains a larger proportion of agricultural land and wetlands (Tran et al., 2024), whereas the Cape Fear watershed has a higher percentage of forested land, pastureland, and urban areas. Additionally, the Cape Fear watershed has a greater elevation range than the Tar-Pamlico watershed.
Furthermore, the climatic conditions in these watersheds vary, with average annual precipitation in the Cape Fear watershed ranging from 1200 to 1300 mm (Griffin et al., 2013), while the Tar-Pamlico watershed receives between 1300 and 1500 mm annually (Tran et al., 2024). The average temperatures in both watersheds are characterized by warm, humid summers and mild winters. The annual average temperature in the Cape Fear watershed is around 15.5 • C (Griffin et al., 2013), while the Tar-Pamlico watershed averages a slightly warmer temperature at approximately 16.5 • C (Tapas, 2024a(Tapas, , 2024b)). The major characteristics of both watersheds are summarized in Table 1. The Cape Fear watershed has steeper slopes due to its varied elevation, while the Tar-Pamlico River basin features more gentle slopes (Fig. 1). The Cape Fear River basin is characterized by a mix of sandy and clay-rich soils (Benedetti et al., 2006), while the Tar-Pamlico River basin primarily has sandy loam soils (Tang, 2010). The main rivers draining these watersheds are the Cape Fear and the Tar River, respectively (Bales, 2003).
SWAT is a semi-distributed hydrological model developed in the early 1990s and is one of the most widely used tools for water quality and quantity simulation (Arnold et al., 1998;Gassman et al., 2022;Guo and Su, 2019;Do et al., 2024a;Aryal et al., 2023). Technological advancements since the original model was developed have made more detailed hydrological modeling possible, and complexity has been added to the model (Bieger et al., 2017;Efstratiadis and Koutsoyiannis, 2010;Tumsa et al., 2022). SWAT+ is an advanced version of SWAT with respect to data management, data analysis, data visualization, and process interaction at a watershed and sub-watershed level (SWAT+++, 2020;Tapas et al., 2024bTapas et al., , 2024c)). SWAT+ divides the watershed into sub-watersheds, sub-watersheds into landscape units, and landscape units into HRUs. An HRU is the smallest spatial unit in the SWAT+ , i.e., there is at least one HRU per sub-basin. SWAT+ considers land cover, soil, and slope data for creating the HRUs (SWAT+++, 2020). SWAT+ is capable of implementing landscape units as well as flow and pollutant routing over the landscape (SWAT+++, 2020;Tapas et al., 2024d). Although the basic equations for calculating flow in SWAT+ , such as the water balance equation, Manning's equation for channel flow, and the SCS curve number method for runoff estimation, have not been changed, significant changes have been made in the code and the formatting of the input files (SWAT+++, 2020;White et al., 2022;Tapas, 2024a). SWAT+ is also more appealing for spatial and temporal data analysis and data visualization (Bieger et al., 2017). In this study, version 2.2.0 of SWAT+ (released Feb 2023) was used.
We obtained both watershed boundary shapefiles using the United States Geological Survey (USGS) StreamStats website (Ries et al., 2017). We used QSWAT+ (Dile et al., 2019) to create streams, sub-basins, and Hydrological Response Units (HRU) using the Digital Elevation Model (DEM), land cover, and soil files (Table 2). We simulated both watersheds with a two-year warm-up period to initiate and stabilize several hydrological processes (Oo et al., 2020). Before beginning the model optimization across different scenarios, we verified that the watershed outputs for the uncalibrated model were within reasonable limits (S1 Supplementary Material), including metrics such as plant yield, denitrification, leaf area index (LAI), and biomass (S1 Supplementary Material) for the Cape Fear and Tar-Pamlico watersheds (Fig. 1).
The exact calibration of the model for plant yield, denitrification, and different flow paths (surface runoff, lateral flow, and groundwater flow) will require further simulations, followed by fine-tuning parameter ranges and subsequent simulations for a soft calibration (SWAT+++, 2020). The main motivation of this study is to assess the generalizability and applicability of different configurations of rainfall datasets and autocalibration techniques on the SWAT+ model performance across different watersheds, highlighting overarching trends that can inform global watershed modeling efforts, especially in data-scarce regions.
2.3.2.1. ERA-5. ERA-5 is the latest climate reanalysis product from the European Centre for Medium-Range Weather Forecasts (ECMWF) and stands as their fifth-generation reanalysis (Allabakash and Lim, 2020). It provides hourly data encompassing atmospheric, land-surface, and sea-state parameters, complete with uncertainty estimates (C3S, 2017;Hersbach et al., 2020;Muñoz-sabater et al., 2021). The ERA-5 analysis is produced at hourly intervals using an advanced 4Dvar integration approach (Table 3). This dataset can be found in the Climate Data Store (CDS), structured on regular latitude-longitude grids with a resolution of 0.25 • × 0.25 • . For our study, we obtained the daily ERA-5 precipitation data by aggregating it from the hourly dataset. The ERA-5 data used in our study is available from the ECMWF website (https://cds.climate.copernicus.eu/).
gridMET is a 0.04 • resolution daily meteorological dataset providing climate data such as temperature, precipitation, wind, humidity, and radiation for the contiguous United States (CONUS). It combines the spatial characteristics of gridded climate data from the Parameter-elevation Regressions on Independent Slopes Model (PRISM) with the favorable temporal features (and additional variables) of the regional reanalysis of North American Land Data Assimilation System Phase 2 (NLDAS-2) through climatically aided interpolation (Abatzoglou, 2013;Abatzoglou and Ficklin, 2017). The gridMET data used in this study can be accessed from the Climatology Lab website (https://www.climatologylab.org/gridmet.html).
GPM IMERG is a multi-satellite precipitation dataset that provides global precipitation data with high temporal and spatial resolution (Hou et al., 2014). It is a product of NASA's GPM mission, which aims to provide accurate and comprehensive precipitation measurements for the entire globe (Li et al., 2021). The GPM IMERG dataset combines data from various sources, including passive microwave sensors, infrared sensors, and radar, to provide precipitation estimates with a spatial resolution of 0.1 • and a temporal resolution of 30 minutes (Hou et al., 2014). The GPM IMERG product used in this study can be obtained from the NASA website (https://gpm.nasa.gov/data).
To better understand the differences in the rainfall data from three satellite-based products-ERA5, gridMET, and GPM IMERG-we compared them to observed data from USGS stations in Greenville, NC (USGS ID: 02084000), spanning from May 2010 to December 2020, and Kelly, NC (USGS ID: 02105769), spanning from October 17, 2003, to December 31, 2019, based on the availability of observed data at these stations. The seasonal analysis was calculated by grouping the rainfall data by season (Winter, Spring, Summer, and Fall) and then computing the Pearson correlation coefficients-a statistical measure that ranges from -1-1, indicating the strength and direction of a linear relationship between two variables-between each rainfall dataset and the observed data for each season.
For calibration of the SWAT+ model, we used daily observed flow data from USGS. We used the USGS station at the Tar River in Greenville, NC (USGS station id: 02084000) for Tar-Pamlico watershed calibration. We used the USGS station at Kelly, NC (USGS station id: 02105769) for calibration of the model for the Cape Fear watershed (Fig. 2). Considering the missing observed flow data and available weather data period (S3 Supplementary Material), we simulated the Tar-Pamlico watershed from Jan 2001 to Sept 2016 and the Cape Fear watershed from Jan 2003 to Dec 2018 with two years as a warm-up period.
Calibration is an important step in hydrological modeling in which users try to match simulated data with observed data by changing multiple parameters in the model (Sunmin, 2021;Tolson and Shoemaker, 2007). We used LHS (Stein, 1987), DDS (Tolson and Shoemaker, 2007), and GLUE (Blasone et al., 2008) automatic calibration approaches. LHS is a powerful method of selecting parameter values for a given number of simulations. It divides n number of parameters in m equal ranges so that each region is targeted. This method tests the combination of parameters in fewer simulations relative to random sampling.
The DDS algorithm is a heuristic-artificial intelligence technique to find the best option amongst given optionsthe algorithm searches for a globally optimal solution at the start and narrows to the optimal local solution towards the end of user-specified iterations (Tolson and Shoemaker, 2007). For the DDS method, we performed parallel processing with 175 iterations on 12 threads with a total of 2100 simulations. GLUE is a simple and flexible algorithm widely used by environmental system modelers (Beven and Binley, 1992;Tolson and Shoemaker, 2007). In the GLUE method, numerous parameter combinations are randomly selected. Each combination is assigned a likelihoodrepresenting its chance of appearing in multiple model setsbased on the principle of equifinality, which acknowledges that multiple parameter sets can similarly well simulate system behavior despite the differences in the parameter values themselves. (Blasone et al., 2008;Mirzaei et al., 2015). For GLUE and LHS methods, we also used a parallel processing approach with a total of 2100 simulations.
Model performance indicators used in this study.
Metric equation
Note: Q obs is observed streamflow, Q sim is simulated streamflow, i is i th simulation, and Q is the mean value, and n is the total number of values.
The goodness-of-fit of model output compared to observed data is measured using performance indicators (Jain and Sudheer, 2008;Wealands et al., 2005). We used the three most widely used performance indicators: NSE (Nash and Sutcliffe, 1970), KGE (Gupta et al., 2009), and R 2 (Barrett, 1974;Ozer, 1985). All of them have an ideal value of one for comparison, and each measures different aspects of goodness-of-fit (Ritter and Muñoz-Carpena, 2013). NSE indicates the variation between observed and residual data variance (Nash and Sutcliffe, 1970). KGE is a comprehensive score considering the ratios of means and dispersion, along with the correlation of observed and simulated datasets (Gupta et al., 2009). The R 2 explains how well the simulated variable explains the observed variable's variation (Ozer, 1985). Most hydrological models use NSE for testing model performance (Jain and Sudheer, 2008); however, NSE can be biased depending on data repetition, magnitude, and the number of measurements (Jain and Sudheer, 2008;Ritter and Muñoz-Carpena, 2013). Thus, it is advised to assess the model goodness-of-fit using multiple performance indices (Jain and Sudheer, 2008). In this study, we compared model results using three performance indices. We optimized the model for NSE and found the maximum value from all 2100 simulations using each autocalibration technique. The formulae for calculating NSE, KGE, and R 2 are shown in Table 4.
Over the last decade, there has been a significant development in autocalibration tools and platforms for hydrological models. The current SWAT+ calibration tools are SWAT+ Toolbox (Celray, 2022;Gassman et al., 2007), SWATplusR (Schuerz, 2022), SWATplus-CUP (WWEE., 2023), R-SWAT (Nguyen et al., 2022), and IPEAT+ (Yen et al., 2019). SWAT+ toolbox is coded in the C language and is not available for the Linux and Mac operating systems. There is no explicit SWAT+ toolbox user portal functional for researchers to discuss modeling challenges. SWATplusR is a package in R that allows users to identify reasonable parameter ranges along with parallel processing. SWATplusR currently does not have a graphical user interface and could be challenging to use for those with no experience using R software. SWATplusCUP has a graphical user interface and can be used by beginners. SWATplusCUP premium features, such as parallel processing, are not open source. IPEAT+ in the current version does not provide uncertainty analysis or a graphical user interface (Nguyen et al., 2022;Yen et al., 2019).
R-SWAT has a graphical user interface within R and provides an open-source parallel processing option for SWAT+ . R-SWAT was developed using the Shiny-R-package (Nguyen et al., 2022) and has a user group for researchers to discuss questions. These advantages are why R-SWAT was used in this study. Parameters can be altered in 3 ways in SWAT+ : (1) absolute value change, (2) change the value by relative percentage, and (3) change the value by a specified amount. For instance, if the original parameter value in SWAT+ is x, then the new value of a parameter (x') after assigning change (C) under different change types will be absolute value change (x' = x + c), percentage change (x' = x + c*x), replace (x' = c) (Nguyen et al., 2022). The calibration (CAL) file included with SWAT+shows the absolute minimum and maximum range of the parameters in the model. We used the replace parameter change option for the basin-wide parameters and a few aquifer-level parameters based on previous literature (Table 5). We justify using replace options for aquifer-level parameters due to SWAT+ defaulting to assigning uniform values for all aquifers in a given watershed; thus, we have no loss of resolution at the aquifer level. For parameters with a narrow range (range 0-1), we used absolute value change, and for the parameters with a wide range, we used the relative change option (Table 5). This allowed us to target a large number of parameter combinations with limited simulations. Both watersheds were automatically calibrated using three different methods (LHS, DDS, GLUE) with three rainfall datasets (ERA-5, gridMET, GPM IMERG) in R-SWAT. Based on the literature review (Bieger et al., 2017;Arnold et al., 2012;Pulighe et al., 2021) and watershed characteristics, we selected the 17 most widely used flow parameters (Table 5) for sensitivity analysis and model optimization. Sensitivity analysis was conducted utilizing the Latin Hypercube Sampling (LHS) in parallel mode through R-SWAT, performing a total of 2100 simulations with the selected calibration parameters. An α value of 0.05 was used to determine sensitivity, with parameters having p-values less than 0.05 considered sensitive and confirmed with high t-statistic (t-stat) values. We analyzed models with three performance indices (NSE, KGE, R²) and visual interpretation to explore the goodness-of-fit of SWAT+ simulated flow for different scenarios. Our preliminary analysis using the GPM IMERG rainfall dataset identified consistent sensitive parameters for both the Tar-Pamlico and Cape Fear watersheds. We selected diverse parameters from the aquifer, soil, channel, HRU, and basin levels to comprehensively model flow (Bailey et al., 2020;Nguyen et al., 2022;Srinivasan et al., 2010;Tan and Yang, 2020).
This section presents a comparison of Pearson correlation coefficients between various rainfall satellite datasets and USGS observed rainfall data to evaluate their accuracy and reliability for rainfall representation in different regions. Pearson correlation coefficients indicated that the gridMET dataset consistently exhibited the highest correlation with observed data in both locations, with values of 0.76 and 0.87 for Kelly and Greenville, respectively. IMERG followed closely, showing strong correlations of 0.73 for Kelly and 0.76 for Greenville. In contrast, the ERA5 dataset showed the weakest correlation, particularly in Greenville, where the correlation coefficient was only 0.092. These results suggest that gridMET and IMERG are more reliable datasets for rainfall representation in these regions, while ERA5 may not be as representative of observed data.
For Kelly, gridMET exhibited strong correlations with observed data across all seasons, especially in fall (0.88) and winter (0.79), based on the Pearson correlation coefficient. IMERG also performed well, with significant correlations in fall (0.85) and winter (0.73). Similarly, in Greenville, gridMET showed the highest correlations in fall (0.91) and winter (0.84), while IMERG displayed robust correlations across all seasons, particularly in fall (0.82) and winter (0.76). ERA5, however, demonstrated weaker and occasionally negative correlations, especially in winter for both locations, suggesting potential misalignment in capturing seasonal rainfall patterns. For clearer visual interpretation, we plotted the rainfall data for the year 2011 (Fig. 3), which includes Hurricane Irene.
The sensitivity analysis conducted for the Tar-Pamlico and Cape Fear watersheds identified key parameters affected by different rainfall datasets. For the Tar-Pamlico watershed with the (a) ERA-5 rainfall data we found the sensitive parameters to be cn2, revap_co, flo_min, revap_min, awc, alpha, perco, and epco (b) gridMET dataset, the sensitive parameters were cn2, revap_co, flo_min, revap_min, awc, epco and cn3_swf (c) GPM IMERG dataset we found the sensitive parameters to be cn2, revap_co, flo_min, revap_min, awc, alpha, and esco (Fig. 4).
In the case of the Cape Fear river basin, for the (a) ERA-5 rainfall data we found the sensitive parameters to be cn2, revap_co, flo_min, revap_min, awc, alpha, perco, and epco (b) gridMET dataset the sensitive parameters were cn2, revap_co, flo_min, revap_min, awc, alpha, perco, epco and cn3_swf (c) GPM IMERG dataset we found the sensitive parameters to be cn2, revap_co, flo_min, revap_min, awc, alpha, perco, and cn3_swf (Fig. 4).
Five commonly sensitive parameters were identified-cn2, revap_co, flo_min, revap_min, and awc-across different rainfall datasets for the Cape Fear and Tar-Pamlico watersheds (Fig. 4). Flow sensitivity was notable for the alpha parameter (indicating baseflow) across all rainfall datasets and both watersheds, with the exception of the Tar-Pamlico watershed model using the gridMET rainfall dataset. The percolation coefficient (perco) parameter was found to be sensitive under all rainfall datasets for the Cape Fear watershed and with the ERA-5 dataset for the Tar-Pamlico watershed. For gridMET, flow simulations in both watersheds showed sensitivity to epco (plant uptake compensation factor) and cn3_swf (soil water factor for the cn3 parameter). This study revealed unique challenges in the selected watersheds due to the complex interactions between land and water processes. Our findings provide a reference for applying these parameters, tested across various SPPs and calibration techniques, while underscoring essential challenges and considerations for accurately modeling coastal watershed dynamics.
We tested nine different simulated flow scenarios with the combination of three rainfall datasets and three autocalibration techniques using SWAT+ for each watershed. One method of assessing model accuracy is visually comparing the simulated and observed flow. As shown in Fig. 5 for the Tar-Pamlico watershed, the high observed flow in 2011 (above 250 m 3 s -1 ) occurred around late August and early September (Fig. 5). This increased flow was due to Hurricane Irene -a category one hurricane -hitting eastern North Carolina. The high flows were measured at USGS Greenville station from August 28th 2011, to September 5th 2011, with the highest observed daily flow on August 31st 2011, of 345 m 3 s -1 (Fig. 5). We simulated the model for Tar-Pamlico from 2001 to 2016, but to get more insights into SWAT+ peak flow and baseflow predictions, we have shown the predictions for the year 2011 in Fig. 5 for the Tar- Pamlico watershed. This helped us understand the combined effects of rainfall datasets and optimization techniques on simulated storm flows. Due to space limitations, all 18 graphs of simulated and observed flow for the 3 rainfall datasets and 3 optimization techniques across the 2 watersheds are shown in Supplementary Material (Figures S1 through S18 in S2 Supplementary Material).
SWAT+ fell short of capturing peak flow duration for all scenarios. The calibrations using the GPM IMERG rainfall product did the best job predicting the magnitude of peak flow, whereas gridMET did the worst. Simulated flows using all three rainfall products under all optimization techniques underestimated the time to peak flow and subsequent hydrograph recession times. For GPM IMERG data, the GLUE optimization technique (358 m 3 s -1 ) performed better than LHS (426 m 3 s -1 ) and DDS (309 m 3 s -1 ) for simulating the peak flow. For ERA-5 data, LHS (311 m 3 s -1 ) and GLUE (372 m 3 s -1 ) optimization procedures performed better than DDS (595 m 3 s -1 ), for simulating the peak flow. For gridMET data, the LHS (243 m 3 s -1 ) optimization method performed much better than both GLUE (974 m 3 s -1 ) and DDS (1013 m 3 s -1 ) for simulating the peak flow, as DDS and GLUE provided unreasonably high flows.
The results in the Cape Fear basin (Figure S19 Supplementary Material) for each calibration method and rainfall dataset were similar to the results in the Tar-Pamlico basin. Fig. 5 and Figure S19 (Supplementary Material) include period of baseflow and storm flow simulation in Tar-Pamlico and Cape Fear using different rainfall data sets. We found that the combinations of the ERA-5 dataset with LHS or GLUE, gridMET dataset with LHS, and GPM IMERG dataset with either GLUE or LHS give better baseflow magnitude prediction results. In the Cape Fear watershed, we observed high baseflows (greater than 40 m 3 s -1 ). LHS performed better than DDS and GLUE in most cases, implying that the LHS technique should be used over DDS and GLUE for groundwater-driven watersheds (Figure S19 Supplementary Material).
In this study, we systematically calibrated flow parameters to maximize the Nash-Sutcliffe Efficiency (NSE) utilizing three distinct rainfall datasets and three optimization techniques within two different watersheds. KGE and R 2 were evaluated based on all 2100 simulations for a watershed with a specific set of rainfall data. Fig. 6 depicts the maximum value of the particular performance metric amongst all simulations for the given scenario.
GPM IMERG data produced optimal flow for both watersheds under LHS and GLUE methods. For the Tar-Pamlico watershed, we found the best KGE (0.424) and R 2 (0.501) for the ERA-5 dataset with the LHS method and the best NSE (0.323) with the GLUE method. The gridMET dataset gave the best KGE (0.188) and R 2 (0.439) with the LHS method and the best NSE (0.059) with the GLUE method. The GPM IMERG rainfall data produced the best KGE (0.619) with the LHS method and the best NSE (0.412) and R 2 (0.498) with the GLUE method (Fig. 6).
For the Cape Fear River basin, we found the best NSE (0.449), KGE (0.522), and R 2 (0.559) for ERA-5 data with the LHS method. The gridMET dataset gave the best NSE (0.341) and R 2 (0.498) with the GLUE method and the best KGE (0.426) with the LHS method. For the GPM IMERG dataset, we found the best NSE (0.498) and KGE (0.676) with the GLUE method and the best R 2 (0.532) with the LHS method (Fig. 6).
The ERA-5 and GPM IMERG datasets performed better than the gridMET rainfall dataset when evaluated using KGE and NSE indices (Fig. 6). On the other hand, R 2 performed better than NSE and KGE for the gridMET dataset (Fig. 6). We found the worst performance index values for gridMET rainfall data when optimized with the DDS technique (NSE, KGE, R 2 -Tar-Pamlico: -0.236, 0.057, 0.212 -Cape Fear: 0.029, 0.148, 0.304).
We found that KGE and R 2 varied less among different scenarios than NSE (Fig. 7). In most cases, we found higher performance indices for ERA-5 and GMP IMERG rainfall datasets with the GLUE optimization technique (Fig. 6). We found comparatively better results for the gridMET dataset with the LHS optimization technique (NSE, KGE, R 2 -Tar-Pamlico: 0.041, 0.188, 0.439 -Cape Fear: 0.314, 0.426, 0.489). NSE gave lower values than KGE and R 2 under all scenarios (Fig. 6). Maximum values of KGE and R 2 varied based on the rainfall dataset and optimization technique. We found R 2 to give consistently higher values than KGE and NSE for the gridMET rainfall dataset (Fig. 6). Overall, our findings suggest that the GPM IMERG rainfall dataset, combined with the GLUE optimization technique, provides the most effective approach for optimal flow simulations in our test watersheds (Fig. 6).
We conducted a parameter uncertainty analysis to assess the variability and reliability of the parameters used in predicting streamflow. This analysis encompassed the nine scenarios, involving three rainfall datasets and three optimization techniques, applied across both watersheds, as depicted in Fig. 7. We compared the optimal values variation of the sensitive parameters under the best NSE for a given scenario. Fig. 7 shows the box plot of the five common sensitive parameters (cn2, revap_co, awc, flo_min, and revap_min) for both watersheds. We used relative change (cn2, flo_min, revap_min) and absolute value change (awc, revap_co) for these parameters (Table 5).
We assumed that the uncertainty in the prediction of these parameters is only from the rainfall dataset and optimization technique used, as we kept all other data (soil map, land cover, observed data, and other SWAT+ input data) constant while building the model. We found the largest variation in the flo_min (Tar-Pamlico: 0.03-0.48; Cape Fear: -0.11-0.49) and revap_min (Tar-Pamlico: -0.25-0.21; Cape Fear: -0.22-0.18) parameters between both watersheds, which implies those are more uncertain than the other three parameters. We found comparatively less variation amongst other sensitive parameters for awc (Tar-Pamlico: 0.01-0.29; Cape Fear: 0.01-0.26), cn2 (Tar-Pamlico: -0.29-0; Cape Fear: -0.29 to -0.06), and revap_co (Tar-Pamlico: 0-0.09; Cape Fear: 0-0.08).
The optimal parameter values for the Cape Fear (Table S1, Supplementary Material) and Tar-Pamlico (Table S2, Supplementary Material) watersheds, under three rainfall datasets and optimization techniques in SWAT+ flow simulations, are presented in the supplementary material. The varied range of optimal parameters (Table S1 and S2 Supplementary Material) under different combinations suggests that the choice of rainfall datasets and optimization techniques significantly affects parameter uncertainty. For instance, the Tar-Pamlico watershed showed considerable variation in the optimal values of alpha.aqu and canmx.hru, while the Cape Fear watershed exhibited substantial differences in revap_min.aqu and bf_max.aqu values across different scenarios (Table S1 and S2 Supplementary Material). These findings underscore the importance of selecting appropriate datasets and techniques tailored to each watershed's characteristics to reduce uncertainty and enhance model reliability.
The results of this study demonstrate a clear preference for the gridMET and IMERG datasets over ERA-5 for accurately representing rainfall in the Kelly and Greenville locations. The superior performance of gridMET, evidenced by its high Pearson correlation coefficients with observed data, highlights its robustness and suitability for hydrological modeling and climate analysis in the study regions (Tapas et al., 2024d). IMERG also showed strong correlations, indicating its reliability. The comparatively weaker performance of ERA-5, especially in Greenville, suggests potential limitations in accurately capturing local rainfall patterns in coastal regions. These findings are critical for researchers and practitioners in selecting appropriate datasets for various applications, emphasizing the need for a thorough evaluation of dataset performance before use.
The seasonal analysis highlights the robustness of gridMET and IMERG across various seasons, with gridMET consistently demonstrating high correlations in fall and winter. This seasonal consistency is vital for applications that require accurate seasonal rainfall data, such as agricultural planning and water resource management. In contrast, the negative correlations observed for ERA-5 during winter at both locations indicate potential weaknesses in capturing seasonal rainfall variability. Recognizing these seasonal discrepancies is crucial for refining dataset algorithms and enhancing predictive accuracy. Future research should aim to address these limitations, investigate the underlying causes of seasonal performance variations, and develop more precise and reliable rainfall datasets.
In analyzing the Tar-Pamlico and Cape Fear watersheds across three distinct rainfall datasets, five sensitive parameters consistently emerged: cn2, revap_co, flo_min, revap_min, and awc. This consistency highlights the critical dependence of flow on these factors, which govern surface runoff and infiltration partitioning, evapotranspiration, and groundwater discharge to streams. The recurring presence of these parameters across varied scenarios underscores their essential role in shaping watershed responses to different rainfall inputs, affirming their significance in SWAT+ modeling and the broader understanding of watershed biogeochemical processes. Our findings align with previous studies, which also identified these parameters as sensitive for streamflow simulation (Wagner et al., 2022;Reza & Salmani, 2023;Yen et al., 2019).
The high sensitivity of the alpha parameter for most scenarios showed the significance of groundwater flow response to recharge changes in flow simulation for ERA-5 and GPM IMERG rainfall datasets and its dependency on the watershed under the gridMET rainfall dataset. In SWAT+ , the perco parameter plays a critical role in streamflow simulation, demonstrating sensitivity across most scenarios. This parameter is derived using the Green-Ampt equation for infiltration, as detailed by Rawls et al. (1983), which significantly influences the model's hydrological response. This parameter represents the groundwater infiltration rate as a function of soil hydraulic properties, such as soil texture, structure, and pore size distribution, as well as soil water content and other factors that affect the soil's ability to transmit water (SWAT+++, 2020).
The epco parameter specifically addresses simulation of plant water uptake in SWAT+ . It is a dimensionless factor that adjusts the plant's root water uptake to account for the reduction in potential evapotranspiration (PET) due to water stress. Sensitivity to the epco parameter can be anticipated in regions with high evapotranspiration rates, such as the Cape-Fear and Tar-Pamlico watersheds. These two watersheds exhibit a humid subtropical climate, resulting in high precipitation rates that produce elevated evapotranspiration rates, making the epco parameter critical in evapotranspiration as it adjusts plant root water uptake under water stress conditions.
The newly introduced parameter, cn3_swf, in SWAT+ facilitates the simulation of infiltration and surface ponding before runoff begins after a dry spell (Wagner et al., 2022). Studies by Eini et al. (2023), Tumsa et al. (2022), Wagner et al. (2022), andWhite et al. (2022) provide a general perspective on the cn3_swf parameter, which adjusts the daily curve number based on soil water content. However, comprehensive documentation on using cn3_swf in SWAT+ calibration remains incomplete. We observed that the Cape Fear River's flow shows sensitivity to cn3_swf under gridMET and GPM IMERG rainfall datasets, likely due to its humid subtropical climate, which results in high annual rainfall. Additionally, the watershed's highly porous soils with permeable sediments may contribute to increased groundwater recharge from rainfall. Furthermore, the Tar-Pamlico watershed demonstrated high sensitivity to the esco parameter (soil evaporation compensation factor) with the GPM IMERG dataset, enhancing SWAT+ 's ability to capture the watershed's humid subtropical climate characteristics under GPM IMERG conditions.
Many SWAT and a few SWAT+ studies mentioned the impacts of watershed characteristics on the performance of rainfall products (Pulighe et al., 2021;Cho et al., 2009). In this study, we also found that the sensitivity of calibration parameters is related to the characteristics of the watershed. Our findings aligned with the studies of Fereidoon et al. (2019) and Guo and Su. (2019) in Tar-Pamlico and Cape Fear, in which curve number (cn2) was found to be the most sensitive parameter with all chosen rainfall datasets and watershed characteristics. In addition, our results aligned with previous works (Migliaccio and Chaubey, 2008;G. Wang et al., 2014;Zhao et al., 2018) that identified revap_co, flo_min, revap_min, awc, alpha, perco, epco, cn3_swf, and esco as significant parameters affecting river flow.
In this study, we analyzed the performance of three rainfall datasets and three optimization techniques on streamflow prediction. We found GPM IMERG rainfall dataset when used with GLUE optimization technique generally yields better results for flow simulations. Our findings partially address the research gaps identified by Yuan et al. (2019), who called for an evaluation of the compatibility between the GPM IERG dataset and the GLUE optimization technique for optimal flow simulations. Mousavi et al., (2023) reported high values of KGE (0.88) for IMERG-Terra model when used with the GLUE optimization technique.
We have found that the selection of a rainfall dataset can play a vital role in flow simulations. This observation aligns with previous literature on SWAT modeling, where multiple rainfall datasets were used to assess their impact on optimal flow simulations (Cho et al., 2009;Do et al., 2022). We found a couple of different combinations for optimal flow simulations based on the specific modeling objective. As mentioned earlier, for achieving optimal overall and peak flow simulation, the combination of GPM IMERG rainfall data with the GLUE optimization technique has proven to be more effective. For optimal base flow simulation, the combinations of the ERA-5 rainfall dataset with either DDS or LHS optimization techniques, or the GPM IMERG rainfall dataset with either GLUE or LHS optimization techniques have been found to work better.
The GPM IMERG rainfall dataset demonstrated the best performance in simulating peak flow magnitude compared to other rainfall datasets. Our results are consistent with previous studies, such as Le et al. (2023), who found that GPM IMERG outperformed five other satellite precipitation products in simulating flood peaks across eleven Vietnamese river basins using the SWAT model. This dataset received high ratings in our assessment due to its use of multiple microwave and infrared sensors, along with post-calibration using gauge-based products (Hou et al., 2014).
In general, SWAT+ performance varied greatly for peak flow magnitude simulation, with the GPM IMERG rainfall dataset and GLUE optimization technique combination yielding the best results. SWAT+ underperformed in capturing peak flow duration. Both SWAT and SWAT+ are designed primarily for long-term continuous predictions and have limitations when it comes to capturing instantaneous peak flows (Yu et al., 2018). When simulating flows with all three rainfall products and optimization methods, we consistently experienced an issue of timing peak flow and the subsequent hydrograph recession occurring too early. This underscores the significant influence of the curve number (cn2) on hydrograph dynamics during events equivalent to hurricane intensity (Caviedes-Voullième et al., 2012;Lyon et al., 2004). The accuracy of streamflow simulations displayed notable differences depending on the rainfall datasets used. This reinforces the observation that in humid climates, SWAT+ flow simulations exhibit sensitivity to the specific rainfall data (Tan and Yang, 2020).
The Cape Fear watershed is predominantly characterized by sandy soils, conducive to infiltration. It is considered as groundwaterdriven watershed (Ewen et al., 2011;Valley, 1988), leading to pronounced baseflows sourced from a combination of deep, confined, and shallow aquifers (Ewen et al., 2011;Valley, 1988). However, for the crucial output of baseflow, SWAT+ underperformed in this specific watershed. The model did not accurately predict the magnitude of baseflow. This could be explained by the model's simplified groundwater flow equations. These equations may not fully capture the intricacies of baseflow in a watershed that's so heavily influenced by groundwater, as indicated by Bailey et al. (2020).
Our findings indicate that there's no single combination of rainfall dataset and optimization technique is universally superior to other which aligns with previous literature (Dos Santos et al.,. 2022;Mararakanye et al., 2020). The combination should be selected based on data availability, computational power availability, watershed location, and watershed characteristics. GPM IMERG (Hou et al., 2014) and ERA-5 (Muñoz-sabater et al., 2021) data are available worldwide, whereas gridMET data is available only for the contiguous United States (Abatzoglou, 2013). There are not regional limitations for choosing optimization techniques; all of them can be accessed worldwide through various open-source applications. We compared these techniques using an equal number of simulations to demonstrate how they perform with similar computational power.
Performance metrics (NSE, KGE, and R 2 ) are used to measure the goodness-of-fit of the simulation to observed data and could be biased based on data magnitude, outliers, repetition, and other factors (Ritter and Muñoz-Carpena, 2013). In this study, ERA-5 and GPM IMERG datasets performed better than the gridMET rainfall dataset when evaluated using R 2 , KGE, and NSE indices (Fig. 6). The worst performance index values for gridMET rainfall data were indicated when using with the DDS technique (NSE, KGE, R 2 -Tar-Pamlico: -0.236, 0.057, 0.212 -Cape Fear: 0.029, 0.148, 0.304). Our findings contrast with findings from Tolson and Shoemaker (2007), which determined that DDS was superior in efficiency compared to GLUE. For DDS, our experiment was performed with 175 iterations across 12 threads, yielding a total of 2100 simulations. We hypothesize that the chosen 175 iterations or the 12-thread configuration might not be adequate to encompass the extensive variety of parameter combinations when employing the DDS's second parallel approach, as mentioned by Nguyen et al. (2022). The method, which involves parallel processing by assigning parameter values to subsequent cores, remains an area where further validation is needed (Nguyen et al., 2022).
Our findings, as shown in Fig. 6, revealed that KGE and R 2 exhibited higher values across different scenarios compared to NSE when assessing the performance of three rainfall datasets and three optimization techniques in both watersheds. The performance of KGE, especially when compared to NSE, can be attributed to its complex assessment of hydrological model outputs. Specifically, KGE incorporates three different components: correlation (r), variability (alpha), and bias (beta) (Knoben et al., 2019). In contrast, NSE is calculated using the mean squared error between the simulated and observed flows (Moriasi et al., 2007). This more thorough assessment typically translates to hydrological simulations that are both more precise and dependable (Gupta et al., 2009;Moriasi et al., 2007). It is worth noting that had we adjusted the model specifically for KGE and R 2 , similar to our optimization for NSE, their performance metrics likely would have been higher.
In this study, we performed experiments with the assumption that the only differences in selecting parameter values stemmed from the rainfall dataset and the optimization techniques used. Our analysis revealed considerable variations, especially in the flo_min parameters for both Tar-Pamlico (ranging from 0.03 to 0.48) and Cape Fear (-0.11-0.49) as well as in the revap_min parameters for Tar-Pamlico (-0.25-0.21) and Cape Fear (-0.22-0.18). The observed variations suggest higher uncertainties in these parameters compared to the other three that were sensitive in all scenarios, emphasizing the need for cautious consideration and potential refinement.
The optimization methods for SWAT+ that we examined were successful in parameterizing revap_min and facilitated efficient water movement from shallow aquifers to the unsaturated zone, given the appropriate threshold (SWAT+++, 2020). The minimal variation observed in the awc parameter highlights SWAT+ 's effectiveness in accurately modeling both the flow movement through landscapes and its uptake by vegetation. It has implications for predictions related to runoff and evapotranspiration. The minor discrepancy in the cn2 values across the two watersheds might be attributed to similarities in either land slope or land cover across these regions. Supporting this, Chordia et al. (2022) identified a lesser degree of variation in cn2 in comparison to eight other different parameters over four watersheds. In conclusion, SWAT+ parameterizes key components, such as cn2, the groundwater revap coefficient, and the threshold depth of shallow aquifers where revap takes place, underscoring its effectiveness in SWAT+ modeling.
When correlated with observed rainfall, GridMET performed the best due to its integration of USGS rainfall gauge data, followed by IMERG, and ERA-5, which showed the poorest correlation. However, this direct rainfall comparison does not necessarily translate to the best performance when using rainfall data to force a hydrological model. In flow simulations, IMERG gave the highest flow performance indices, followed by ERA-5, with GridMET performing the worst. This non-linear relationship occurs because the calibration process of hydrological models involves adjusting parameters to compensate for errors in the rainfall input (Le et al., 2020). This discrepancy highlights that while gridMET excels in direct rainfall comparisons, the calibration process in hydrological modeling can mitigate errors in rainfall input, leading to varied performance outcomes. Le et al. (2020) had similar results, showing that CHIRPS had better direct rainfall correlations with rain gauges compared to the Tropical Rainfall Measuring Mission products, 3B42V7 (research-quality) and 3B42RT (real-time). However, 3B42V7 and 3B42RT performed better in streamflow simulations due to their higher temporal resolution. GPM IMERG's superiority in flow performance can be attributed to its high temporal resolution and comprehensive coverage, enabling it to capture rainfall events more accurately. Additionally, the finer temporal resolution of ERA-5 and IMERG might improve accuracy away from the rain gauges. Despite the ERA-5 rainfall dataset showing poorer correlations with observed data, its flow performance was not the worst. This can be explained by its consistent temporal resolution, which may help capture broader rainfall patterns adequately. In contrast, GridMET performed the worst in flow simulations, possibly due to its blending of PRISM and NLDAS-2 datasets, which might introduce inconsistencies that negatively affect flow simulations. Overall, the robust performance of GPM IMERG in simulating high flows underscores its capability to accurately represent extreme rainfall events, which is crucial for peak flow simulations in hydrological models. This study's limitation is its broad calibration approach across two watersheds, without fine-tuning parameters for localized hydrological processes, which may impact flow prediction precision. This approach ensures a consistent framework for comparing different rainfall datasets and autocalibration techniques, highlighting the generalizability of our findings across varied hydrological settings. It helps identify overarching trends and impacts on SWAT+ model performance, offering insights not confined to localized conditions. This method is valuable for USA and global watershed modeling with limited observed data, demonstrating how combinations of different SPPs and optimization techniques affect model performance without extensive local conditions inputs.
This study demonstrated the significant impact of different rainfall datasets and autocalibration techniques on the accuracy of flow simulations in the SWAT+ model. Flow simulation was influenced more by rainfall forcing than the autocalibration method. The five most sensitive parameters identified were cn2, revap_co, flo_min, revap_min, and awc, with cn2 being the most sensitive. The newly added SWAT+ parameter cn3_swf was sensitive in the groundwater-driven watershed, highlighting the interconnectedness of surface and subsurface processes. Significant variation in simulated flow was observed under different scenarios, with overall better predictions for ERA-5 and GPM IMERG datasets using the GLUE optimization technique. The gridMET dataset performed better with the LHS technique.
Autocalibration techniques play an essential role in simulating SWAT+ baseflow accurately compared to storm flows, while the rainfall dataset is crucial for both storm flows and baseflow predictions. Overall, SWAT+ captures baseflow magnitude better than peak flow magnitude. GPM IMERG outperformed ERA-5 and gridMET in predicting peak flow magnitude. We suggest using the ERA-5 dataset with DDS or LHS, the gridMET dataset with LHS, and the GPM IMERG dataset with either GLUE or LHS for improved baseflow predictions. SWAT+ struggled to predict peak flow duration during extreme events like Hurricane Irene. KGE and R 2 showed higher values across most scenarios compared to NSE, with R 2 providing consistently higher values for the gridMET dataset. Optimization techniques can improve flow simulations even for poorly correlated SPPs. This study will assist hydrological modelers worldwide in selecting the optimal combination of rainfall dataset, calibration method, and performance index for more accurate flow simulations.
Future research could focus on refining localized parameter calibration, incorporating decomposition methods to better capture data noise and variability. Additionally, exploring the applicability of these findings across diverse climatic regions could enhance the robustness and generalizability of SWAT+ model predictions, ensuring they perform well under various environmental conditions. Expanding the study to include additional rainfall datasets like TRMM and CHIRPS could also provide a broader perspective, enabling comparisons with other widely used datasets.
Journal of Hydrology: Regional Studies 57 (2025) 102134