Web-Based Baseflow Estimation in SWAT Considering Spatiotemporal Recession Characteristics Using Machine Learning

Abstract

The increasing frequency and severity of hydrological extremes due to climate change necessitate accurate baseflow estimation and effective watershed management for sustainable water resource use. The Soil and Water Assessment Tool (SWAT) is widely utilized for hydrological modeling but shows limitations in baseflow simulation due to its uniform application of the alpha factor across Hydrologic Response Units (HRUs), neglecting spatial and temporal variability. To address these challenges, this study integrated SWAT with the Tree-Based Pipeline Optimization Tool (TPOT), an automated machine learning (AutoML) framework, to predict HRU-specific alpha factors. Furthermore, a user-friendly web-based program was developed to improve the accessibility and practical application of these optimized alpha factors, supporting more accurate baseflow predictions, even in ungauged watersheds. The proposed HRU-specific alpha factor approach in the study area significantly enhanced the recession and baseflow predictions compared to the traditional uniform alpha factor method. This improvement was supported by key performance metrics, including the Nash–Sutcliffe Efficiency (NSE), the coefficient of determination (R²), the percent bias (PBIAS), and the mean absolute percentage error (MAPE). This integrated framework effectively improves the accuracy and practicality of hydrological modeling, offering scalable and innovative solutions for sustainable watershed management in the face of increasing water stress.

1. Introduction

Recent climate change has led to a notable increase in the frequency and magnitude of natural disasters, such as floods and droughts, highlighting the urgency of effective water resource management [1,2]. Among the various components of watershed hydrology, baseflow is a critical metric of groundwater discharge rate that constitutes a significant portion of the streamflow in most catchments, playing an essential role in hydrologic analysis, groundwater characterization, hydrologic modeling, and water resource management [3,4,5]. Moreover, the accurate estimation of baseflow plays a key role in maintaining both water quality and quantity, and ensuring there are adequate resources to support ecosystems, agriculture, and human needs [6,7,8]. Among watershed models, the Soil and Water Assessment Tool (SWAT) has been widely assessed for its applicability across various watersheds, with various studies focusing on baseflow analysis [9,10,11]. However, SWAT’s performance in baseflow analysis often suffers from challenges in parameter calibration and inherent model limitations [12]. A SWAT parameter, the alpha factor, governs the recession curve of groundwater discharge and significantly influences baseflow predictions [13]. Although SWAT is already designed to assign specific values for each hydrologic response unit (HRU), research on the methods used to apply alpha factor values for each HRU remains insufficient. In regard to this, previous studies evaluated the baseflow by applying the same alpha factor to each HRU without considering temporal and spatial variations [14,15,16]. This approach can lead to inaccuracies in baseflow predictions, particularly in watersheds with heterogeneous hydrological and complex geological conditions. To deal with this limitation, Lee et al. [12] proposed a method to address the spatial and temporal variability of the alpha factor. Their study demonstrated that incorporating spatially distributed alpha factors could significantly improve the baseflow estimation accuracy, especially in complex watershed systems. However, the Lee et al. [12] approach is limited by its lack of accessibility, as it does not provide a direct mechanism for users to access or apply these spatially and temporally resolved alpha factors in their analyses. This means that this lack of accessibility reduces the broader applicability and utility of their findings for real-world water resource management.

To address these limitations, recent advancements in machine learning (ML) approaches are widely recognized for their strong predictive capabilities for analyzing large datasets and capturing complex non-linear relationships [17,18,19,20]. Moreover, integrating physical knowledge into ML models, such as hybrid ML–physical models and physics-informed ML approaches, has been explored to improve the predictive performance [21,22]. Although these advancements have improved the applicability of ML in hydrological modeling, comparing multiple ML models to identify the best performer remains both time-consuming and labor-intensive [23]. Automated ML (AutoML) offers solutions to these challenges. AutoML frameworks, such as the Tree-Based Pipeline Optimization Tool (TPOT), provide a streamlined solution by automating tasks like feature engineering, model selection, and hyperparameter tuning, enhancing efficiency and reducing the burden on researchers [24,25,26]. By integrating AutoML with the SWAT, it is possible to develop a more robust framework for baseflow analysis, reducing uncertainties and enhancing the predictive accuracy. Moreover, developing user-friendly interfaces is needed to enable non-expert users to access spatially and temporally optimized alpha factors and apply baseflow analysis techniques [27,28,29]. Such developments are crucial for bridging the gap between advanced research and practical water resource management. By combining the SWAT with machine learning techniques, this research proposes a method for improving baseflow estimations.

Accordingly, the objectives of this study were as follows: (1) Utilize AutoML, such as the TPOT, to train and validate HRU-specific alpha factors based on the findings from Lee et al. [12] and develop a web-based program for improved accessibility and practicality. (2) Compare baseflow estimates obtained using conventional uniform alpha factor methods (Case 1) with those derived from the proposed HRU-specific alpha factor approach (Case 2). Figure 1 illustrates the procedure followed in this study.

2. Methods

2.1. Study Watershed

The Gapcheon watershed, a South Korean sub-basin of the Geum River watershed, covers an area of 602 km² (Figure 2) and is home to over 1 million residents. Over the past two decades, the region has experienced significant development, highlighting the need for effective water resource management [13]. In this study, the land use within the watershed was primarily composed of forests, accounting for 71.07% (428.00 km²) of the area, followed by urbanized areas at 12.59% (75.76 km²), agricultural land at 9.31% (56.07 km²), and pasture at 4.05% (24.39 km²), among other uses. Moreover, according to the Korea Meteorological Administration, the study area experienced an average annual precipitation of 1273.6 mm and an average annual temperature of 13.2 °C over the eight years from 2010 to 2017 (

Table 1. The trend of temperature and precipitation in the study areas.

Year	Annual Precipitation (mm)	Monthly Precipitation (mm)		Annual Temperature (°C)
		Max.	Min.	Max.	Min.
2010	1419.7	376.4	16.4	24.7	1.3
2011	1943.4	587.3	4.0	23.4	2.4
2012	1409.5	463.6	2.5	24.2	2.3
2013	1120.2	218.7	19.9	25.9	2.0
2014	1117.7	240.9	6.5	24.8	3.1
2015	822.7	145.6	27.0	25.4	2.7
2016	1228.4	367.9	11.6	26.4	2.6
2017	1127.5	434.5	11.6	25.3	2.3

). During this time, the yearly minimum and maximum temperatures were 1.3 °C and 26.4 °C, respectively. As shown in Figure 2, the study area has ten streamflow monitoring stations (Yongchon, Dugye, Munam, Inchang, Gasuwon, Boksu, Hanbat, Mannyeon, Daedeok, and Wonchon), with the Wonchon station as the final main outlet. These data were obtained from the Water Management Information System.

2.2. SWAT Input Data and Overview

The SWAT is a continuous, long-term, semi-distributed hydrological model developed by the United States Department of Agriculture Agricultural Research Service (USDA-ARS) to simulate streamflow and water quality at the watershed scale while incorporating various land management practices [30]. The model effectively simulates the hydrological cycle using HRUs as its fundamental computational unit. Each HRU represents a unique combination of land use, soil type, and slope within individual sub-basins, allowing for detailed spatial representation of watershed characteristics [31]. Additionally, SWAT automatically estimates soil-related variables for each soil layer using soil property data stored in its internal database, enhancing the model’s ability to simulate soil–water interactions [32]. For baseflow, SWAT assumes an aquifer system composed of both shallow and deep aquifers.

In this study, the Gapcheon watershed was divided into ten sub-basins, corresponding to the locations of streamflow monitoring stations (Figure 2), to improve spatial accuracy in hydrological simulations. SWAT requires various input data, including meteorological, topographic, and geographical information on land use and soil (Figure 3).

Table 2. Description of data and sources used for model setup.

Data Type	Name	Source
Spatial data	Digital Elevation Model (DEM)	National Geographic Information Institute
	Land use	Korea Ministry of Environment
	Soil type	Korea Rural Development Administration
Meteorological data	Precipitation, wind speed, maximum and minimum temperature, relative humidity, and solar radiation (daily timeseries, 2010–2017)	Korea Meteorological Administration
Hydrological data	Streamflow (daily timeseries, 2010–2017)	Water resource Management Information System

presents the climate data from the Daejeon station of the Korea Meteorological Administration (KMA) as daily data (maximum temperature, minimum temperature, humidity, wind, solar radiation, and precipitation) from 1 January 2010 to 31 December 2017. The streamflow data were obtained from the Water Resource Management Information System (WAMIS), which provides daily streamflow records for ten stations (Gasuwon, Daedoek, Mannyeon, Dugye, Munam, Boksu, Yongchon, Inchang, Hanbat, and Wonchon) during the same period. Moreover, the National Geographic Information Institute provided the study watershed’s Digital Elevation Model (DEM) with a 1:5000 scale numerical map with a 10 m grid resolution. The model analysis spans eight years (2010–2017), with a four-year warm-up period (2010–2013) to minimize the influence of initial conditions and ensure model stability. This comprehensive setup allowed SWAT to effectively simulate hydrological processes and assess baseflow variability within the Gapcheon watershed.

2.3. Machine Learning for Spatiotemporal Alpha Factor Estimation in SWAT: A TPOT-Based Approach

Building machine learning models is often labor-intensive and time-consuming due to model selection and hyperparameter tuning complexity. However, AutoML, such as TPOT, effectively streamlines this process by automatically exploring and optimizing various models and configurations, significantly reducing manual effort and improving overall efficiency. TPOT, an AutoML algorithm based on genetic programming, automates the complex and time-consuming tasks of model selection and hyperparameter tuning by intelligently exploring thousands of ML pipelines to identify the most effective combination for the data [33]. The best-performing pipelines and parameters found while TPOT is working are stored and used as representative pipelines at the end of the run. In TPOT, in addition to finding the optimal algorithm, the optimal parameters will also be given. The baseflow recession alpha factor equation is expressed as follows (Equation (1)):

$${\mathrm{Q}}_{\mathrm{t}}=\mathrm{}{\mathrm{Q}}_{\mathrm{t}-1}{\mathrm{e}}^{-\mathsf{\alpha }∆\mathrm{t}}$$

(1)

where ${\mathrm{Q}}_{\mathrm{t}}$ is the baseflow at time t, ${\mathrm{Q}}_{\mathrm{t}-1}$ is the baseflow at the previous time step, $\mathsf{\alpha }$ is the baseflow recession, and $∆\mathrm{t}$ is the time step. The alpha factor controls the rate at which groundwater contributes to streamflow. A higher alpha value indicates a faster decline in baseflow, whereas a lower value suggests prolonged groundwater discharge. This recession constant is influenced by HRU-specific characteristics, including soil properties, slope, and HRU area. Highly permeable soils generally result in lower alpha values, prolonging baseflow contributions, whereas urbanized areas or steep slopes tend to have higher alpha values due to faster runoff and reduced infiltration. Conventional SWAT implementations apply a single, watershed-wide alpha factor, which may not adequately capture spatial and temporal variability. To address this limitation, our study introduces a method that estimates HRU-specific alpha factors while accounting for monthly variations. Unlike conventional approaches that assume a stationary recession coefficient, this method incorporates spatiotemporal variability, leading to a more dynamic and precise estimation of baseflow recession characteristics across diverse hydrological conditions.

This study trained and validated a TPOT model for HRU-specific alpha factor estimation using key hydrological variables, including soil water holding capacity, HRU slope, and HRU watershed area ratio, as proposed in Lee et al. [12]. Additionally, monthly precipitation (January to December) was incorporated in this study to enhance model accuracy by capturing seasonal variations in hydrological processes. The dataset was divided into 19,620 HRU-specific monthly alpha factor data points from seven watersheds for model training, while the remaining 7992 HRU-specific monthly alpha factor data points from three watersheds were used for validation. This division allowed the model to effectively learn spatial and temporal variability across diverse HRUs while ensuring robust generalization. To optimize model performance, TPOT was configured with 10 generations and a population size of 5, using 10-fold cross-validation to systematically evaluate multiple machine learning models and hyperparameter configurations. The framework automatically explored Decision Trees, RandomForest, XGBoost, and Gradient Boosting, applying a genetic algorithm to iteratively refine model structures. The selection process prioritized models that maximized the coefficient of determination (R²) and Nash–Sutcliffe Efficiency (NSE) values while minimizing percent bias (PBIAS), ensuring the most accurate predictions [34,35]. These metrics are commonly utilized in machine learning studies to assess the accuracy of model predictions by comparing estimated values with actual observations [36,37,38,39,40]. After the optimization process, TPOT identified RandomForestRegressor as the optimal model, fine-tuning its hyperparameters through evolutionary search.

To facilitate practical implementation, this study modified SWAT’s source code to allow direct integration of HRU-specific alpha factors, ensuring seamless application in hydrological modeling. Unlike previous studies that required extensive manual preprocessing, the modified SWAT framework enables users to incorporate optimized alpha factors without additional coding. Furthermore, a user-friendly web-based program was developed to provide access to monthly HRU-specific alpha factors, making the method accessible to both researchers and practitioners. This approach effectively overcomes the limitation of estimating alpha factors solely at streamflow observation, enabling accurate estimation even at ungauged stations. By incorporating the spatiotemporal recession characteristics of each HRU, this method ensures a more precise and comprehensive representation of hydrological variability across the watershed, leading to improved baseflow predictions and more reliable water resource management. These advancements collectively enhance the applicability of HRU-specific alpha factor estimation in SWAT, supporting more informed hydrological modeling and decision-making.

2.4. Web-Based Program for Spatiotemporal Alpha Factor Estimation in SWAT Using Automated Machine Learning (TPOT)

This study developed an advanced method for estimating the alpha factor at the HRU level in SWAT by integrating spatiotemporal recession characteristics proposed by Lee et al. [12]. While the conventional method in SWAT can reflect spatial variability, it struggles to capture temporal fluctuations in hydrological responses. Additionally, although SWAT allows for spatially varying alpha factor assignments, previous approaches lacked a systematic method for determining how to allocate alpha factors spatially, and they did not incorporate a way to apply them dynamically over time. To address this limitation, this study employed TPOT to refine the estimation of the alpha factor by incorporating key hydrological parameters such as HRU-specific watershed area ratios, HRU slope, monthly precipitation, and soil water holding capacity. While monthly precipitation influences temporal variability, the estimated alpha factor reflects hydrological factors such as soil water retention capacity, groundwater contributions, and recession characteristics. Although the HRU slope remains constant, it indirectly affects hydrological responses by influencing water infiltration and runoff characteristics rather than directly driving temporal variations in alpha factors. By leveraging machine learning, specifically TPOT, we systematically capture non-linear relationships between recession characteristics and watershed attributes, allowing for a more precise estimation of baseflow recession. This method enhances SWAT’s capability to represent spatiotemporal variability, overcoming the conventional model’s limitations.

To address the challenges of accurately capturing spatial precipitation variability using data from a single climate station, this approach integrates additional hydrological parameters, such as soil water holding capacity and HRU-specific watershed area ratios, which influence runoff generation and baseflow contribution. By incorporating these factors, the estimation of HRU-specific alpha factors is adjusted to better reflect local hydrological conditions. This approach can help supplement the model’s applicability across diverse hydroclimatic conditions, even in cases where spatial precipitation heterogeneity is not explicitly accounted for. Moreover, the developed web-based program enables users to seamlessly integrate these optimized alpha factors into SWAT simulations, improving baseflow estimation even in ungauged watersheds. By combining machine learning with hydrological modeling, this study significantly advances the accuracy and applicability of SWAT for baseflow analysis in diverse watershed conditions. The developed web-based program allows users to upload input datasets for alpha factor estimation as follows (Figure 4). CSV file includes HRU-specific watershed area ratios, monthly precipitation, soil water holding capacity, and HRU slope. Watershed map (shape file) represents the spatial distribution of HRUs within the study area. Using these data, the program computes HRU-specific alpha factors for each month (January to December) by accounting for spatiotemporal variability. The results are provided in both visual and tabular formats, along with a modified code for seamless integration with the SWAT. To enhance user understanding, the program visually presents the calculated alpha factors for each HRU. The web interface employs structured HTML components, such as the <section> tag and FontAwesome (fa-stack) icons, to intuitively categorize and display alpha factor results. This user-friendly layout allows users to easily interpret and compare the spatiotemporal variability of HRU-specific alpha factors. In a previous study using SWAT, estimating alpha factors in ungauged watersheds was challenging due to the lack of streamflow data. This study overcomes this limitation by enabling reliable alpha factor estimation through the developed web-based program. By utilizing various hydrological inputs, the program ensures accurate predictions of baseflow characteristics, even in ungauged regions.

2.5. Assessment of Recession and Baseflow Estimation in Case 1 and Case 2

SWAT involves numerous parameters associated with hydrologic processes, such as rainfall runoff, that require calibration to ensure accurate prediction. In this study, SWAT-CUP, an automated calibration and uncertainty analysis tool, was used to calibrate streamflow and evaluate recession and baseflow. The Sequential Uncertainty Fitting Ver.2 (SUFI-2) algorithm, which has been widely applied in various studies for its effectiveness in hydrological modeling, was also employed in this study [41,42,43,44,45,46]. In Cases 1 and 2, SWAT-CUP calibrated the streamflow in nine sub-basins and one watershed. In other words, calibration was performed across nine sub-basins and the overall watershed to ensure a reliable representation of hydrological responses throughout the Gapcheon watershed. In Case 1, the conventional method was applied by assigning different alpha factors to each sub-basin, while the same alpha factor was uniformly applied to all HRUs within each sub-basin. This means that although each sub-basin had a unique alpha factor, all HRUs within a given sub-basin shared the same value. This approach allows for the differentiation between sub-basins but does not account for variability at the finer HRU level.

In contrast, Case 2 incorporated the spatiotemporal recession characteristics, as proposed by Lee et al. [12], by applying distinct alpha coefficients for each HRU within the sub-basins. These HRU-specific alpha values were calculated using a web-based program and subsequently applied within the SWAT, allowing for a more detailed representation of the hydrological variability across different HRUs. After calibration using SWAT-CUP, the alpha factors for Case 1 and Case 2 were set as fixed parameters and applied in SWAT to predict recession characteristics. In this study, model performance was evaluated using several statistical metrics, including R², NSE, index of agreement (IOA), mean absolute percentage error (MAPE), and PBIAS [35] (Equations (2)–(6)).

Baseflow, reflecting recession characteristics, was analyzed using Pass1 of BFlow2021 (https://app.envsys.co.kr/bflow2021/) (accessed on 19 February 2025), which separated baseflow from the observed streamflow in the study watersheds. Moreover, the baseflow index (BFI), representing the contribution of baseflow to the streamflow, was computed for each of the study watersheds. To compare the observed streamflow with the baseflow estimated using BFlow2021, the streamflow simulated using SWAT (in both Case 1 and Case 2) was also separated into baseflow components using BFlow2021. Furthermore, the observed recession patterns were extracted and compared with the simulated to assess improvements in baseflow recession predictions between Case 1 and Case 2.

$$R^2={({\displaystyle \frac{{\sum }_{i=1}^n(O_{obs,i}-{\overline{O}}_{obs})\left(S_{sim,i}-{\overline{S}}_{sim}\right)}{\sqrt{\sum _{i=1}^n{(O_{obs,i}-{\overline{O}}_{obs})}^2}\sqrt{\sum _{i=1}^n{(S_{sim,i}-{\overline{S}}_{sim})}^2}}})}^2$$

(2)

$$NSE=1-\left[{\displaystyle \frac{{\sum }_{i=1}^n{(O_{obs,i}-S_{sim,i})}^2}{{\sum }_{i=1}^n{(O_{obs,i}-{\overline{O}}_{obs})}^2}}\right]$$

(3)

$$IOA=1-\left[{\displaystyle \frac{{\sum }_{i=1}^n{(O_{obs,i}-S_{sim,i})}^2}{{\sum }_{i=1}^n{(\left|S_{sim,i}-{\overline{O}}_{obs}\right|+\left|O_{obs,i}-{\overline{O}}_{obs}\right|)}^2}}\right]$$

(4)

$$MAPE={\displaystyle \frac{100}{n}}{\sum }_{i=1}^n\left|{\displaystyle \frac{O_{obs,i}-S_{sim,i}}{O_{obs,i}}}\right|$$

(5)

$$PBIAS=100\times \left[{\displaystyle \frac{{\sum }_{i=1}^n(O_{obs,i}-S_{sim,i})}{{\sum }_{i=1}^n{O}_{obs,i}}}\right]$$

(6)

where $O_{obs,i}$ (m³/s) is the observation, $S_{sim,i}$ (m³/s) is the simulation, ${\overline{O}}_{obs}$ (m³/s) is the mean of the observations, ${\overline{S}}_{sim}$ (m³/s) is the mean of the simulations, and n is the total number of observations.

3. Results and Discussion

3.1. Development and Evaluation of a Web-Based Program for Spatiotemporal Alpha Factor Estimation in SWAT Using Automated Machine Learning (TPOT)

The performance evaluation of TPOT was conducted using established hydrological performance metrics, treating heterogeneous alpha factors as observed values and TPOT-calculated alpha factors as predictions. Among the models tested, the RandomForestRegressor was selected as the optimal model due to its ability to capture non-linear relationships and handle high-dimensional data effectively. The best pipeline was configured with the following hyperparameters: max_features = 0.1502, min_samples_leaf = 18, and min_samples_split = 3, where max_features controls the proportion of features used for splitting, min_samples_leaf sets the minimum samples required in a leaf node to enhance generalization, and min_samples_split defines the minimum samples needed to split an internal node, preventing overfitting. Additionally, random_state = 42 was used to ensure reproducibility. This configuration demonstrated a superior predictive performance compared to the other tested models. To validate the model, the HRU-specific alpha factors estimated using the TPOT were compared with those calculated in the study by Lee et al. [12]. The TPOT was first trained using key hydrological variables such as monthly precipitation, soil water holding capacity, HRU slope, and watershed area ratio. The validation was conducted using data from three watersheds selected from ten study watersheds excluded from the trained model to compare the HRU-specific alpha factors estimated using the TPOT with the HRU-specific alpha factor calculated in the study by Lee et al. [12]. The validation results for Boksu (R² = 0.86, NSE = 0.74, PBIAS = 4.71%), Mannyeon (R² = 0.77, NSE = 0.74, PBIAS = −6.34%), and Gasuwon (R² = 0.88, NSE = 0.86, PBIAS = −4.02%) demonstrated reliable predictive performance. These results indicate strong model reliability, with a high R², acceptable bias (PBIAS), and substantial efficiency (NSE). Specifically, the high R² values across all watersheds suggest that the model effectively captured the variance in observed alpha factors, indicating a solid fit between observed and predicted data. The NSE values, all above 0.7, demonstrate that the model had strong predictive power and efficiently explains the variability of the observed data. Additionally, the PBIAS values falling within an acceptable range indicate minimal systematic overestimation or underestimation, confirming a balanced model performance across different hydrological conditions [34]. After successful validation, a machine learning model was developed using the TPOT, specifically optimizing a RandomForestRegressor to estimate HRU-specific alpha factors. To enhance accessibility, this pre-trained TPOT RandomForest engine was integrated into a web-based program (https://app.envsys.co.kr/Alpha_es/) (accessed on 19 February 2025), allowing users to obtain HRU-specific alpha factor data tailored to their study watersheds. Unlike conventional methods that require observed streamflow measurements for alpha factor estimation, this web-based program allows users to estimate alpha factors even in ungauged watersheds. In the web-based program, users only need to input relevant hydrological variables, and the pre-trained TPOT RandomForest engine automatically estimates the alpha factors without requiring additional model training. The developed web-based program requires users to upload the HRU shapefile from the SWAT model of the study area. Additionally, users input key hydrological variables such as the monthly precipitation, soil water holding capacity, HRU slope, and HRU area ratio. Based on these inputs, the program runs the optimized TPOT RandomForest model to generate monthly HRU-specific alpha factor visualizations (January to December). This allows users to intuitively analyze the baseflow distribution and prioritize areas for management. Notably, significant spatial variations in alpha factors were observed between the January–June and September periods, as highlighted in Figure 5. These variations indicate seasonal fluctuations in baseflow recession across the watershed, influenced by key factors such as precipitation, HRU area, HRU slope, and soil properties. Unlike conventional baseflow estimation models that assume a stationary alpha factor, our approach accounts for spatiotemporal variability in recession characteristics. By leveraging the TPOT’s machine learning capabilities, we effectively capture these variations and estimate HRU-specific alpha factors, leading to a more realistic representation of watershed hydrology. Additionally, the spatial variability in estimated alpha factors underscores the importance of seasonally adjusted recession and baseflow assessments in hydrological modeling.

While the variation in HRU-specific alpha factors observed in Figure 5 may appear small, minor changes can significantly impact baseflow modeling. Slight variations in alpha factor values influence the BFI and recession characteristics, which, in turn, affect groundwater recharge rates and the duration of baseflow contribution to streamflow. Since baseflow is a cumulative process, these variations can lead to notable differences in the total stream discharge over time. Compared to the conventional method, which applies a uniform alpha factor across the watershed, our web-based program calculates tens of thousands of monthly HRU-specific alpha factors. This reduces prediction errors common in traditional approaches, leading to more accurate baseflow modeling and facilitating the implementation of effective water quality management strategies. Furthermore, given the close connection between baseflow and water quality issues—such as nitrate pollution loadings—our approach provides a quantitative framework to proactively assess and manage water quality concerns. Additionally, Figure 5 highlights specific patterns and anomalies in alpha factor trends corresponding to periods of higher variability or unexpected changes in baseflow recession. These anomalies may be driven by seasonal hydrological shifts, extreme weather events, or differences in watershed characteristics. Since this program does not rely on observed streamflow data, it is particularly useful for estimating baseflow characteristics in ungauged watersheds, allowing for alpha factor estimations even in regions without direct observations. This capability extends its applicability to various study areas and enhances its practical value for watershed management. The integration of the TPOT with a web-based program provides a scalable, robust, and user-friendly solution for spatiotemporal alpha factor estimation. This advancement significantly enhances baseflow predictions and supports informed decision-making in watershed management. However, additional model retraining may be necessary to ensure reliable performance when it is applied to regions with significantly different hydrological conditions. To address this limitation, region-specific model adjustments and expanded training datasets incorporating diverse hydrological characteristics are recommended. Future research will incorporate additional environmental variables, such as soil permeability and aquifer properties, to improve the model adaptability across diverse hydroclimatic settings. These refinements will enhance the model’s robustness and applicability in various watershed conditions, even where spatial precipitation variability is not explicitly represented. Furthermore, incorporating a broader range of environmental variables and evaluating model adaptability in various watershed settings will strengthen the model robustness and improve its applicability across different hydrological regimes, even in cases where spatial precipitation variability is not explicitly represented.

3.2. Comparison of Recessions Estimation in Case 1 and Case 2

Before evaluating the baseflow predictions of Case 1 and Case 2, it is essential to verify whether the model accurately simulates streamflow. In Case 1, SWAT-CUP (SUFI-2) calibrated the model for nine sub-basins and the entire Gapcheon watershed, ensuring a comprehensive representation of hydrological processes. As a result, all streamflow stations in the study area were included in the calibration, either at the sub-basin or watershed scale. Since streamflow calibration was performed using Case 1 and subsequently applied consistently to both Case 1 and Case 2, the calibrated results influenced the streamflow simulations in both cases (

Table 3. The model performance results for daily streamflow in the study watersheds.

Study Watershed	R2		NSE		IOA		MAPE (%)		PBIAS (%)
	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2
Yongchon	0.58	0.58	0.57	0.56	0.83	0.83	66.17	65.87	21.54	21.48
Dugye	0.78	0.78	0.73	0.74	0.93	0.93	65.47	65.34	35.26	35.28
Munam	0.56	0.57	0.56	0.56	0.84	0.83	68.97	67.15	9.51	9.02
Inchang	0.57	0.58	0.56	0.56	0.81	0.83	72.51	70.13	9.66	9.14
Gasuwon	0.63	0.63	0.62	0.63	0.88	0.88	57.71	56.98	−8.36	−7.67
Boksu	0.66	0.67	0.62	0.63	0.90	0.91	68.63	66.91	−8.82	−8.03
Hanbat	0.57	0.58	0.58	0.57	0.86	0.86	63.02	61.53	12.92	12.14
Mannyeon	0.66	0.66	0.64	0.65	0.86	0.85	65.16	64.75	7.25	7.13
Daedoek	0.64	0.64	0.57	0.57	0.83	0.64	76.63	75.02	43.98	41.72
Wonchon	0.54	0.54	0.53	0.54	0.84	0.83	67.32	66.81	−0.63	1.44

). To assess model performance, statistical metrics, such as NSE, R², IOA, MAPE, and PBIAS, were used to evaluate streamflow predictions across all ten monitoring stations. Additionally, the streamflow was analyzed to determine whether the calibration process similarly affected both streamflow components. A comparison of the two cases indicates that Case 1 and Case 2 showed similar accuracy. However, as this study focused primarily on the impact of the alpha factor on baseflow recession characteristics, the streamflow calibration results were assessed mainly to ensure their reliability rather than to emphasize significant differences. The results confirmed that the differences in streamflow between the two cases were minimal, further validating the robustness of the calibration process. The analysis results prioritized recession characteristics and baseflow evaluation rather than directly comparing daily streamflow values, as the alpha factor adjustments more directly influence these. Therefore, Table 3 presents the calibrated streamflow metrics for all monitoring stations to maintain consistency within the calibration framework and ensure a meaningful assessment of model performance. According to the Moriasi et al. [34] guidelines for SWAT calibration evaluation, NSE values above 0.5 are considered ‘satisfactory’, which applies to the results shown here. Furthermore, IOA values close to 1 indicate a strong agreement between the model predictions and observed data. Other evaluation metrics, including R², IOA, and PBIAS, confirmed that the model calibration was generally successful. These findings underscore the importance of using multiple evaluation metrics for a comprehensive assessment. For instance, while the Dugye watershed exhibited a high R² value of 0.78 and the highest IOA (0.93) among the ten study watersheds, its PBIAS of 35.26% indicates significant bias. This suggests that despite the strong NSE and IOA values, the model may consistently overestimate or underestimate observed values. Additionally, the baseflow was analyzed using BFlow2021, comparing observed baseflow with SWAT-simulated baseflow to assess recession characteristics.

This study analyzed the baseflow recession of representative upstream and downstream points from the ten stations. In Case 2, where the alpha factor was set for each HRU, the recession simulation predictions demonstrated an improvement over Case 1, as evidenced by a reduction in MAPE errors.

Table 4. Comparison of baseflow recession patterns between Case 1 and Case 2 was performed in the study watersheds.

Study Watershed	R2		NSE		IOA		MAPE (%)		PBIAS (%)
	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2
Yongchon	0.78	0.66	0.44	0.57	0.79	0.83	46.66	37.84	40.45	22.12
Dugye	0.33	0.62	0.17	0.60	0.77	0.86	64.93	50.23	−20.24	−11.46
Munam	0.71	0.73	0.49	0.54	0.78	0.78	33.03	30.26	35.94	25.69
Inchang	0.85	0.75	0.61	0.64	0.81	0.85	39.45	20.17	13.73	9.19
Gasuwon	0.68	0.76	0.62	0.72	0.88	0.90	43.46	35.15	−23.04	−12.50
Boksu	0.63	0.74	0.53	0.62	0.76	0.83	35.87	33.90	32.23	17.40
Hanbat	0.76	0.72	0.62	0.65	0.84	0.87	34.32	30.67	24.57	19.54
Mannyeon	0.56	0.78	0.51	0.66	0.78	0.86	45.21	32.18	−11.48	−8.93
Daedoek	0.51	0.78	0.42	0.51	0.72	0.76	51.20	40.16	−54.28	−25.08
Wonchon	0.78	0.80	0.68	0.74	0.86	0.90	35.56	33.20	12.31	16.70

emphasizes that the MAPE effectively evaluates the prediction accuracy by representing errors as percentages, facilitating the interpretation of error magnitudes regardless of the data scale. Unlike PBIAS, which can be affected by the direction and magnitude of errors, the MAPE focuses on absolute errors, offering a more balanced assessment of model performance across datasets. This characteristic ensures that the MAPE reflects the overall prediction accuracy without being unduly influenced by individual overestimations or underestimations. Furthermore, improvements in model predictions were observed across additional performance metrics. For R², the predictions improved in most watersheds except for Yongchon, Hanbat, and Inchang. In the Wonchon watershed, however, the PBIAS values in Case 2 were worse than those in Case 1, indicating increased bias in the simulated results. The positive PBIAS in Case 2 suggests significant underestimation, warranting further investigation into specific periods or conditions where the model may be underestimated. Despite this, other metrics, such as NSE, R², and IOA, demonstrated improvements in Case 2, indicating that the model better captured the spatial and temporal variability of observed recession patterns (Figure 6). Since the primary focus of this study was accurately simulating recession, the improved NSE, R², and IOA metrics in Case 2 suggest an enhanced model performance.

However, future studies should carefully evaluate this trade-off between metrics to ensure a comprehensive model assessment. These findings highlight the necessity of employing HRU-specific alpha factors instead of using a uniform value across sub-basins. The lack of improvement in the MAPE results and the significant estimation uncertainty are likely attributed to applying a uniform alpha factor from the Case 1 distribution. In this context, the MAPE, which quantifies the model accuracy by measuring how closely predicted values align with observed data, revealed that the predictions still show considerable deviations from actual values. This outcome indicates that applying a uniform alpha factor for the entire period may not effectively capture the variations within the recession process. While this study focused on the SWAT, similar methodologies for optimizing the alpha factor, which is a baseflow recession parameter, can be explored in other hydrological models. Physically based models, such as SWAT-MODFLOW, VIC, and MIKE SHE, can incorporate comparable parameter adjustments, including modifications to the aquifer hydraulic conductivity, specific storage, and specific yield to improve recession estimates. Therefore, future research will focus on adapting this methodology to other hydrological models by identifying analogous parameters, evaluating their sensitivity to the baseflow recession, and developing a generalized calibration approach.

3.3. Comparison of Baseflow Estimation in Case 1 and Case 2

After evaluating the recession simulations from Case 2 compared to Case 1, the baseflow was assessed in both cases. This study enhanced the simulation analysis of baseflows at specific stations by applying HRU-specific alpha factors differentially (Case 2) rather than using a uniform value across all HRUs (Case 1). Consistent with the trends observed in the simulations, the baseflow results of Case 2 showed an improvement in several watersheds compared to Case 1 (

Table 5. Comparison of baseflow results between Case 1 and Case 2 was conducted after extracting baseflow from SWAT streamflow using the BFlow2021.

Study Watershed	R2		NSE		IOA		MAPE (%)		PBIAS (%)
	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2	Case 1	Case 2
Yongchon	0.63	0.61	0.46	0.57	0.81	0.86	58.09	54.21	46.58	22.72
Dugye	0.50	0.56	0.40	0.52	0.82	0.85	57.94	50.58	30.48	18.02
Munam	0.64	0.66	0.58	0.64	0.82	0.86	71.57	61.62	31.63	21.19
Inchang	0.65	0.67	0.63	0.64	0.88	0.89	65.14	64.27	26.92	30.28
Gasuwon	0.67	0.67	0.51	0.52	0.89	0.88	57.51	45.65	14.60	−5.46
Boksu	0.60	0.63	0.59	0.62	0.86	0.87	61.96	56.28	21.85	14.51
Hanbat	0.56	0.58	0.48	0.53	0.84	0.86	59.95	40.32	33.96	26.14
Mannyeon	0.61	0.62	058	0.60	0.87	0.88	59.95	51.22	15.55	−1.05
Daedoek	0.55	0.72	0.48	0.51	0.83	0.90	60.59	27.77	31.59	−27.41
Wonchon	0.66	0.76	0.65	0.74	0.89	0.93	58.65	38.73	16.57	−6.88

). While MAPE values generally decreased, indicating better performance, certain watersheds exhibited mixed trends. In particular, the R² for Yongchon was higher in Case 1 than in Case 2, and the PBIAS for Inchang remained relatively high, suggesting that while the differentiation of alpha factors improved the accuracy in many cases, some inconsistencies still exist. Overall, the results highlight the benefits of HRU-specific alpha factors, as the overall evaluation metrics indicate that Case 2 performed better than Case 1. The study by Lee et al. [12] also highlighted that alpha factor variations among HRUs influence baseflow changes. The BFI value was analyzed based on the measured streamflow of study watersheds using BFlow2021. The BFI represents the proportion of baseflow in the total streamflow and the characteristics of aquifers in the watershed, playing an essential role in determining runoff patterns. Analyzing the BFI for the study watershed’s simulation period revealed that more than 50% of the total streamflow is baseflow (Munam 0.49; Yongchon 0.59; Inchang 0.52; Gasuwon 0.61; Dugye 0.71; Boksu 0.52; Mannyeon 0.55; Hanbat 0.52; Daedoek 0.60; Wonchon 0.52). Among these, Dugye, where the baseflow rate is highest compared to direct runoff, had the largest BFI value, while Munam had the smallest. To determine whether a stream is increasing or decreasing, several hydrological methods can be utilized. In this study, we primarily relied on the BFI analysis and alpha factor estimations to assess the groundwater contributions. Streams exhibiting higher baseflow contributions generally indicate gaining conditions, while lower baseflow proportions may suggest losing conditions. For streams that exhibit losing conditions (influent) during certain periods of the year, the HRU-specific alpha factor approach applied in this study can still be effective. Since this method captures spatiotemporal variations in baseflow recession, it can reflect seasonal transitions between gaining and losing conditions. This detailed analysis of BFI across different watersheds further highlights the importance of HRU-specific alpha factor estimation in improving the baseflow prediction accuracy. Understanding the proportion of baseflow in total streamflow allows for a better assessment of groundwater contributions, which is crucial for evaluating gaining and losing stream conditions. However, additional calibration using groundwater-level observations may enhance accuracy in hydrological systems. Future research will integrate numerical groundwater models (e.g., MODFLOW) with SWAT to refine the determination of gaining and losing conditions, thereby enhancing the predictive accuracy of recession characteristics and baseflow estimation through improved hydrological modeling. Additionally, integrating dynamic groundwater data will provide a more comprehensive approach to assessing seasonal shifts in gaining/losing stream behavior. This study also proposes the use of a web-based program to improve model usability, enabling the prediction of hydrological behavior (e.g., baseflow recession) in ungauged stations with limited observational data. Integrating such web-based programs into decision support systems provides a user-friendly interface accessible even to non-experts, allowing water resource managers to make informed decisions. Besides, the system facilitates better water resource management by delivering efficiently interpretable results. Therefore, this study suggests that integrating automated machine learning into hydrological modeling can enhance the predictive efficiency, accuracy, and usability of the models. However, since this study focused primarily on the methods of alpha factor application considering recession characteristics, other hydrograph components were not fully addressed. While the simulated recession exhibited high predictability, the baseflow predictions did not show significant improvements. Future research should explore additional factors influencing baseflow predictions and evaluate other parts of the hydrograph, such as peak flows and total runoff. This comprehensive approach could further enhance the robustness and applicability of the proposed methods.

4. Conclusions

In this study, efforts were made to overcome the limitations of the conventional method (applying a uniform alpha factor in the SWAT) by calculating the alpha factor while considering both the spatial and temporal variability of recession characteristics and providing results through a web-based program. This approach aimed to improve the recession and baseflow prediction accuracy. Unlike conventional approaches, which apply a temporally and spatially uniform alpha factor, our study explicitly incorporated both spatial and temporal variability by estimating HRU-specific and monthly varying alpha factors. Based on the findings of Lee et al. [12], various factors were utilized as input data for the AutoML machine learning algorithm (TPOT) to train and test the prediction of HRU-specific and temporally varying alpha factors. A web-based program was developed to enhance user convenience and enable efficient baseflow analysis by integrating the machine learning engine (TPOT) into the system, allowing for the efficient application of the trained model. In this study, we developed a web-based program that enables users to visualize the spatiotemporal variability of recession characteristics in the study watershed and receive modified SWAT engine data along with the results. Moreover, since the web-based program does not require users to undergo complex computational processes, even non-experts can efficiently and quickly utilize it. Therefore, this web-based program enables the application of the alpha factor in the SWAT by effectively estimating HRU-specific alpha factors that vary over time, making it widely applicable for analyzing baseflow both domestically and internationally. Overall, this study improved the simulation of the baseflow recession characteristics, thus enhancing the predictability of baseflow simulations compared to conventional methods. Case 2, which applied the HRU-specific and temporally varying alpha factor, showed improved baseflow recession and baseflow predictions with reduced error rates compared to Case 1, as evidenced by various indicators such as the R² value and the MAPE. Furthermore, it demonstrated the advantage of integrating spatiotemporal variability in recession characteristics over the conventional uniform alpha factor approach. These findings provide crucial data and can be used to understand the organic relationship between direct runoff and groundwater. Moreover, given the close connection between baseflow and water quality issues, such as nitrate pollution loadings, spatiotemporal alpha factor estimation can offer quantitative assessments and support the proactive management of water quality concerns, including regarding nitrate contamination. In this study, the HRU-specific alpha factor was calculated, but in future research, more dynamic, event-based alpha factor estimation will be conducted to consider the temporal characteristics of recession more accurately. Moreover, factors such as soil hydraulic conductivity and organic matter content will be further enhanced and considered to improve the predictability of the TPOT monthly alpha factor using HRU calculation results for more precise spatiotemporal alpha factor estimation.