Calibration and validation strategies of SWAT+ for streamflow simulation: a systematic review

Highlights

• PRISMA-based review of SWAT+ streamflow calibration and validation strategies.
• SWAT+ Toolbox is widely used but lacks multi-objective calibration.
• Highlights multi-site and multivariable calibration for reliable model simulations.
• Shows increasing adoption of KGE alongside NSE evaluation.

1 Introduction

Accurate hydrological process modeling is essential for sustainable water resource management and policy development under changing land use and climate conditions (Du et al., 2025a; Tigabu et al., 2024). Hydrological models are critical for quantifying water balance components and assessing the spatial and temporal effects of land use and climate change on catchment hydrology (Jahanshahi et al., 2025; Mishra et al., 2025; Tarkegn et al., 2025a). Among these models, the Soil and Water Assessment Tool (SWAT) has been widely used worldwide over the past few decades (Arnold et al., 2012). The model was improved to better simulate complex hydrological processes, leading to the development of a fully revised version called SWAT+ (Bieger et al., 2017). SWAT+ model included several key advancements, including better differentiation between upland regions and floodplains (Chawanda et al., 2020; N’guessan et al., 2024), more flexible model configuration, dynamic reservoir operations, and customizable aquifer boundaries, enabling more accurate simulation of complex hydrological processes (Barresi Armoa et al., 2024; Tapas, 2024). These improvements have positioned the model as one of the most comprehensive tools for diverse applications (Mohamed et al., 2025; Tapas et al., 2025; Tarkegn et al., 2025b).

The model has been applied across diverse geographic and climatic settings, with key modeling procedures such as calibration and validation ensuring that simulated hydrological processes accurately represent observed watershed behavior (Abbas et al., 2024; Harifidy et al., 2024). Calibration is the process of adjusting the most sensitive parameter’s value to better represent the model for a given set of local conditions. Whereas validation is the process of comparing model results with independent data without modifying the previously adjusted parameters (Arnold et al., 2012). Effective calibration and validation are vital for reducing model uncertainty and increasing confidence in its simulation capability for both present and future conditions (Chen et al., 2024; Harrison et al., 2022; Yen et al., 2019).

The growing number of studies using SWAT+ for hydrological simulations has relied on diverse data sources, calibration and validation tools, and calibration strategies, all of which strongly influence model accuracy (Cecconello et al., 2025; Jiménez-Navarro et al., 2024; Tigabu et al., 2024). These studies have used DEM, land use/land cover (LULC), and soil datasets from different sources and at varying resolutions. Although streamflow data remain the most widely used input for SWAT+ calibration, some studies have incorporated multiple variables to better represent hydrological processes (Arnold et al., 2012; Jiménez-Navarro et al., 2024; Mohamed et al., 2025). Calibration approaches also vary widely, ranging from fully automated algorithms to manual trial-and-error methods and hybrid techniques (Harifidy et al., 2024; Nguyen et al., 2022; Yen et al., 2019).

This systematic review is motivated by the rapidly growing use of SWAT+ in hydrological modeling. As an updated and more advanced version of SWAT, SWAT+ has been widely adopted; however, no comprehensive review has specifically examined calibration and validation strategies for streamflow simulation, which underpins other SWAT+ modeling applications. Therefore, this study aims to (1) systematically identify the main and alternative input data sources, calibration datasets, calibration and validation software/tools, and recent advances in calibration approaches; and (2) provide structured recommendations to improve reproducibility and reduce model uncertainty in future SWAT+ applications.

2 Methodology

2.1 Databases and search strategy

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines to ensure a comprehensive and transparent methodology and to evaluate the quality of the studies (Page et al., 2021).

The literature search was conducted across two major bibliographic databases, including Scopus, WOS, and the SWAT model literature databases, which were selected for their global reach and scientific credibility in environmental science and hydrology. For Scopus and WOS, search keywords were systematically developed using Boolean operators, truncation, and phrase matching to optimize both sensitivity and specificity. The final search string was (“Soil and Water Assessment Tool plus” OR “SWAT+*” OR “SWAT plus” OR “Revised version of SWAT”) AND (“Streamflow” OR “surface runoff” OR “hydrology” OR “water flow”) AND (“modeling” OR “simulation” OR “model performance” OR “calibration” OR “validation” OR “parameter estimation” OR “uncertainty analysis”).

The literature search covered the period from 2017 to 2025 to align with the initial release and subsequent adoption of SWAT+. Only peer-reviewed original research articles published in English were included, and early access articles were excluded. To ensure comprehensive coverage of relevant studies, SWAT+ records were also manually extracted from the SWAT+ literature database.

2.2 Inclusion and exclusion criteria, and data extraction

This review focused exclusively on hydrological simulation studies that used SWAT+, the updated version of the SWAT model, to evaluate model calibration and validation strategies. Peer-reviewed original research articles were included if they used streamflow data to calibrate and validate the SWAT+ model, even when they included additional analyses. Articles were excluded if they: (1) focused on the older SWAT model, (2) used a coupled SWAT+ model, or (3) did not conduct calibration and validation.

The review protocol, including the predefined inclusion and exclusion criteria and data extraction framework, was prospectively registered on the Open Science Framework.¹ The full-text screening were conducted by the corresponding author (B. F. A.) and co-author (R. A.) according to the predefined eligibility criteria. Any disagreements regarding study eligibility were resolved through discussion between the authors.

A total of 227 articles were retrieved from Scopus, Web of Science, and the SWAT literature databases. The records were exported to Zotero, where 92 duplicates were automatically removed. The remaining 136 studies were screened by title and abstract and assessed in full text by the corresponding author using predefined eligibility criteria. Of these, 35 articles were excluded during title and abstract screening because they were not relevant to the review objective. Ultimately, 101 studies underwent full-text review, and 21 were excluded with reasons based on the inclusion and exclusion criteria. Finally, 80 studies were included in this systematic review, and data were extracted. The complete search strategy and inclusion and exclusion process are summarized in the PRISMA flow diagram (Figure 1).

To ensure methodological rigor and reproducibility, all included studies underwent quality assessment using a structured checklist (Supplementary Table S1). The assessment showed that the studies clearly documented streamflow data sources, formal calibration procedures, independent validation, and the use of multiple model performance metrics. However, minor reporting gaps were identified in relation to methodological reproducibility, particularly the explicit reporting of HRU definition thresholds and the methods used to calculate ETa. All included studies achieved a cumulative quality score above 70%, providing a reliable methodological evidence base for this synthesis.

Data was systematically extracted from each eligible study to capture key bibliographic, geographic, and methodological information. Bibliographic details included authors, publication year, country, and total study area coverage. For model configuration, we extracted information on the source and spatial resolution of geospatial inputs, including DEM, LULC, soil maps, and climate datasets. To support reproducibility, HRU thresholds and the equations used for ETa estimation were documented as part of the quality assessment criteria. Calibration data sources were also recorded to determine whether studies relied only on streamflow or included additional variables. Temporal information, including the calibration time step (daily or monthly), warm-up period, and calibration and validation durations, was extracted. Finally, details on calibration software, optimization algorithms, objective functions, and reported performance metrics were compiled.

For studies that included multiple watersheds, calibration scenarios, algorithms, or performance metrics, each independent modeling case was treated and analyzed as a separate unit rather than as a single article-level observation. Accordingly, the frequencies reported in the figures and tables represent the number of modeling cases, providing a more accurate depiction of methodological patterns across the reviewed studies. The extracted data were analyzed using Excel and R to generate descriptive statistics, assess publication trends, and produce a geographic distribution map. The PRISMA flow diagram were generated using Biblioshiny and organized with the support of AI.

3 Results and discussion

3.1 Publication trend and geographical distribution of the included studies

To examine the evolution of SWAT+ applications in streamflow simulation, a temporal analysis of publication trends from 2017 to 2025 was conducted (Figure 2a). During the initial years following the release of SWAT+, the number of published studies remained limited, with early applications primarily focusing on model development, testing, and field-scale evaluation by the model developers (Bieger et al., 2017, 2019). As illustrated in Figure 2a, publication output increased markedly from 2021 onward, indicating the progressive adoption of SWAT+ by the wider hydrological modeling community.

The geographical distribution of publications indicates that Ethiopia, China, and the United States contributed the largest number of SWAT+ streamflow simulation studies (Figure 2b). Additional SWAT+ streamflow studies have been conducted in other regions, including the Beas Basin in India (Singh et al., 2023), the Longaví Catchment in Chile (Moya et al., 2022), the Karnali River Basin in Nepal (Pradhananga et al., 2025), central Morocco (Alitane et al., 2022), a Mediterranean watershed in Italy (Pulighe et al., 2021), and the Philippines (Apdohan et al., 2025), highlighting the global applicability of the model. This distribution likely reflects a combination of factors, including strong research capacity in hydrological modeling, national investments in watershed management and water resources research, the availability of long-term hydro meteorological monitoring data, and increasing efforts to address water security and climate change adaptation.

3.2 Data inputs and their impact on SWAT+ model performance

The reliability and accuracy of SWAT+ simulations are fundamentally determined by the quality and spatial resolution of the input datasets, specifically DEM, LULC, soil, and climate data, (Abbaspour et al., 2019; Bieger et al., 2017). The selection of these data sources not only influences the physical representation of the watershed but also dictates the success of subsequent model calibration (Rasheed et al., 2024).

3.2.1 Digital elevation model

The influence of DEM resolution on streamflow simulation was found to be strongly scale-dependent. Across the reviewed studies, the 30 m SRTM DEM was the most frequently adopted elevation dataset, followed by the 90 m SRTM and 30 m ASTER GDEM (Table 1). In small and topographically complex watersheds, finer-resolution DEMs improved the representation of drainage networks, slope gradients, flow paths, and watershed boundaries (Ahmad et al., 2025). In contrast, the advantages of higher-resolution DEMs were generally less pronounced in larger or relatively flat catchments, where the influence of local topographic variability on runoff generation is reduced. Several studies also reported that the 30 m SRTM DEM produced more reliable stream network delineation than the 30 m ASTER GDEM, primarily because of its lower vertical error and reduced topographic artifacts (Sharma and Tiwari, 2014; Tan et al., 2015). Furthermore, the prevalence of SRTM 30 m over coarser 90 m datasets indicates the need for finer topographic representation to capture hydrological heterogeneity (Rasheed et al., 2024).

Table 1. Frequency of data types used for SWAT+ model setup in the reviewed streamflow simulation studies.

The table summarizes the occurrence of the different input datasets required for model configuration, indicating how frequently each data type was used across the reviewed studies.

In contrast, studies conducted in large river basins reported limited improvements in streamflow performance from using very high-resolution DEMs. At these spatial scales, streamflow dynamics are primarily governed by climate forcing and basin characteristics rather than fine-scale topographic variability (Dile et al., 2024; Tapas, 2024). Consequently, hydrologically conditioned datasets, such as the 90 m MERIT DEM, provided satisfactory model performance while substantially reducing computational requirements (Mishra et al., 2025). These findings demonstrate that DEM resolution alone does not determine SWAT+ performance; instead, its influence depends on watershed scale, terrain complexity, and the hydrological processes being represented.

3.2.2 Soil and land use/land cover data

Land use/land cover (LULC) and soil datasets define the spatial heterogeneity of HRUs and therefore strongly influence the representation of hydrological processes in SWAT+ (Bieger et al., 2017). The reviewed studies employed a wide range of national and global LULC and soil datasets with varying spatial resolutions (Table 1). However, relatively few studies systematically evaluated the influence of alternative datasets on model performance, indicating that the impacts of input data selection remain insufficiently quantified (Al-Khafaji et al., 2020).

The studies indicate that the influence of LULC datasets on SWAT+ performance depends primarily on watershed heterogeneity rather than spatial resolution alone. In small and fragmented catchments, high-resolution LULC datasets consistently improved HRU delineation by preserving fine-scale landscape variability, resulting in more realistic simulations of runoff generation and evapotranspiration (Ait M’Barek et al., 2023; Al Khoury et al., 2023). However, these advantages diminish in larger watersheds where spatial aggregation effects allow coarser products to achieve comparable predictive performance while substantially reducing computational demand. Notably, the choice between different high-resolution sources can yield systematic differences; for instance, the FORM 10 m dataset has been shown to outperform ESA 10 m in specific streamflow simulations (Zhang et al., 2025).

The choice between static and multi-period LULC datasets is a critical strategy for reducing parameter uncertainty in long-term simulations (Kiprotich et al., 2021; Gong et al., 2024). While static maps are often assumed sufficient if climate is the main driver, utilizing dynamic datasets allows the model to mechanistically account for land-use transitions like urbanization or deforestation (Ayalew et al., 2023). This temporal consistency ensures that the CN2 dynamics reflect the evolving physical reality of the basin, preventing the model from over-relying on climate forcing to explain trends actually caused by land surface changes (Kiprotich et al., 2021; Jahanshahi et al., 2025).

Similar patterns were observed for soil datasets. Most included studies employed national soil databases where available, whereas others relied on global products such as the Harmonized World Soil Database (HWSD) and the FAO/UNESCO Soil Map of the World (Table 1). High-resolution soil datasets generally enhanced the representation of spatial variability in soil texture and hydraulic properties, leading to more realistic HRU parameterization and potentially reducing structural uncertainty (Bouslihim et al., 2025; Fernando et al., 2025). In contrast, coarse-resolution global soil datasets often achieved satisfactory streamflow performance after calibration but were less effective in representing internal hydrological processes, such as actual evapotranspiration and groundwater recharge, due to the loss of spatial heterogeneity and increased parameter equifinality (Noa-Yarasca et al., 2025).

3.2.3 Climate data

More than half of the reviewed studies used ground-based weather station data as climate inputs for SWAT+ simulations. However, these datasets often lack the spatial coverage needed to capture localized rainfall variability, particularly in complex terrain and data-scarce regions (Table 1; Ahmad et al., 2025). Consequently, the use of gridded climate products, such as CHIRPS, CFSR, and ERA5, has increased in SWAT+ applications because they provide spatially and temporally consistent meteorological information. Similarly, Kassaye et al. (2024a) employed bias-corrected CMIP6 climate data in the Baro River Basin, Ethiopia, demonstrating the utility of global climate model outputs for future hydrological projections in data-scarce regions. Similarly, Tumsa et al. (2022) quantified LULC impacts on water balance in the Guder catchment, Ethiopia, using SWAT+ with a time-series-based calibration approach, reporting NSE values of 0.70–0.79 during calibration and 0.86 during validation. These products can serve as useful alternatives where ground-based observations are limited, although their performance should be evaluated and biases corrected before their application (Abate et al., 2024; Abate et al., 2023; Eini et al., 2023; Kardhana et al., 2024). For instance, the accuracy of precipitation inputs serves as the fundamental control on the model’s ability to replicate physical rainfall-runoff transformations (Mohamed et al., 2025).

Comparative studies have shown that the performance of gridded climate datasets varies considerably across regions, climatic conditions, and methodological frameworks. In Ethiopia’s Upper Danakil Basin, CHIRPS showed the strongest agreement with observed rainfall (R² = 0.853), outperforming CFSR (R² = 0.688) and ERA5 (R² = 0.494) (Abate et al., 2023; Du et al., 2025b). Similarly, in Madagascar, CHIRPS produced better streamflow simulations than CFSR across 12 basins, achieving satisfactory performance (NSE > 0.45) in five basins compared with three basins for CFSR (Harifidy et al., 2024). However, in Malaysia’s Kelantan River Basin, APHRODITE showed the best performance for monthly streamflow simulation (NSE = 0.61–0.71; R² = 0.74–0.76), whereas CHIRPS performed poorly during validation (NSE = −0.27), ERA5-Land showed limited performance (calibration NSE = 0.45; validation NSE = 0.30), and NASA POWER achieved moderate skill (calibration NSE = 0.65; validation NSE = 0.45) (Du et al., 2025b).

Similarly, Eini et al. (2022) reported that PERSIANN-CDR achieved NSE values ranging from 0.32 to 0.68 across Polish watersheds, while PERSIANN-CCS exhibited considerably lower performance (NSE = −0.11 to 0.41), underscoring substantial intra-product variability. Kardhana et al. (2025) further demonstrated that raw GSMaP precipitation resulted in unsatisfactory streamflow simulations (NSE < 0.20); however, after bias correction, model performance improved markedly, reaching NSE > 0.60 and approaching the performance obtained using ground-based observations. Adding further complexity, Tapas et al. (2025) showed that GPM IMERG achieved the highest efficiency under GLUE calibration (NSE = 0.50; KGE = 0.68), whereas ERA5 and GridMET performed best when calibrated using DDS, indicating that the suitability of climate datasets is also influenced by the calibration approach.

3.3 Calibration and validation software/tools

Calibration and validation of SWAT+ can be conducted using manual, automated, or hybrid approaches that combine expert judgment with optimization algorithms (Kardhana et al., 2024). Several software packages are available to support sensitivity analysis, parameter optimization, and model evaluation, including SWAT+ Toolbox, SWATplus-CUP, and IPEAT+. More recently, programming interfaces in R and Python have expanded calibration capabilities by enabling reproducible workflows, parallel computing, and integration with advanced optimization algorithms (Schürz, 2019).

SWAT+ Toolbox was the most frequently applied open-source calibration platform among the reviewed studies, most commonly in combination with the DDS algorithm (Table 2; N’guessan et al., 2024; Zhang et al., 2025). Additional applications of the SWAT+ Toolbox include studies in the Baro River Basin, Ethiopia (Kassaye et al., 2024b), demonstrating its widespread adoption in data-scarce regions. Bihonegn and Gebeyehu Awoke (2025) further evaluated spatiotemporal sediment yield variability in the Upper Blue Nile Basin, Ethiopia, complementing their earlier work in the Upper Awash Basin (Bihonegn and Gebeyehu Awoke, 2023). DDS is well suited to high-dimensional optimization problems because it efficiently explores large parameter spaces while reducing computational demands (Gao et al., 2025). Comparative evaluations indicate that SWAT+ Toolbox coupled with DDS can achieve robust calibration and validation performance. For example, Cecconello et al. (2025) reported monthly calibration statistics of NSE = 0.69 and KGE = 0.78, together with daily validation statistics of NSE = 0.74 and KGE = 0.78, outperforming R-SWAT (SUFI-2) under the conditions evaluated. Similarly Harifidy et al. (2024) found that SWAT+ Toolbox (DDS) produced more stable and consistent automatic calibration results than R-SWAT using LHS.

Table 2. Classification of the reviewed SWAT+ streamflow simulation studies by calibration strategy, software/tools, optimization algorithms, and representative references.

Full reference are provided in Supplementary Table S2. Frequencies of software/tools and optimization algorithms represent independent modeling cases; comparison scenarios were treated as separate cases.

Another SWAT+ calibration software program is SWATplus-CUP, an extension of SWAT-CUP to support calibration and uncertainty analysis. It integrates optimization algorithms such as Sequential Parameter Estimation (SPE), formerly known as SUFI-2, and particle swarm optimization (PSO) (Abbaspour, 2022). SWATplus-CUP is specifically designed to address the high parameterization demands of SWAT+ streamflow simulation (Asadi et al., 2024; Tarkegn et al., 2025a). Although its use in the reviewed studies was limited, this does not indicate lower calibration capability, as software adoption is also influenced by licensing requirements, computational resources, and user familiarity (Figure 3).

Integrated Parameter Estimation and Uncertainty Analysis Tool Plus (IPEAT+) is a freely available calibration and uncertainty analysis platform developed specifically for SWAT+ (Yen et al., 2019). It incorporates probabilistic parameter estimation and uncertainty analysis methods to evaluate parameter sensitivity and predictive uncertainty. This tool application is limited in the literature, likely due to its higher computational demand and the dominance of more established tools such as the SWAT+ Toolbox and SWATplus-CUP (Kiprotich et al., 2021; Naik and Abiodun, 2024; Zhu et al., 2023).

In addition to dedicated calibration software, an increasing number of studies have adopted programming-based workflows using R and Python (Table 2). R-SWAT is an open-source R-based tool with a graphical user interface (GUI), whereas packages such as SWATplusR, SWATrunR, and SWATdoctR support script-based workflows for automated model execution and parameter optimization(Ayalew et al., 2024; Estrada et al., 2024; Schürz, 2019). In particular, SWATrunR enables users to develop customized calibration workflows tailored to specific modeling objectives (Panique-Casso et al., 2025; Sánchez-Gómez et al., 2025). Other studies have employed general-purpose frameworks such as the FME package for sensitivity analysis and uncertainty quantification(Mahmoodi et al., 2024; Tigabu et al., 2025), while SPOTPY has been used for SWAT+ sensitivity analysis and calibration through Python-based optimization workflows, (Barresi Armoa et al., 2024; Barresi Armoa et al., 2023). Collectively, these studies indicate a growing shift toward flexible, script-based calibration environments that facilitate customized and reproducible SWAT+ workflows.

PEST (Parameter Estimation) is a model-independent optimization framework widely used for parameter estimation and uncertainty analysis in environmental models (Doherty, 2015). Within the reviewed literature, its application to SWAT+ remains limited but demonstrates its potential as an alternative calibration framework. Alp et al. (2022) evaluated three optimization algorithms implemented within PEST Levenberg Marquardt (LM), Shuffled Complex Evolution (SCE), and Covariance Matrix Adaptation Evolution Strategy (CMAES), and found that the SCE and CMAES algorithms outperformed LM, achieving NSE values of 0.56 during calibration and 0.38 during validation. Similarly, Tapas et al. (2025) reported that calibration performance varied with both the optimization algorithm and the precipitation dataset, with GLUE producing the highest efficiencies (NSE = 0.50, KGE = 0.68) when combined with GPM IMERG, whereas DDS performed slightly better with ERA5 and GridMET precipitation. In contrast, LHS consistently produced lower model efficiencies. These findings indicate that calibration performance depends not only on the optimization algorithm but also on its interaction with the precipitation forcing and watershed characteristics.

Beyond the choice of calibration software, several studies have adopted hybrid calibration strategies that combine automatic optimization with subsequent manual refinement to further improve model performance (Jiménez-Navarro et al., 2021; Mohamed et al., 2025; Wendell et al., 2024). Although this approach can enhance calibration accuracy by incorporating hydrological expertise, it may also introduce subjectivity and reduce reproducibility if manual adjustments are not adequately documented. Therefore, researchers should explicitly report the manually adjusted parameters, their calibration ranges, and the criteria used to terminate the refinement process to ensure transparency, reproducibility, and scientific rigor.

3.4 Calibration and validation strategies

The selection of an appropriate calibration strategy is primarily determined by watershed scale, spatial heterogeneity, and the objectives of the modeling study (Noa-Yarasca et al., 2025). In small catchments, calibration commonly focuses on accurately reproducing different flow regimes, particularly low-flow and baseflow dynamics (Kardhana et al., 2024). In contrast, large and hydrologically heterogeneous basins require additional spatial constraints through multi-site and multi-variable calibration to better represent internal watershed processes. Furthermore, Smit et al. (2024) used hydropedological insights for improved calibration, while Sharma et al. (2024) assessed drainage water management applications using SWAT+.

Streamflow is the primary observational variable used for SWAT+ calibration (Table 3). However, calibrating the model using only outlet streamflow in large basins may produce acceptable statistical performance while failing to adequately represent internal watershed processes (Abbaspour et al., 2019; Abate et al., 2023; Mohamed et al., 2025). Furthermore, Yulianti et al. (2025) demonstrated the added value of isotope data for improving water balance simulations in a pastoral catchment using SWAT+, while Julich et al. (2022) evaluated the impact of overgrazing on water fluxes in a semi-arid watershed. This limitation reflects the problem of parameter equifinality, whereby multiple parameter sets generate similar outlet hydrographs despite unrealistic representations of internal hydrological processes (Mliyeh et al., 2026; Shah et al., 2021). Consequently, incorporating observations from multiple gauging stations and complementary hydrological variables provides stronger constraints on model parameterization.

Table 3. Summary of calibration and validation strategies used in the reviewed SWAT+ streamflow simulation studies, showing the calibration approach, variables, data sources, and representative references.

Where, ETa, Actual evapotranspiration = Soil moisture; HI, Hydrological index; SWE, Snow Water Equivalent; GLEAM, Amsterdam Global Terrestrial Evaporation Model; GloFAS, Global Flood Awareness System; MOD16A2, MODIS/Terra + Aqua Global Terrestrial Evapotranspiration 1-Day L4 Global 500 m SIN Grid; CCI SM, Climate Change Initiative Soil Moisture; USGS, United States Geological Survey.

Despite the widespread use of streamflow-only calibration, relatively few studies adopted multi-variable calibration by integrating streamflow with evapotranspiration, satellite-derived soil moisture, or hydrological indices(Abate et al., 2024; Abate et al., 2023; Eini et al., 2023; Jiménez-Navarro et al., 2024; Mohamed et al., 2025; Rodríguez-Castellanos et al., 2025). This approach provides independent constraints on different components of the hydrological cycle, thereby improving model realism and reducing uncertainty associated with internal watershed processes, particularly in large and spatially heterogeneous basins (Barresi Armoa et al., 2024; Mohamed et al., 2025). The increasing availability of remotely sensed and reanalysis products has facilitated the use of these complementary datasets, particularly in data-scarce regions where ground-based observations are limited or unavailable (Mliyeh et al., 2026).

Despite these advantages, multi-variable calibration also presents methodological challenges. Integrating remotely sensed and reanalysis datasets requires careful consideration of data uncertainty and potential inconsistencies between observational products and model simulations, which may influence parameter estimation and calibration performance (Eini et al., 2023; Mohamed et al., 2025).

3.5 Calibration and validation approach

Beyond the selection of calibration software and variables, the reviewed studies adopted several calibration and validation approaches to improve model performance and better represent watershed processes. One widely adopted approach is multi-site calibration, in which model parameters are optimized simultaneously using observations from multiple gauging stations rather than only at the watershed outlet (Barresi Armoa et al., 2023; Tapas et al., 2025). This strategy better captures the spatial heterogeneity of rainfall, soils, and LULC, while reducing parameter equifinality by constraining parameter values across multiple locations (Abbaspour et al., 2019; Sánchez-Gómez et al., 2025). Although particularly beneficial for large and heterogeneous watersheds, multi-site calibration also increases calibration complexity (Manawko Tefera and Kasiviswanathan, 2023). In addition, Similarly, some studies employed multi-period calibration using dynamic LULC datasets to better represent long-term hydrological changes associated with land-use change and climate variability (Estrada et al., 2024; Kiprotich et al., 2021; Xiao et al., 2023).

Another important aspect is the calibration time scale, involving monthly or daily increments (Liqun et al., 2025; Setti et al., 2022). Comparative SWAT+ studies have shown that calibration at the monthly time scale consistently yields higher performance metrics than daily calibration (Cecconello et al., 2025; Estrada et al., 2024; Tan et al., 2023). This reflects the smoothing effect of temporal aggregation, which reduces short-term runoff variability and compensates for discrepancies in rainfall timing and runoff generation, leading to higher values of NSE, KGE, and R² than daily calibration (Al Khoury et al., 2023; Cecconello et al., 2025). Conversely, daily calibration remains essential for flood prediction and peak-flow analysis, as it avoids the masking of temporal errors inherent in aggregation (Arnold et al., 2012; Ayalew et al., 2023). Consequently, several recent studies adopted hierarchical calibration strategies in which monthly calibration was followed by daily refinement, preserving long-term water balance while improving the simulation of daily hydrograph dynamics (Castellanos-Osorio et al., 2023).

Another important consideration is the temporal sequencing of calibration and validation datasets. Most reviewed studies followed the conventional practice of calibrating SWAT+ using earlier streamflow records and validating the model with more recent observations to assess predictive performance under temporally independent conditions (Lorente-Ramos et al., 2025). However, several recent studies prioritized data quality over chronology by using more recent, higher-quality datasets for calibration (Tan et al., 2023; Tarkegn et al., 2025a, 2025b; Tigabu et al., 2025; Tigabu et al., 2024). This does not imply that more recent observations are inherently superior. Instead, calibration datasets should be selected based on data completeness, consistency, and measurement reliability, as these characteristics have a direct influence on parameter estimation.

3.6 Objective functions

Objective functions are statistical metrics used to quantify the agreement between simulated and observed hydrological variables, such as streamflow, evapotranspiration, and sediment, during model calibration (Nash and Sutcliffe, 1970; Arnold et al., 2012). Because objective functions guide parameter optimization, their selection directly influences calibration outcomes and the reliability of SWAT+ simulations (Gupta et al., 2009; Il Choi, 2024). Among the reviewed studies, Nash-Sutcliffe Efficiency (NSE) (Eq. (1) was the most frequently used objective function, followed by the Kling-Gupta Efficiency (KGE) (Eq. (2), whereas relatively few studies adopted multiple objective functions simultaneously (Figure 4).

The predominance of NSE across the reviewed studies reflects its long-standing role as the standard objective function for hydrological model calibration. However, the synthesis also highlights its inherent limitations. Because NSE is based on squared residuals, it disproportionately emphasizes high-flow events, making model performance strongly dependent on peak discharge simulation while reducing sensitivity to low-flow conditions and seasonal baseflow dynamics (Gupta et al., 2009; Knoben et al., 2019). Therefore, satisfactory NSE values do not necessarily guarantee an accurate representation of watershed processes or water balance, particularly in catchments characterized by intermittent streamflow and strong seasonal variability (Gupta et al., 2009; Knoben et al., 2019; Llanos-Paez et al., 2023).

In contrast, KGE has gained increasing attention because it evaluates model performance through correlation, bias and variability simultaneously, allowing researchers to distinguish whether model deficiencies arise from timing errors, systematic bias or incorrect representation of flow variability (Gupta et al., 2009). The increasing adoption of KGE observed in the reviewed studies therefore reflects a broader shift from purely statistical goodness-of-fit assessment toward a more process-oriented evaluation of hydrological model performance.

Analysis of the reviewed literature further revealed that objective-function selection should reflect the dominant hydrological characteristics and modelling objectives of the watershed rather than relying on a universal criterion. Watersheds characterized by intermittent streamflow and pronounced seasonal variability increasingly adopted KGE or logarithmic NSE to improve the representation of low-flow dynamics and water-balance processes (Estrada et al., 2024; Leone et al., 2024). Conversely, studies targeting multiple hydrological processes, such as streamflow and evapotranspiration, more frequently employed multivariable or multi-objective calibration strategies that combined complementary objective functions to constrain different components of watershed response (Jiménez-Navarro et al., 2021; Tigabu et al., 2025).

Overall, the increasing use of complementary objective functions demonstrates a gradual shift from single-objective toward multi-objective calibration frameworks in SWAT+. Rather than maximizing a single goodness-of-fit statistic, these approaches seek to improve parameter identifiability, reduce equifinality, and enhance the representation of multiple hydrological processes, thereby improving the physical realism, robustness, and transferability of SWAT+ simulations (Knoben et al., 2019)

NSE = 1 − ∑ i = 1 n ( Y i obs − Y i sim ) 2 ∑ i = 1 n ( Y i obs − Y obs mean ) 2 (1)

KGE = 1 − ( r − 1 ) 2 + ( α − 1 ) 2 + ( β − 1 ) 2 (2)

where n is number of observations Y_i^obs is the observed streamflow at time i, while Y_i^sim is the simulated streamflow at time i. r linear correlation coefficient between simulated and observed values, α is measure of relative variability in the simulated and observed values, β is the ratio between the mean simulated and mean observed flows.

3.7 Model performance evaluation metrics

In SWAT modeling, statistical metrics are used as objective functions (section 3.5) to optimize model parameters during calibration and subsequently to evaluate simulation accuracy during validation using independent datasets. Across the reviewed studies, NSE was the most frequently reported performance metric (92.5%), followed by PBIAS (80%), R² (57.5%), and KGE (38.8%), whereas RSR, RMSE, MSE, MAE, and the correlation coefficient (r) were used less frequently (Figure 5). The most common combination of metrics was NSE, R², and PBIAS (25% of studies); followed by NSE, R², PBIAS, and RSR (10%), NSE, PBIAS, KGE, and RSR (6.5%), and several other combinations that each appeared in less than 5% of the reviewed studies. Approximately one-quarter of the studies adopted unique combinations of performance metrics, including NSElnQ, MSE, MAE, r, bR², and SPAEF, to evaluate specific aspects of model performance (Castellanos-Osorio et al., 2023; Du et al., 2025b; Eini et al., 2023). In addition, 7.5% of the studies supplemented, and 6.3% replaced, NSE with KGE, typically in combination with PBIAS and R² (Attar et al., 2025; Tarkegn et al., 2025a).

Because each performance metric evaluates a different aspect of model behavior, no single metric is sufficient to assess overall model reliability. PBIAS quantifies volumetric error and is therefore essential for assessing water-balance accuracy, whereas R² measures the strength of the linear association between observed and simulated values without indicating whether simulations are unbiased (Moriasi et al., 2007). KGE complements these metrics by simultaneously evaluating correlation, bias, and variability, while NSE remains useful for assessing overall agreement between simulated and observed streamflow (Gupta et al., 2009).

Based on the evidence synthesized from the reviewed studies, no single performance metric is appropriate for all modelling objectives. Instead, complementary metrics should be selected according to the dominant hydrological process and simulation purpose. Table 4 summarizes a proposed minimum set of performance metrics for different SWAT+ applications. These recommendations are derived from the collective evidence presented in this review and established hydrological evaluation guidelines in literatures.

Table 4. Summary of recommended objective functions and performance metrics under different modelling objectives.

3.8 The overall SWAT+ model performance for streamflow simulation

To better understand the overall performance of the SWAT+ model across the included studies, the values of commonly used performance metrics (NSE, KGE, R², PBIAS, and RSR) were analyzed using a comprehensive statistical summary, including the minimum, Quartile 1 (Q1, the 25th percentile), Quartile 2 (Q2, the median or 50th percentile), and Quartile 3 (Q3, the 75th percentile), and maximum values. The analysis was conducted separately for daily and monthly time steps during both calibration and validation periods (Table 5). Because the compiled studies represent highly heterogeneous modeling conditions, including differences in hydro climate, basin size, flow regulation, monitoring density, and calibration period length, the pooled statistics should be interpreted as general performance benchmarks rather than direct model comparisons. Higher values of NSE, R², and KGE indicate better performance, whereas lower values of RSR and the absolute PBIAS values indicate better performance based on range in Supplementary Table S3 (Moriasi et al., 2007, 2015).

Table 5. Summary statistics of SWAT+ model performance metrics across the reviewed streamflow simulation studies during calibration and validation.

Cal, Calibration, and Val, Validation. Number of study represent independent modeling cases, with each distinct watershed, calibration scenario, software/tool, or modeling experiment treated as a separate case.

During daily calibration, NSE values ranged from 0.38 to 0.89 (Bieger et al., 2019; Kardhana et al., 2024), with a median of 0.65 (Q1 = 0.56, Q3 = 0.76), indicating that 50% of the studies achieved NSE values above 0.65, corresponding to good performance. The top 25% of studies achieved NSE > 0.76, highlighting high calibration performance. The interquartile range of 0.22 indicates moderate variability in daily simulation accuracy across studies. Similar patterns were observed for other performance metrics. KGE values ranged from 0.49 to 0.91 (Asadi et al., 2024; Wagner et al., 2022) (median = 0.72). R² values ranged from 0.4 to 0.90 (median = 0.67), indicating strong agreement between observed and simulated streamflow. RSR (0.4–0.77, median = 0.58) indicates moderate residual error consistent with NSE values. The median absolute PBIAS was 8%, indicating low systematic bias in at least half of the studies during calibration.

During daily validation, NSE (median = 0.65), KGE (median = 0.65), and R² (median = 0.70) confirmed the model reliability across the reviewed studies. RSR value (median = 0.6) indicates moderate residual error, while the absolute PBIAS values (median = 10%) indicate that 50% of the studies showed less than 10% systematic bias during validation. These results demonstrate that SWAT+ performs reliably at daily time steps, although variability among studies suggests sensitivity to calibration data, methodological approaches, and catchment characteristics.

During the monthly calibration period, the NSE values reported across studies ranged from −0.38 to 0.92, with a median of 0.68 (Q1 = 0.56, Q3 = 0.82), indicating that 50% of studies achieved good to very good model performance (NSE > 0.65) (Moriasi et al., 2007). KGE values ranged from 0.48 to 0.95 (median = 0.71),and R² values ranged from 0.3 to 0.97 (median = 0.74), indicating that half of the studies achieved good to very good correlation between observed and simulated streamflow. The median RSR was 0.57, indicating that 50% of the studies reported RSR values below this threshold, suggesting relatively low residual error. Similarly, the median absolute PBIAS was 13.7%, indicating that half of the studies exhibited relatively low systematic bias.

During the monthly validation period, more than 50% of the studies reported NSE and KGE values above 0.65, indicating good to very good model performance. However, one study reported a negative NSE value (−5.21), suggesting poor model performance under conditions of limited data availability or complex hydrological processes (Harifidy et al., 2024). R² median = 0.73, with approximately 50% of studies confirming a strong correlation between observed and simulated streamflow during validation. The median R² value was 0.73, indicating that approximately 50% of the studies showed a strong correlation between observed and simulated streamflow during validation. The values of RSR and absolute PBIAS are summarized in Table 5. Overall, the synthesized results were higher on the monthly time scale than on the daily time scale. These performance benchmarks are consistent with other SWAT+ studies that reported satisfactory streamflow simulations under diverse hydro-climatic conditions. For example, Gautam and Corzo (2023) evaluated flood and drought mitigation in Nepal, Leone et al. (2023) assessed ecological flow regimes in a Mediterranean basin, Grosser and Schmalz (2025) projected climate change impacts on low flows in Germany, Xiao et al. (2025) used CMIP6-SWAT+-BMA for transboundary lake basin assessment in China, and Zhu et al. (2025) investigated hydrological drought responses in northwest China using an improved glacier-integrated SWAT+ model. Collectively, these studies reinforce the model’s robustness across varied settings.

3.9 Future outlook

Based on the findings of this review, several priority directions can guide future SWAT+ research. One of these is that improving the quality and availability of climate input data remains essential, as model performance strongly depends on reliable datasets. Many regions still have limited ground-station coverage, which affects the reliability of model calibration. Consequently, several studies have explored alternative datasets, including satellite and reanalysis products, to compensate for sparse observations (Eini et al., 2022; Abate et al., 2023; Kardhana et al., 2025). For example, CHIRPS precipitation and CFSR reanalysis datasets have shown promise as reliable alternatives for hydrological modeling in data-scarce regions (Dogiso et al., 2025; Harifidy et al., 2024; Zhu et al., 2023). Beyond input data, an important frontier in SWAT+ research is the development of hybrid modeling frameworks that combine the physical rigor of process-based models with the predictive strength of machine learning. Recent studies show that algorithms such as Informer, Bi-LSTM, and Temporal Fusion Transformers can substantially outperform standalone SWAT+ simulations by capturing latent nonlinear residuals and temporal dependencies (Cao and Ying, 2025). Explainable AI techniques, such as SHapley Additive exPlanations, are also essential for maintaining model transparency and ensuring that machine-learning enhancements remain hydrologically plausible (Asadi et al., 2024).

Beyond traditional static calibration, sequential data assimilation techniques are also needed. Although most SWAT+ studies rely mainly on streamflow calibration, recent evidence indicates that incorporating additional variables, such as ETa or soil moisture, can better represent hydrological processes and improve model reliability (Abate et al., 2023; Eini et al., 2023; Jiménez-Navarro et al., 2024). Because model performance also depends on input data quality, further research should examine the effects of higher-resolution DEMs in hydrologically sensitive regions, as well as the influence of different soil databases, LULC datasets, and streamflow records. Although the SWAT+ Toolbox is widely used for calibration and validation, and NSE remains the most commonly applied objective function, using multiple objective functions is recommended to address the limitations of relying on a single metric. However, the current version of the SWAT+ Toolbox does not support multi-objective functions for streamflow calibration and validation, although it can implement several single-objective functions. Enhancing the software to support multi-objective calibration would therefore improve its capacity to generate more realistic streamflow simulations.

Emerging calibration strategies that use recent, high-quality observations for calibration and earlier records for validation have shown promise for improving parameter estimation and model reliability (Tan et al., 2023; Tigabu et al., 2024, 2025; Tarkegn et al., 2025a, 2025b). Therefore, before calibrating a model with earlier data and validating it with recent data, researchers should first assess data quality by comparing observed and simulated values and prioritize the use of higher-quality datasets to achieve more accurate streamflow simulations. Overall, future SWAT+ applications will benefit from integrating reliable climate and streamflow data, adopting multivariable calibration approaches, and applying advanced multi-objective and automated optimization techniques to improve model performance across diverse hydro climatic environments.

4 Conclusion

This systematic review shows that SWAT+ offers greater flexibility through its modular representation of landscape units and improved connectivity. However, model reliability still depends strongly on the suitability of input data, calibration methods, and evaluation frameworks for specific watershed conditions. The reviewed studies indicate that meteorological forcing and spatial datasets substantially affect model performance, although their suitability varies with basin characteristics, data availability, and spatial heterogeneity. High-resolution datasets may improve simulations in small or complex catchments, while coarser global products can be useful in larger basins where spatial aggregation reduces input variability. In data-scarce regions, satellite and reanalysis climate products, such as CHIRPS, ERA5, and CFSR, provide valuable alternatives; however, their performance should be evaluated locally and bias-corrected before use.

Calibration strategies should also align with study objectives and watershed characteristics. Although automated optimization tools such as the SWAT+ Toolbox with DDS are widely used for their computational efficiency, the reviewed studies emphasize the need to combine optimization with hydrological knowledge to improve process representation. In heterogeneous basins, multi-site and multivariable calibration approaches that incorporate additional observations, such as evapotranspiration or soil moisture, can provide complementary constraints and reduce parameter uncertainty. However, their effectiveness depends on the quality, scale compatibility, and uncertainty of the added datasets.

Furthermore, the selection of objective functions is critical during model calibration, as it guides parameter optimization and influences the resulting parameter sets. In contrast, performance metrics are essential for evaluating the accuracy and reliability of calibrated SWAT+ simulations. Although NSE remains frequently used, its sensitivity to high-flow conditions highlights the importance of complementary metrics such as KGE and low-flow indicators, particularly when seasonal dynamics or extreme events are important. Therefore, multi-metric evaluation frameworks are recommended to provide a more comprehensive assessment of model behavior rather than relying on a single performance indicator. Overall, future SWAT+ applications should adopt context-dependent strategies that consider watershed characteristics, data availability, and research objectives. These tailored approaches can enhance the credibility and practical value of SWAT+ as a decision-support tool for sustainable water resource management.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI was solely to assist with language editing and the organization of figures.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frwa.2026.1882804/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Jiangjiang Zhang, Hohai University, China

Reviewed by: Lei Yao, Hohai University, China

Samanta Tolentino Cecconello, Federal Institute of Rio Grande do Sul (IFSul), Brazil

Footnotes

https://osf.io/mdkbq/overview?view_only=1a3dd39ae0d7439794df53a438ea8764