Zhang et al. (2026) Coupling SWAT and Interpretable Deep Learning to Improve Streamflow Simulation in Arid Regions

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Q. Zhang, Junhu Wu, Tianle Wang, Chenzhan Tang, Weibin Liu
• DOI: 10.1007/s11269-025-04422-0

Research Groups

State Key Laboratory of Water Engineering Ecology and Environment in Arid Area, Xi’an University of Technology, Xi’an, PR China

Short Summary

This study couples the SWAT hydrological model with three interpretable Transformer-integrated deep learning models (TCN-Transformer, KAN-Transformer, xLSTM-Transformer) to enhance streamflow simulation accuracy and interpretability in arid regions. The coupled models significantly outperform the standalone SWAT model, with SHAP analysis revealing precipitation as the primary driver and SWAT-derived hydrological features playing a dominant role.

Objective

• To explore the performance differences between SWAT and deep learning (DL) models in streamflow simulation.
• To construct coupled models of SWAT with different Transformer ensemble models and evaluate their performance in daily streamflow simulation for the Wei River Basin (WRB).
• To use the SHAP method to explain the coupled models and investigate the main driving factors influencing streamflow simulation performance.

Study Configuration

• Spatial Scale: Wei River Basin (WRB) in China, covering approximately 134,800 square kilometers, divided into 52 sub-basins. Four hydrological stations (Linjiacun, Zhuangtou, Xianyang, Tongguan) were analyzed.
• Temporal Scale: Daily streamflow simulations from 1997 to 2023, with a warm-up period (1997–2001), a calibration period (2002–2015), and a validation period (2016–2023).

Methodology and Data

• Models used:
  • SWAT (Soil and Water Assessment Tool) hydrological model
  • Transformer (baseline deep learning model)
  • TCN-Transformer (Temporal Convolutional Network-Transformer)
  • KAN-Transformer (Kolmogorov-Arnold Networks-Transformer)
  • xLSTM-Transformer (Extended Long Short-Term Memory-Transformer)
  • SHAP (SHapley Additive exPlanations) for model interpretability
• Data sources:
  • Digital Elevation Model (DEM)
  • Land Use/Land Cover (LULC) data (2023)
  • Soil attributes data (13 distinct soil textural classes)
  • Meteorological data: precipitation (PCP), temperature (TEMP), wind speed (WS), relative humidity (RH), solar radiation (SR)
  • Observed daily streamflow data from four hydrological stations within the Wei River Basin.

Main Results

• The three innovative Transformer integrated models (TCN-Transformer, KAN-Transformer, xLSTM-Transformer) individually outperformed the baseline Transformer model in streamflow simulation.
• The coupled models (SWAT-T, SWAT-TT, SWAT-KT, SWAT-xT) consistently outperformed the standalone SWAT model and the baseline Transformer models during both calibration and validation periods.
• The SWAT-KT model demonstrated the best overall performance among the coupled models, achieving Nash–Sutcliffe Efficiency (NSE) values between 0.76 and 0.86 and coefficients of determination (R²) between 0.81 and 0.90 during the validation period across the four hydrological stations.
• The standalone SWAT model showed a tendency to underestimate peak flows and overestimate baseflows, while the coupled models exhibited improved performance in predicting both.
• SHAP analysis of the best-performing SWAT-KT model identified precipitation (PCP) as the key driving factor for streamflow, with an importance ranging from 14.33% to 22.55% across the stations.
• SWAT-derived hydrological features (groundwater discharge (GWQ), surface runoff (SURQ), percolation (PERC), lateral flow (LATQ), evapotranspiration (ET), soil water content (SW)) collectively contributed 57.13% to 66.79% to the overall streamflow simulation, highlighting their dominant role.
• Spatial heterogeneity in feature contributions was observed; for instance, GWQ was particularly important at Zhuangtou station (20.28% importance), while PCP, GWQ, and SURQ were significant positive contributors at Linjiacun.

Contributions

• Proposes and evaluates three novel Transformer-integrated deep learning models (TCN-Transformer, KAN-Transformer, xLSTM-Transformer) coupled with the SWAT hydrological model for enhanced streamflow simulation.
• Significantly improves streamflow prediction accuracy in arid regions by integrating the physical mechanisms of SWAT with the advanced sequence modeling capabilities of Transformer architectures.
• Enhances the interpretability of deep learning hydrological models through the application of SHAP, providing quantitative insights into the contributions of meteorological and SWAT-derived hydrological driving factors.
• Identifies the dominant role of SWAT-derived hydrological features in coupled model performance, bridging the gap between process-based and data-driven models.
• Demonstrates the robustness and adaptability of the proposed coupled models, particularly SWAT-KT, across multiple hydrological stations in a complex arid river basin.

Funding

• Scientific research project of Shaanxi Provincial Department of Education (17JS096)
• Research Fund Program of State Key Laboratory of Eco-hydraulics in Northwest Arid Region (2016ZZKT-9)

Citation

@article{Zhang2026Coupling,
  author = {Zhang, Q. and Wu, Junhu and Wang, Tianle and Tang, Chenzhan and Liu, Weibin},
  title = {Coupling SWAT and Interpretable Deep Learning to Improve Streamflow Simulation in Arid Regions},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04422-0},
  url = {https://doi.org/10.1007/s11269-025-04422-0}
}