He et al. (2026) Enhancing runoff simulation in data-scarce mountainous regions: a coupled SWAT and transferable transformer approach

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-27
• Authors: Yi He, Rui Yan, Yanhong Tang, Jun Zhang, Hui Qian, Dejing Chen, Li Zhu, Xin Cao
• DOI: 10.1016/j.jhydrol.2026.135396

Research Groups

• Shaanxi Key Laboratory of Earth Surface System and Environmental Carrying Capacity, College of Urban and Environmental Sciences, Northwest University, Xi’an 710127, China
• College of Computer Science, Northwest University, Xi’an, Shaanxi 710127, China
• School of Water and Environment, Chang’an University, Xi’an 710054, Shaanxi, China

Short Summary

This study develops the SWAT-HydroTransformer, a physics–data hybrid framework that integrates SWAT-simulated hydrological processes as physical constraints into a multi-scale Transformer architecture. The framework aims to enhance runoff prediction in data-scarce mountainous regions, demonstrating superior predictive skill and robust transferability.

Objective

• To enhance runoff simulation in data-scarce mountainous regions by developing the SWAT-HydroTransformer, a physics–data hybrid framework that integrates SWAT-simulated hydrological processes as physical constraints into a multi-scale Transformer architecture, addressing challenges posed by complex hydro-meteorological interactions and limited observations.

Study Configuration

• Spatial Scale: Shiquan River Basin, China (main application); regional transferability evaluated across three additional data-limited basins.
• Temporal Scale: Not explicitly specified, but the study addresses challenges related to limited long-term observations.

Methodology and Data

• Models used: SWAT, Transformer, SWAT-HydroTransformer (developed model), Long Short Term Memory (LSTM), Random Forest (RF), Extreme Gradient Boosting (XGBoost).
• Data sources: Hydro-meteorological observations.

Main Results

• The SWAT-HydroTransformer achieved substantially higher predictive skill compared to standalone SWAT, LSTM, RF, XGBoost, and the basic Transformer.
• Nash–Sutcliffe Efficiency (NSE) values for the SWAT-HydroTransformer were 0.843 under baseline conditions and 0.786 under non-stationary climatic conditions.
• Model interpretability analyses (SHAP and spectral decomposition) demonstrated strong physical consistency: seasonal components corresponded to precipitation-driven variability, while trend components captured baseflow and storage dynamics.
• After transfer learning (fine-tuning) across three data-limited basins, NSE values improved to 0.729, 0.826, and 0.634, indicating robust cross-basin adaptability.

Contributions

• Development of the SWAT-HydroTransformer, a novel physics–data hybrid framework for runoff prediction in data-scarce mountainous regions.
• Integration of SWAT-simulated hydrological processes as physical constraints into a multi-scale Transformer architecture.
• Incorporation of a trend–seasonal decomposition mechanism and an enhanced AutoCorrelation attention module to explicitly separate hydrological signals.
• Demonstration of superior predictive skill, enhanced interpretability, and robust regional transferability of the proposed framework.

Funding

• Not specified in the provided text.

Citation

@article{He2026Enhancing,
  author = {He, Yi and Yan, Rui and Tang, Yanhong and Zhang, Jun and Qian, Hui and Chen, Dejing and Zhu, Li and Cao, Xin},
  title = {Enhancing runoff simulation in data-scarce mountainous regions: a coupled SWAT and transferable transformer approach},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135396},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135396}
}