An interpretable daily runoff simulation method in data-scarce watersheds by coupling SWAT and LSTM-Transformer

Abstract：To address the challenge of simulating runoff in ungauged regions, a hybrid physical–data-driven framework was developed by coupling Soil and Water Assessment Tool (SWAT) with an LSTM–Transformer. SWAT-derived process variables were fused with meteorological forcing to form a physically informed feature set for the Transformer-enhanced LSTM. The framework was first calibrated at a gauged station and then transferred to ungauged basins to evaluate its spatial generalizability. At the gauged station, the SWAT–LSTM–Transformer achieved the highest accuracy among all tested models, yielding an NSE of 0.587 and an R² of 0.728 on the validation dataset. It also maintained a better balance between calibration fit and validation robustness than SWAT–LSTM, SWAT–RF, SWAT–SVM, and stand-alone SWAT. SHAP-based interpretation revealed stable and hydrologically coherent predictor dependencies: temperature, lateral flow, and evaporation emerged as dominant drivers of the model’s runoff simulations, whereas precipitation and soil moisture exerted shorter-term and event-focused influences. When transferred to ungauged stations in the same watershed, the model reproduced seasonal runoff variations and event-scale fluctuations with high accuracy, with NSE ranging from 0.80 to 0.94 and R² from 0.83 to 0.92. Under cross-watershed transfer, the model continued to capture the main temporal patterns, with NSE and R² ranging from 0.62 to 0.83 and 0.60 to 0.84, respectively, although performance declined during extreme events. Overall, the coupled SWAT–LSTM–Transformer framework provides a robust and transferable approach for daily runoff simulation in data-scarce watersheds.

Key words: SWAT; LSTM-Transformer; runoff simulation; ungauged watersheds