Coupling SWAT with recurrent and convolutional deep learning architectures for catchment-scale streamflow simulation and interpretability in the Upper Blue Nile River Basin

Study region The Upper Blue Nile River Basin (UBNRB), Ethiopia. Study focus Complex hydrological processes and limited observations hinder reliable streamflow simulation in the UBNRB. This study coupled the Soil and Water Assessment Tool (SWAT) with three deep learning (DL) architectures LSTM, GRU, and TCN to exploit the complementary strengths of physically based and data driven modeling. Four standalone models and three coupled models (SLSTM, SGRU, and STCN) were evaluated at two contrasting stations, Al-Diem and Kessie, addressing the limited benchmarking of multiple DL architectures and station specific SHAP interpretability in SWAT–DL coupling. New hydrological insights for the region During the evaluation period, STCN achieved RMSE/NSE of 590.4 m³ s⁻¹ /0.921 at Al-Diem, compared with 723.8/0.881 for SWAT. At Kessie, STCN achieved 292.0/0.894 versus 294.1/0.893 for SWAT, indicating that coupling benefits were station and metric dependent rather than universal. GradientExplainer showed temporally stable attribution patterns, SWATQ consistently ranked first at Al-Diem, whereas SWATQ and LATQ formed the leading attribution group at Kessie. Spearman analysis confirmed strong predictor dependence, and grouped permutation produced 2.2 and 4.7 fold larger RMSE deterioration when SWAT derived rather than meteorological predictor blocks were perturbed at Al-Diem and Kessie, respectively. These findings indicate stronger predictive reliance on SWAT derived representations, not independent causal importance, and provide a controlled historical benchmark for data scarce UBNRB streamflow modeling.