Streamflow Prediction in Snow-Dominated Semiarid Watersheds Using a Cascade-Forward Deep Learning Approach

Abstract Accurate streamflow prediction in snow-dominated regions is challenging due to complex nonlinear interactions and significant temporal lags between snow accumulation and melt. This study evaluates the efficacy of a spatially explicit cascade-forward deep learning framework compared to traditional lumped methods in Cannonball Watershed, North Dakota, US. We assessed four architectures: long short-term memory (LSTM), gated recurrent units (GRU), transformer, and LSTM-PLUS, using a 365-day lookback period to capture seasonal snowpack memory. Results demonstrate that the transformer model uniquely benefited from the spatially explicit cascade-forward approach, outperforming all other models by effectively learning upstream–downstream connectivity and routing delays through its multihead self-attention mechanism. While the LSTM-PLUS variant excelled at volumetric water-balance closure, the cascade-forward transformer yielded the most robust streamflow predictions with metric indexes of percent bias (PBIAS) of − 5.172 % , R 2 of 0.816, and Nash–Sutcliffe Efficiency (NSE) of 0.811. The cascade-forward approach more accurately represented streamflow, reflecting the advantages of semidistributed machine-learning (ML) modeling, analogous to those observed in physically based distributed hydrologic models. To ensure interpretability and physical credibility, SHapley Additive exPlanations (SHAP) analysis was applied, which demonstrated that snow- and energy-balance variables, such as liquid snow water content, shortwave, and sensible heat fluxes, are dominant controls of streamflow predictions relative to rainfall input, highlighting snow accumulation and melt as the primary drivers of basin-scale runoff dynamics. The findings highlight that integrating spatial routing through cascade-forward and attention-based ML architectures enhances model resolution and predictive performance in a manner analogous to the improvements observed in semidistributed, physically based hydrologic modeling.