Runoff variability modeling is critical for water resource management and extreme hydrological event warning under climate change. Traditional hydrological models have substantial limitations in capturing the highly nonlinear, multi-scale processes in the western United States due to complex topography and diverse climate types. This study proposes an LSTM-CNN hybrid deep learning framework integrating meteorological observations, remote sensing data, topographic features, and soil attributes to model runoff variability across 12 representative watersheds over 21 years (2000–2020). The hybrid architecture uses LSTM networks to capture 90-day temporal dependencies, multi-scale 1D-CNNs to extract short-, medium-, and long-term temporal patterns, and an MLP embedding for static basin attributes, with an attention mechanism for adaptive spatiotemporal integration. The model achieved mean Nash–Sutcliffe Efficiency (NSE) of 0. 87 and Root Mean Square Error (RMSE) of 0. 41 mm/day on the chronological test set, outperforming SWAT, Random Forest, and a single-branch LSTM (the \ (\) NSE of \ (+0. 04\) over the single-branch LSTM, under identical inputs, isolates the hybrid architecture’s contribution) ; leave-one-basin-out evaluation yielded mean NSE of 0. 77, with climate-typical basins retaining NSE \ (0. 80\). SHAP analysis revealed precipitation (0. 42), antecedent precipitation index (0. 28), and snow cover (0. 24) as primary drivers. The model demonstrated strong performance in alpine and Mediterranean watersheds (NSE > 0. 88) with a mean peak flow prediction error of 12. 1% ± 5. 4% (range 4. 3–22. 1% across the 12 basins) (MAPE on annual maximum events). This study provides a new technical approach for runoff prediction; even in the more challenging semi-arid and arid basins, where skill is reduced but remains useful, the framework offers significant implications for water resource management and climate change impact assessment.
Yun Liu (Thu,) studied this question.