Аннотация
Deep learning models such as Long Short-Term Memory networks have shown strong potential for streamflow forecasting. However, their limited interpretability poses challenges for operational use. This study examines how predictor importance varies across flow regimes and whether this can inform the development of more efficient, transparent models. Using the Uba River Basin as a case study, we apply a stratified modeling approach with Integrated Gradients to analyze feature relevance during high, normal, and low flows. Models trained on reduced feature sets derived from the analysis performed competitively with the baseline model using all features. The model with universal and high-flow-specific predictors achieved higher accuracy (NSE 0.787 vs. 0.745), while the normal-flow model reduced high-flow bias. Meanwhile, the baseline remained the most balanced overall, achieving the highest KGE score (0.829). These findings suggest that regimeaware feature selection can improve interpretability and efficiency without substantially compromising performance and contribute to the broader operational adoption of deep learning in hydrological forecasting.