Author(s): Zhuo Yang; Dong Wang; Chenlu Yu; Xiaoyu Ye

Linked Author(s):

Keywords: Bidirectional Gated Recurrent Unit; Dynamic feature fusion; Feature extraction; Precipitation runoff modeling; SHAP analysis

Abstract: Deep learning has become a powerful approach for rainfall–runoff prediction, showing remarkable skill across diverse basins by capturing complex nonlinear relationships between hydrological drivers and runoff. The increasing availability of satellite remote sensing and reanalysis data now enables multimodal fusion, offering a more complete view of hydrological processes. Yet, existing multimodal frameworks often treat spatial feature extraction and temporal modeling as separate tasks, struggle with the instability of early convolutional architectures, and lack unified mechanisms to interpret the contributions of different input modalities. Here we develop a residual-based dynamic fusion framework that integrates spatial and temporal representations through task-oriented optimization. Learnable attention parameters within an improved ResNet module adaptively link feature extraction with sequence modeling, while residual structures enhance stability and representation of complex multimodal interactions. A unified contribution quantification scheme further provides physically interpretable insights into the role of each data source. Applied to two climatically contrasting river basins in China, ReDF-Net coupled with BiGRU achieved the highest skill. SHAP-based analysis confirmed the interpretability of the framework. Our results demonstrate that ReDF-Net provides a generalizable and physically meaningful paradigm for multimodal runoff prediction under changing environmental conditions.

DOI:

Year: 2026