Yang et al. (2026) ReDF-net: a feature extraction and dynamic fusion framework based on residual networks for runoff forecasting

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-28
• Authors: Zhuo Yang, Di Wang, Vijay P.Singh, Along Zhang, Chenlu Yu, Xiaoyu Ye, Qingwen Deng, Xiankui Zeng, Jianguo Jiang, Jian Wu
• DOI: 10.1016/j.jhydrol.2026.135422

Research Groups

• Key Laboratory of Surficial Geochemistry, Department of Hydrosciences, School of Earth Sciences and Engineering, Nanjing University, Nanjing, Jiangsu, China
• Formerly Texas A&M University System, Texas A&M University, College of Agriculture and Life Sciences, Department of Biological and Agricultural Engineering, College Station, TX, USA
• Formerly Texas A&M University System, Texas A&M University, College of Engineering, Zachry Department of Civil and Environmental Engineering, College Station, TX, USA
• United Arab Emirates University, National Water & Energy Center, Al Ain, United Arab Emirates

Short Summary

This paper introduces ReDF-Net, a residual network-based dynamic fusion framework for multimodal runoff forecasting that adaptively couples spatial feature extraction with temporal modeling and quantifies input contributions. It demonstrates significantly enhanced accuracy (NSE > 0.97) and improved interpretability across two Chinese basins, outperforming various conventional and state-of-the-art models.

Objective

• To develop ReDF-Net, a residual network-based dynamic fusion framework, to address limitations in existing multimodal runoff forecasting by adaptively coupling spatial feature extraction with temporal modeling, enhancing training stability, and providing a unified mechanism to quantify the relative contributions of different input modalities for improved accuracy and interpretability.

Study Configuration

• Spatial Scale: Two Chinese basins with contrasting climates.
• Temporal Scale: Runoff forecasting, involving temporal prediction and modeling of hydrological processes.

Methodology and Data

• Models used: ReDF-Net (proposed framework), integrated with LSTM, GRU, and BiGRU for temporal forecasting. Baselines included Random Forest, physically based hydrological modeling (SWAT), and iTransformer. SHAP was used for interpretability analysis.
• Data sources: Multimodal hydrometeorological information, including satellite remote sensing products and reanalysis products.

Main Results

• ReDF-Net consistently enhanced runoff forecasting accuracy across multiple hydrological stations.
• ReDF-Net coupled with BiGRU achieved the best performance, with Nash–Sutcliffe efficiency (NSE) exceeding 0.97 in both evaluated Chinese basins.
• The framework significantly outperformed conventional data-driven baselines (e.g., Random Forest), physically based hydrological modeling (SWAT), baseline models, and the state-of-the-art iTransformer.
• ReDF-Net enabled global contribution assessment of cross-modal features, improving model interpretability and credibility through SHAP-based analysis.

Contributions

• Introduction of ReDF-Net, a novel residual network-based dynamic fusion framework for multimodal runoff forecasting.
• Development of learnable attention parameters within an improved ResNet module for adaptive coupling of spatial feature extraction with temporal modeling.
• Optimization of residual blocks to enhance the representation of complex multimodal features and improve training stability.
• Implementation of a unified contribution quantification module for systematic evaluation of relative input importance, strengthening physical consistency and interpretability.
• Demonstration of superior performance (NSE > 0.97) compared to conventional and state-of-the-art models across diverse basins.

Funding

• Funding information is not available in the provided paper text.

Citation

@article{Yang2026ReDFnet,
  author = {Yang, Zhuo and Wang, Di and P.Singh, Vijay and Zhang, Along and Yu, Chenlu and Ye, Xiaoyu and Deng, Qingwen and Zeng, Xiankui and Jiang, Jianguo and Wu, Jian},
  title = {ReDF-net: a feature extraction and dynamic fusion framework based on residual networks for runoff forecasting},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135422},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135422}
}