Chai et al. (2026) A differentiable Xin’anjiang model incorporating transfer learning for cross-basin flood prediction under data scarcity

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-02
• Authors: Yikai Chai, Wenyu Ouyang, Xuezhi Gu, Yong Peng, Xiaolan Chen, Yulin Yang, Minglei Ren
• DOI: 10.1016/j.jhydrol.2026.136321

Research Groups

• School of Infrastructure Engineering, Dalian University of Technology (China)
• Ningbo Institute of Dalian University of Technology (China)
• College of Hydrology and Water Resources, Hohai University (China)
• China Institute of Water Resources and Hydropower Research (China)

Short Summary

This paper presents a differentiable Xin'anjiang model (dXAJ) that integrates transfer learning to enhance regional adaptability for cross-basin flood prediction under data scarcity. The dXAJ achieves robust temporal generalization and strong spatial transferability.

Objective

• Investigate the feasibility of using a differentiable hydrological model integrated with transfer learning for improving flood prediction in data-scarce regions.

Study Configuration

• Spatial Scale: Regional scale, focusing on 18 diverse basins in Anhui Province, China.
• Temporal Scale: Long-term, evaluating the performance of the dXAJ over 663 flood events.

Methodology and Data

• Models used: Differentiable Xin'anjiang model (dXAJ), Xin'anjiang model (XAJ), and Long Short-Term Memory (LSTM) network.
• Data sources: Flood event data from 18 basins in Anhui Province, China; multi-source remote sensing data.

Main Results

• The dXAJ achieves a median Kling-Economou-Gupta index (KGE) of 0.785 on the test set, rising to 0.811 for large floods.
• In zero-shot transfer, the pre-trained dXAJ attains a median KGE of 0.694 and a median absolute peak error of 21.4%.
• With transfer learning, fine-tuning using only 33% of local events yields a KGE of 0.785 and reduces peak error to 15%.

Contributions

• The study demonstrates that differentiable hydrological modeling integrated with transfer learning provides a practical and interpretable framework for improving flood prediction in data-scarce regions.
• The findings highlight the potential of transfer learning to enhance regional adaptability and reduce data demand.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 51979003) and the China Postdoctoral Science Foundation (Grant No. 2020M672242).

Citation

@article{Chai2026differentiable,
  author = {Chai, Yikai and Ouyang, Wenyu and Gu, Xuezhi and Peng, Yong and Chen, Xiaolan and Yang, Yulin and Ren, Minglei},
  title = {A differentiable Xin’anjiang model incorporating transfer learning for cross-basin flood prediction under data scarcity},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136321},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136321}
}