Author(s): Yuxuan Gao; Hao Hu; Dongfang Liang; Edoardo Borgomeo

Linked Author(s): Dongfang Liang

Keywords: Data-scarce catchments; Explainable AI; Hydrological prediction; Physics-informed neural networks; Transfer learning

Abstract: The reliable hydrological prediction is challenging in regions where streamflow gauging records are insufficient. Recent studies have resorted to transfer learning (TL), a machine learning technique that leverages information from data-rich (source) catchments to improve prediction accuracy in data-scarce (target) catchments. However, the existing TL approach is purely data-driven, leading to limited physical consistency and low interpretability. To address these challenges, we propose a physics-informed transfer learning (PITL) framework. In this framework, the physically meaningful parameters in the Hydrologiska Byråns Vattenbalansavdelning (HBV) model can be automatically determined and finetuned by linking with a deep learning (DL) network. The PITL explicitly tracks how physical parameters adapt during transfer, thereby demystifying the TL and enabling hydrologically consistent interpretation. From our results, across different target regions and data-scarcity levels, PITL outperforms three baseline models (LSTM-Local, LSTM-TL, and δHBV-Local), with the highest gains achieved under the most severe scarcity. For example, under the 6-month target-training scenario, PITL increases the median NSE of catchments in Chile from 0.607 (LSTM-TL) to 0.673. Beyond the improved skill, the learned dynamic soil-moisture parameter β and evapotranspiration parameter γ are found to capture the cross-region differences and exhibit seasonal consistency with the wetting–drying transitions. Overall, PITL improves both accuracy and explainability in the hydrological predictions in data-scarce regions, supporting the reliability assessment and physical parameter diagnosis during knowledge transfer.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.174

Year: 2026