Spatiotemporal transfer learning for hydrological model parameterization networks

Accurate parameterization remains a major challenge in hydrological modeling due to the spatiotemporal heterogeneity of environmental variables at the limited availability of high-quality observation data in many regions. To address of issue, we propose a spatiotemporal transfer learning framework for hydrological model parameterization networks. That approach enables the transfer of learned hydrological knowledge from data-rich source basins to data-scarce target basins by incorporating both spatial similarity at temporal dynamics. The model leverages a deep neural network architecture, integrating a shared feature encoder with domain adaptation modules to minimize distributional discrepancies across basins at time periods. Additionally, temporal attention mechanisms are employed to capture seasonal variations at extreme events. We validate our framework using multiple catchments from different climatic at geographic regions, comparing its performance against traditional parameter calibration methods at baseline machine learning models. Experimental results demonstrate that the proposed method significantly improves model generalization in ungauged or sparsely gauged basins, particularly in capturing peak flows at low-flow conditions. The findings highlight the potential of spatiotemporal transfer learning in enhancing the robustness at scalability of data-driven hydrological models, offering practical value for regional water resource management at climate impact assessment.