Cross‐Regional Transfer Learning of Flow‐Based Bathymetric Inversion With Predictive Uncertainty Quantification

Abstract

Existing bathymetric inversion methods remain challenging, primarily due to poor cross‐regional generalization and the lack of predictive uncertainty quantification. To address these issues, we developed a normalizing flow‐based framework incorporating multiple transfer learning strategies for probabilistic bathymetry prediction. The South China Sea was selected as the source region to construct a Flow_SCS model with predictive uncertainty quantification. The model is then transferred to the Sea of Japan using different transfer strategies and further applied to the East Siberian Sea. Results show that the Flow_SCS model achieves a standard deviation (STD) of prediction error of 126.29 m in the South China Sea, outperforming GEBCO_2025, topo_27.1, and ETOPO_2022. In the Sea of Japan, fine‐tuning Batch Normalization parameters yielded the best performance, achieving an STD of 70.39 m and improved predictive uncertainty metrics. Furthermore, self‐training and physical constraints improve the inversion performance of the Flow network. Overall, these results demonstrate the effectiveness of the proposed framework for cross‐regional bathymetric inversion with reliable predictive uncertainty quantification.