Cross-basin incremental learning for tropical cyclone intensity estimation

Deep learning has attracted more and more attention in the field of tropical cyclone (TC) intensity estimation (TCIE). It is able to achieve promising results when the testing data follows the same distribution as the training data. However, due to the difference of geographical locations, TC intensity distributions, and imaging sensors, TC in different basins often show diverse distributions, making deep learning models trained on one basin can hardly be generalized to other basins. In this paper, we propose a cross-basin incremental learning model (CBIL-TCIE) to estimate the intensity of TC in multiple basins. CBIL-TCIE consists of domain-shared and domain-specific layers within the framework of multi-task learning. The domain-shared layers learn the common knowledge of all basins, and the domain-specific layers learn the specific knowledge of the current basin. Additionally, most of the existing studies have primarily focused on utilizing either maximum sustained wind (MSW) or minimum sea level pressure (MSLP) to represent TC intensity. Differently, in order to better characterize the intensity of TCs, our model can output MSW and MSLP concurrently as the TC intensity in different basins. To test the performance of our proposed model, we conduct experiments on a widely used dataset named GridSat, which consists of TC data across multiple basins. The performance of the CBIL-TCIE in multiple basins can improve by 19.2 % compared to the widely used fine-tuning method. Furthermore, the experiment demonstrates that concurrently outputting MSW and MSLP can effectively facilitate the ability of TC intensity estimation.