Causality-Inspired Cross-Satellite Continual Learning for Tropical Cyclone Intensity Prediction

Deep Learning (DL) has been widely used to predict the Tropical Cyclone (TC) intensity from satellite images in recent years. It can achieve satis factory performance when the testing images follow the same distribution as the training images. How ever, due to variations in sensors and observation conditions, the images across different satellites of ten exhibit significantly different distributions, making DL models trained on one satellite difficult to generalize to another one. In this paper, we propose a causality-inspired Cross-Satellite Continual Learning (CSCL) model to address this issue. Specifically, we firstly analyze the cross-satellite generalization challenge. Then, we develop a causality-inspired continual learning architecture, which mainly consists of Princi pal Component Learning Module (PCLM) and Feature Distillation Module (FDM). PCLMaimsatintroducing extra parameters into the model trained on the original satellite for learning new knowledge from the new satellite, while FDM attempts at preserving the old knowledge learned from the original satellite. To verify the effectiveness of CSCL model, we conduct experiments using three widely used geostationary satellites in the Northwest Pacific: Himawari-8 (H8), Fengyun-2 (FY2), and Fengyun-4A (FY4A). Experimental results show that CSCL achieves superior generalization performance across all three satellites, with Root Mean Squared Errors (RMSEs) of 8.45/8.00/7.70 m/s and Mean Absolute Errors (MAEs) of 6.28/6.07/5.92 m/s for H8/FY2/FY4A respectively in 24-hour predictions for 2022, comparable with the official operational forecast methods employed worldwide.