An Autoencoder-Based Framework for Multimodal Fusion in Forecasting Tropical Cyclone-Induced Sea Surface Height Responses

Current physical models often struggle to accurately capture the spatial structure of sea surface height (SSH) changes induced by tropical cyclones (TCs) because they do not adequately account for TC features. To address this limitation, a new multimodal fusion framework, namely FusionNet, is proposed to forecast the spatial structure of SSH changes induced by TCs. By integrating TC features and pre-storm ocean conditions whilst preserving structural information and spatial patterns, the proposed model effectively captures the amplitudes of TC-induced SSH changes across various TC intensity categories, accurately forecasting both troughs and sea level rise along TC tracks. Moreover, FusionNet consistently outperforms the baseline derived from the widely used ocean reanalysis fields in capturing and modelling the spatial variations, as evidenced by results from an independent testing set and a case study of Super Typhoon Mangkhut (2018). Ablation analysis further disclosed the respective contributions of TC features and pre-storm ocean conditions in modelling the physical processes underlying TC-induced SSH responses. Notably, incorporating TC features significantly enhances the model's ability to represent trough amplitudes, particularly near the storm center and across different TC intensity groups.