Spatiotemporal deep learning fusion of radar, infrared, and geospatial data for typhoon rainfall estimation in Taiwan

Abstract

Taiwan, located in the Northwest Pacific typhoon corridor, faces frequent tropical cyclones that trigger extreme rainfall and heighten flood risks, underscoring the urgency of accurate short-term forecasts. This study presents a deep learning framework-the Recurrent Spatiotemporal Fusion Module (RSFM)-which integrates rainfall accumulation, radar reflectivity, infrared satellite imagery, and geospatial data to enhance typhoon-induced rainfall prediction. Trained on 36 typhoon events from 2013 to 2023, RSFM employs semantic segmentation combined with recurrent encoding to effectively capture spatiotemporal precipitation patterns. Among three input scenarios, the full multi-source configuration reduced RMSE by up to 9%, achieving values between 3.9 and 6.6 mm/h over 1-6 h forecasts. It also improved regional accuracy, with Bias Ratios approaching 1.0 and ETS exceeding 0.55 in terrain-affected southern and eastern Taiwan. Compared to traditional radar QPE, satellite-based PERSIANN-CCS, and ConvLSTM models, RSFM demonstrated superior skill in localizing convective cores and adapting to diverse typhoon structures, as confirmed through case studies of Typhoons Soulik, Trami, Soudelor, and Megi. These results highlight RSFM's promise as a robust tool for operational early warning systems and hydrological planning in typhoon-prone, mountainous regions, offering valuable support for disaster preparedness under a changing climate.