An Explainable Deep Learning Method With Squeeze and Excitation Block for 6‐Hour Forecast of Tropical Cyclone Remote Precipitation

Abstract This study presents a deep learning framework based on a VGG‐like architecture for forecasting +6‐hr tropical cyclone remote precipitation (TCRP), a major precipitation phenomenon in East Asia. A three‐stage optimization strategy was implemented: (a) K‐fold cross‐validation was used to select 10 optimal ensemble members, which were further optimized; (b) atmospheric physics knowledge was integrated into activation function design and objective functions, resulting in a 19.23% increase in the Squeeze and Excitation Block's (SE‐Block) mean attention allocation to topographic features during the study period (2021–2023); (c) multi‐modal inputs combined with custom objective functions enhanced heavy TCRP prediction accuracy by 0.18 compared to a control experiment. The July 2021 Henan extreme rainfall event demonstrated that replacing traditional activation functions with physics‐informed designs and using custom objective functions increased SE‐Block attention allocation to topographic features by 21.74% and improved heavy TCRP forecasting accuracy by 0.44 for optimal models compared to the control experiment. Results showed superior performance in heavy TCRP prediction and enhanced attention allocation to topographic features during extreme rainfall events, exceeding average values in 2021–2023. This work provides an explainable deep learning approach for TCRP forecasting, advancing both methodology and physical interpretability, suggesting promising solutions for extreme precipitation forecast through hybrid approaches combining atmospheric physics with data‐driven techniques.