AGToLightM: A Deep Learning Model for 60-Minute Thunderstorm Probability Nowcasting by Fusing FY-4B Satellite Infrared and Ground-Based Lightning Data

Lightning activity reflects the occurrence of severe convective events and represents the most prominent physical feature of thunderstorm weather. This study utilizes FY-4B/AGRI multi-channel infrared brightness temperature data (with a temporal resolution of 15 minutes and spatial resolution of 4 km) combined with ground-based lightning observation data to construct the AGToLightM model for predicting the probability of thunderstorm occurrence within the next 60 minutes. Based on lightning event characteristics, the model incorporates a Convolutional Block Attention Mechanism (CBAM) to enhance its ability to extract key spatial and spectral features at cloud tops. An adaptive weighted loss function is employed to address the class imbalance issue caused by sparse positive lightning samples. Three study regions—North China, East China, and South China—were selected, utilizing summer 2025 data for model training and validation. Results demonstrate that AGToLightM effectively captures the spatial distribution and evolution trends of thunderstorms, achieving a best Critical Success Index (CSI) of 0.327. In case studies, areas with forecast probabilities exceeding 60% showed spatial consistency with regions exhibiting radar echoes above 50 dBZ. This study validates the effectiveness of the AGToLightM model in integrating multi-source meteorological data for severe convective forecasting, providing technical guidance for enhancing the reliability of short-term thunderstorm forecasts.