A multisource data‐based lightning nowcasting method integrating the three‐dimensional attention mechanism

Abstract Accurate and timely lightning nowcasting has important application value in preventing meteorological disasters and reducing economic losses. To address the issues of the disjointed processThe Inner Mongolia Climate Center, Nei Mongol Autonomous Region of spatiotemporal feature extraction and the low quality of multisource data feature fusion in existing lightning nowcasting models, an early‐warning model integrating a three‐dimensional attention mechanism is proposed. Spatiotemporally coordinated multisource input samples are constructed using radar composite reflectivity and lightning location data. Spatiotemporal features are extracted through a three‐dimensional convolutional layer to achieve spatiotemporal integrated feature modeling. The three‐dimensional attention mechanism is used to dynamically assign weights to multisource data features, realizing adaptive feature fusion and improving the quality of multisource data fusion. Experiments show that this method has better performance compared with traditional two‐dimensional models. The threat score (TS) is increased by 33.8% on average compared with other models, and the false alarm rate is reduced by 19.3%.