Deep Learning-Based Lightning Nowcasting with Embedded Attention Mechanisms

Lightning events exhibit significant variability in spatial scale, occur suddenly, have short lifespans, and evolve rapidly, making high-resolution forecasting particularly challenging. This study leverages the data-driven capabilities of deep learning to construct a multi-layer UNet neural network architecture with an embedded attention mechanism, resulting in the AME-UNet model for lightning nowcasting in North China. The model integrates lightning location data from the State Grid of China with high spatiotemporal resolution data from the FY-4A geostationary meteorological satellite. Brightness temperature channel differences, which effectively represent cloud-top development heights and freezing levels with clear physical significance, are introduced as predictors for pixel-wise lightning nowacasting for 0-1h and 1-2h timeframes. Results indicate that the AME-UNet model holds promising potential for lightning nowcasting, achieving a maximum hit rate of 0.46 and a false alarm rate of 0.29 for the 0-1 hour forecast, and a hit rate of 0.41 with a false alarm rate of 0.44 for the 1-2 hour forecast. This study offers innovative approaches to deep learning-based lightning nowcasting, expanding the toolkit available for severe weather nowcasting.