HRUnet: A Deep Learning Model Designed for Lightning Forecasting at High Spatial Resolution

Lightning constitutes a significant threat to human life and property. Effective lightning forecasting significantly mitigates associated losses. However, a critical gap persists in the lack of methods capable of providing precise lightning forecasts at high spatial resolution (1km×1km). To address this limitation, we propose High-Resolution U-Net (HRUnet), a novel deep learning model integrating the U-Net encoder-decoder architecture with IR-CBAM block for temporal feature enhancement and PixelShuffle for spatial resolution preservation. Furthermore, we introduce a specialized Focus Loss function to alleviate the severe class imbalance inherent in lightning prediction tasks. HRUnet was evaluated using weather radar and lightning data from Binzhou, Shandong Province, China. Experimental results demonstrate that: (1) The IR-CBAM module, PixelShuffle module, and Focus Loss function yield improvements of 7.22%, 0.26%, and 4.02% respectively in average ETS compared to the baseline U-Net; collectively, HRUnet achieves a 10.86% absolute ETS improvement. (2) On the test set, HRUnet attains performance metrics of 0.9843 hit rate, 0.6015 probability of detection, 0.2856 false alarm rate, and 0.4713 ETS. HRUnet outperforms the other four state-of-the-art deep learning models and demonstrates the best capability in forecasting the initiation, development, and dissipation of lightning clusters among them.