A dual-weighted loss function for lightning nowcasting

Lightning associated with severe convective weather poses a significant threat to public safety and infrastructure. While traditional numerical weather prediction and statistical methods have limitations in computational efficiency and accuracy, deep learning (DL) approaches often yield poor predictive performance for rare but critical lightning events due to the severe class imbalance inherent in the data. To address this challenge, this study introduces a dual-weighted cross-entropy loss function (DWCELoss). This novel function combines a static, global class weight with a dynamic, sample-specific grid weight to enhance the model’s sensitivity in lightning-prone regions. Experiments on the test set demonstrate that, compared to using class-weight alone, the DWCELoss-trained model (at a 0.9 probability threshold) increases the probability of detection by 78.74% (from 0.334 to 0.597), the critical success index by 36.80% (from 0.299 to 0.409), and the F1-score by 26.30% (from 0.460 to 0.581). A case study further validates the method’s ability to capture localized lightning activity, with the CSI score increasing from 0.29 to 0.43 for the event. This work provides a novel and effective pathway for DL-based lightning nowcasting, with significant implications for disaster prevention and public safety. 闪电活动对公共安全和基础设施构成重大威胁.传统预报方法在效率与精度上存在局限, 而深度学习在处理样本极度不平衡的稀有闪电事件时效果欠佳.本研究针对这一挑战, 提出了一种创新的双重加权交叉熵损失函数(DWCELoss), 通过结合静态全局类别权重与动态样本网格权重, 显著增强了模型对闪电高发区域的敏感度.实验结果表明, 在0.9的概率阈值下, DWCELoss模型相比仅使用类别权重的模型, 其检测概率提升78.74%, 临近预报成功率提升36.80%, F1分数增长26.30%.本研究为基于深度学习的闪电临近预报开辟了新路径, 对防灾减灾具有重要意义.