【摘要】 闪电的空间尺度变化范围广泛，发生突然、生命史短且演变迅速，其高时空分辨精细预报极为困难。本研究利用深度学习数据驱动的优势，建立适应多种数据源的多层UNet结构神经网络并添加卷积块注意力模块（Convolutional Block Attention Module, CBAM）注意力机制，以此构建华北地区闪电短临预报深度学习模型AME-UNet。使用国家电网闪电定位数据和新一代静止气象卫星FY-4A高时空分辨率数据，引入表征云顶发展高度和冻结的具有明确物理意义的亮温通道差作为预报因子，用以实现未来0～1 h与1～2 h闪电发生与否的逐像素预报。结果表明AME-UNet模型在闪电短临预报任务中展示了良好的应用潜力，0～1 h命中率最高达0.46，虚警率为0.29,1～2 h命中率最高达0.41，虚警率为0.44。本研究为基于深度学习开展闪电短临预报提供了新思路和新方法。更多还原

【Abstract】 Lightning exhibits significant spatial variability, sudden occurrence, and rapid evolution with short life cycles,making high-resolution nowcasting challenging. This study employs a deep learning approach to develop a multi-layer UNet architecture with an embedded CBAM(Convolutional Block Attention Module) attention mechanism, termed AMEUNet, designed for high-resolution lightning nowcasting in North China. The AME-UNet model combines precise lightning-location data from the State Grid Corporation of China and “high spatiotemporal resolution” data from the FY-4A geostationary meteorological satellite, thereby creating a robust, multisource data foundation. Brightness temperature channel differences, which physically characterize cloud-top development heights and freezing levels, are employed as predictors for pixel-wise lightning nowcasting at 0–1 and 1–2 h lead times. The results demonstrate the competitive performance of AME-UNet, with probabilities of detection values of 0.46(0–1 h) and 0.41(1–2 h) while maintaining false alarm rates values of 0.29 and 0.45, respectively. This study presents novel deep learning approaches for lightning nowcasting, advancing the methodological toolkit for severe weather prediction.更多还原