| 王道平,韦伟,张光磊,张翠芳,郭新文,田金虎,余铸平.用于雷暴云识别的大核卷积注意力和U-Net算法[J].电子测量与仪器学报,2026,40(2):117-125 |
| 用于雷暴云识别的大核卷积注意力和U-Net算法 |
| Deep learning approach for thunderstorm cloud identification byintegrating large kernel attention and U-Net |
| |
| DOI： |
| 中文关键词: 深度学习 雷达组合反射率拼图 雷电 雷暴云识别 |
| 英文关键词:deep learning radar composite reflectivity mosaic lightning thunderstorm cloud identification |
| 基金项目: |
|
|
| Author | Institution |
| Wang Daoping | 1.Hunan Provincial Key Laboratory of Meteorological Disaster Prevention and Mitigation, Changsha 410118,China;
2.Hunan Meteorological Disaster Prevention Technology Center, Changsha 410007,China |
| Wei Wei | Beijing HY Orient
Detection Technology Co., Ltd., Beijing 100080,China |
| Zhang Guanglei | Beijing HY Orient
Detection Technology Co., Ltd., Beijing 100080,China |
| Zhang Cuifang | Beijing HY Orient
Detection Technology Co., Ltd., Beijing 100080,China |
| Guo Xinwen | Hunan Provincial Key Laboratory of Meteorological Disaster Prevention and Mitigation, Changsha 410118,China |
| Tian Jinhu | Beijing HY Orient
Detection Technology Co., Ltd., Beijing 100080,China |
| Yu Zhuping | Hunan Petroleum Branch of Sinopec
Marketing Co., Ltd., Changsha 410028,China |
|
| 摘要点击次数: 362 |
| 全文下载次数: 252 |
| 中文摘要: |
| 针对当前雷暴云识别命中率不高的问题，通过深度学习技术提高识别的准确性。聚焦于融合大核卷积注意力机制（large kernel attention, LKA）和U-Net架构，构建了一种新的雷暴云识别模型LkaUNet,以增强模型对雷暴云整体形态特征和长距离的空间依赖的建模。研究过程中，基于湖南省2022~2023年间的S波段雷达基数据和闪电资料，进行了多阶段质量控制实现雷达组合反射率与雷电数据的时空匹配与噪声抑制，生成了高质量的雷电概率和雷达组合反射率拼图数据集。在此基础上，构建了以U-Net为基础框架，并引入大核注意力模块的深度学习模型。大核卷积注意力机制通过扩大模型的感受野，提高了模型捕获长距离依赖，增强模型对雷暴云特征的感知能力。实验结果表明，采用回归损失函数训练的模型，当阈值大于0.4时，LkaUNet模型的临界成功指数（critical success index，CSI）和命中率（hit rate,HR）均高于基准U-Net模型，同时漏报率（negative alarm probability,NAP）更低；采用分类损失函数进行训练，LkaUNet模型临界成功指数达到0.730 1，对应的命中率为86.27%、漏报率为13.73%、虚报率为18.45%。研究表明，LkaUNet能有效建模雷暴云长距离的空间依赖关系，为深度学习在强对流天气的监测提供了有效方案。 |
| 英文摘要: |
| To address the challenge of low hit rate in thunderstorm cloud identification, deep learning techniques are employed to enhance recognition accuracy. The research focuses on developing a novel thunderstorm cloud identification model, LkaUNet, which integrates a large kernel attention (LKA) mechanism with the U-Net architecture. This design enhances the model’s ability to capture global morphological features and long-range spatial dependencies of thunderstorm clouds. The study utilizes S-band radar base data and lightning observation data from Hunan Province (2022~2023), employing multi-stage quality control to synchronize radar composite reflectivity and lightning data while suppressing noise. This process generates high-quality datasets of lightning probability and radar composite reflectivity mosaics as input. The LkaUNet model builds upon the U-Net framework and incorporates large kernel attention modules to expand the receptive field, thereby improving long-range dependency modeling and feature perception. Experimental results demonstrate: When trained with regression loss functions and a threshold exceeding 0.4, the LkaUNet achieves higher critical success index (CSI) and negative alarm probability (NAP), along with a lower false alarm rate (FAR), compared to the baseline U-Net model; Under classification-based loss training, LkaUNet achieved a CSI of 0.730 1, with corresponding detection metrics of 86.27% hit rate, 13.73% miss rate, and 18.45% false alarm rate. The study concludes that LkaUNet effectively models long-range spatial correlations in thunderstorm clouds, providing a robust deep learning solution for monitoring severe convective weather. The approach highlights the potential of attention mechanisms in meteorological applications. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
