基于深度学习网络显著目标检测算法的强降水落区临近预报技术  

A technique for nowcasting heavy rainfall locations using salient object detection algorithm based on deep learning network

作　　者：张亚萍 刘伯骏[1,2] 庞玥 张焱 张勇 黎中菊 ZHANG Yaping;LIU Bojun;PANG Yue;ZHANG Yan;ZHANG Yong;LI Zhongju(Chongqing Meteorological Observatory,Chongqing 401147,China;CMA Economic Transformation of Climate Resources Key Laboratory,Chongqing 401147,China)

机构地区：[1]重庆市气象台,重庆401147 [2]中国气象局气候资源经济转化重点开放实验室,重庆401147

出　　处：《气象学报》2025年第2期334-349,共16页Acta Meteorologica Sinica

基　　金：重庆市自然科学基金项目(CSTB2023NSCQ-MSX0894);中国气象局西南区域气象中心创新团队基金(XNQYCXTD-202203);重庆市气象部门业务技术攻关项目(YWJSGG-202203);中国气象局创新发展专项(CXFZ2022J002)。

摘　　要：降水持续时间相同时,降水强度越大,诱发城市积涝、山洪、泥石流、滑坡等次生灾害的可能性越大。在全球气候变暖的背景下,亚小时降水极端事件比小时以上时间尺度的极端事件增加得更快,有必要研究亚小时尺度上的强降水监测预警技术。选取 2013-2021 年重庆地区 30 次暴雨天气过程,以显著目标检测深度网络 U2-Net 为基础,将 30 min 短时强降水落区作为显 著目标,天气雷达反射率因子拼图作为输入,通过数据驱动方式自动学习某一时次的天气雷达反射率因子空间分布与其后 30 min 的强降水落区的非线性关系,进行强降水落区预报。强降水落区标签按照 10、20 和 30 mm 阈值分为 3 种,由雷达融合地面分钟 级雨量的定量降水估计得到。模型输入为 3、4.5 和 7 km 高度的雷达反射率因子拼图。经过训练和验证,得到针对 3 种强降水阈 值的 3 个强降水落区预报模型。对测试集的检验结果表明,当邻域半径为 5 km 时,10、20 和 30 mm 阈值模型输出的命中率分别 为 0.66、 0.73 和 0.72,虚警率分别为 0.06、 0.32 和 0.57,临界成功指数分别为 0.64、 0.54 和 0.37;强降水落区预报图中的概率越大, 对强降水落区的预报越可靠。综上所述,强降水落区预报模型通过提取单个时次的三维反射率因子多尺度特征,预报未来 30 min 强降水落区,可以有效补充雨量计布设稀疏地区的强降水监测和预报信息,也为需要提取多源探测资料多尺度特征的临近预报 技术研究提供参考。For precipitation with the same duration,the higher the precipitation intensity,the greater the occurrence likelihood of urban waterlogging,flash flooding,mudslides,landslides,and other secondary disasters.In the context of global warming,subhourly extreme precipitation events are increasing much faster than those with longer durations,highlighting the need for advanced monitoring and nowcasting technologies for sub-hourly heavy rainfall.This study selects 30 heavy precipitation events in Chongqing from 2013 to 2021 to train the deep network U2-Net for significant object detection,using weather radar reflectivity mosaics as inputs.The network identifies heavy rainfall areas as salient objects and autonomously learns the nonlinear relationship between the spatial distribution of reflectivity at a given time and the subsequent 30 min heavy rainfall areas and provides forecasts of heavy rainfall region.The sample labels are divided into three categories based on thresholds of 10 mm,20 mm,and 30 mm obtained from radar-rain gauge quantitative precipitation estimates.The model inputs are radar reflectivity mosaics at altitudes of 3 km,4.5 km,and 7 km.After training and validation,three forecasting models corresponding to the three heavy rainfall thresholds are developed.Testing on an independent dataset reveals that,with a neighborhood radius of 5 km,the models achieve hit rates of 0.66,0.73,and 0.72,false alarm rates of 0.06,0.32,and 0.57,and critical success indices of 0.64,0.54,and 0.37 for the 10 mm,20 mm,and 30 mm thresholds,respectively.Higher probabilities in the forecast maps indicate more reliable forecasts of heavy rainfall areas.In conclusion,the heavy rainfall forecasting models can effectively predict 30 min heavy rainfall areas by extracting multi-scale features of three-dimensional radar reflectivity.These models help supplement rainfall monitoring and forecasting in areas with sparse rain gauge networks,providing a valuable reference for nowcasting technologies that require multi-source data feature ext

关 键 词：显著目标检测 深度网络 天气雷达 短时强降水 临近预报 

分 类 号：P412[天文地球—大气科学及气象学]