河北灾害性天气事件图像识别技术研究与应用  

Research and Application of Image Recognition Technology for Disastrous Weather Events in Hebei

作　　者：魏铁鑫 司丽丽 赵亮[1,2,3] 张静 孙斌[1,2,3] WEI Tiexin;SI Lili;ZHAO Liang;ZHANG Jing;SUN Bin(Key Laboratory of Meteorology and Ecological Environment of Hebei Province,Shijiazhuang 050021,China;China Meteorological Administration Xiong'an Atmospheric Boundary Layer Key Laboratory,Xiong'an New Area 071800,China;Hebei Meteorological Disaster Prevention and Environment Meteorology Center,Shijiazhuang 050021,China)

机构地区：[1]河北省气象与生态环境重点实验室,河北石家庄050021 [2]中国气象局雄安大气边界层重点开放实验室,河北雄安071800 [3]河北省气象灾害防御和环境气象中心,河北石家庄050021

出　　处：《热带气象学报》2024年第6期983-992,共10页Journal of Tropical Meteorology

基　　金：河北省气象局科研开发项目(21ky03、19ky06)共同资助。

摘　　要：为有效利用社会化观测灾害性天气事件的大量图像,发展监测新手段,以社会公众发布的灾害性天气事件图像为训练集,基于残差网络ResNet50(Residual Network 50)构建了8类灾害性天气事件图像识别模型,并通过格点化气象要素实况二次订正技术以及线上审核、线下更新的优化模式,提高模型识别精度和运行效率。结果表明,优化后8类模型识别的平均准确率在80%以上,已应用于河北省多源气象灾情监测采集业务,获取的灾害性天气和气象灾害事件图像数量较传统方式有显著提升。To effectively utilize the extensive image data of disastrous weather events observed by the public,and to develop new monitoring methods,we created a training set using images of disastrous weather events shared by the public.Based on the ResNet-50 convolutional neural network,an image recognition model was developed for eight types of disastrous weather events.The model’s accuracy and operational efficiency were enhanced through a secondary correction technique using gridded meteorological parameters,as well as an optimization mode involving online auditing and offline updates.After optimization,the average recognition accuracy rate of the eight models exceeded 80%.This technology has been applied to the multi-source meteorological disaster monitoring and documentation in Hebei Province,significantly increasing the number of images obtained for disastrous weather events compared to traditional methods.

关 键 词：灾害性天气 图像识别 社会化观测 ResNet50残差网络 格点化实况订正 

分 类 号：X43[环境科学与工程—灾害防治]