HINet:一种面向冰雹识别的多源数据融合网络  

HINet:A Multi-source Data Fusion Network for Hail Identification

作　　者：张小雯 郁培雯 商建 华珊 张启绍 ZHANG Xiaowen;YU Peiwen;SHANG Jian;HUA Shan;ZHANG Qishao(National Meteorological Center,Beijing 100081,China;Anyang National Climatological Observatory,Anyang 455000,China;School of Artificial Intelligence,Nanjing University of Information Science&Technology,Nanjing 210044,China;National Satellite Meteorological Center(National Centre for Space Weather),Beijing 100081,China;Innovation Center for FengYun Meteorological Satellite(FYSIC),Beijing 100081,China;Key Laboratory of Radiometric Calibration and Validation for Environmental Satellites,Beijing 100081,China)

机构地区：[1]国家气象中心,北京100081 [2]安阳国家气候观象台,安阳455000 [3]南京信息工程大学人工智能学院,南京210044 [4]国家卫星气象中心(国家空间天气监测预警中心),北京100081 [5]许健民气象卫星创新中心,北京100081 [6]中国气象局遥感卫星辐射测量和定标重点开放实验室,北京100081

出　　处：《遥测遥控》2024年第4期45-56,共12页Journal of Telemetry,Tracking and Command

基　　金：国家重点研发计划项目(2022YFC3004104);中国气象局创新发展专项项目(CXFZ2024J001);中国气象局水文气象重点开放基金项目(23SWQXZ001);风云卫星应用先行计划2023(FY-APP-ZX-2023.01);安阳国家气候观象台开放研究基金课题(AYNCOF202401)。

摘　　要：冰雹天气具有突发性和局地性强,以及破坏力大的特点。尽管地面自动站、雷达和卫星等获取的观测资料在冰雹识别中发挥了一定的作用,但单一观测资料的局限性导致冰雹识别虚警率较高和准确率较低。因此,亟需构建基于多源高分辨率观测的冰雹识别技术。本文提出了一种面向冰雹识别的多源数据融合网络,该深度学习方法利用时空特征提取模块、多源数据特征融合模块和UCUNet(U Connection Unet,U形连接卷积神经网络)识别模块,充分挖掘冰雹发生时FY4B(风云四号B星)、天气雷达和数值模式等多源数据的时空特征,并创新地加入地形高度、坡度、坡向等作为冰雹识别因子。为评估所提网络方法的性能,本文进行了系列实验,并将实验结果与真实标签数据进行对比。结果显示,HINet(Hail Identification Net,冰雹识别网络)能够充分利用多源数据,在复杂地形条件下有效改善冰雹识别结果,在冰雹研究和识别中具有较高的准确性和实用性。Hailstorms are characterized by their suddenness,localized nature and high destructive power.Although observations acquired by ground-based automatic stations,radars and satellites play a certain role in hail identification,the limitation of single observation data leads to a high false alarm rate and low accuracy rate in hail identification.Therefore,there is an urgent need to construct a hail identification technology based on multi-source high-resolution observation.In this paper,a multi-source data fusion network for hail recognition is proposed.The deep learning method utilizes the spatio-temporal feature extraction module,the multi-source data feature fusion module,and the UCUNet(U Connection Unet)recognition module to fully exploit the spatio-temporal features of the multi-source data such as FY4B(FengYun-4B star)satellites,weather radar,and numerical models when hail occurs,and innovatively adds the topographic height,slope,and slope direction as hail recognition factors.In order to evaluate the performance of the proposed network method,this paper conducts a series of experiments and compares the experimental results with real labeled data.The results show that HINet(Hail Identification Net)can make full use of multi-source data and effectively improve the hail identification results under complex terrain conditions.The network model proposed in this paper has high accuracy and practicality in hail research and identification.

关 键 词：冰雹识别 深度学习 时空特征提取 多源数据特征融合 复杂地形 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置] P458.121.2[自动化与计算机技术—控制科学与工程]