基于U-Net神经网络的CALIPSO产品漏检层次分类  

Classification of Missed Layers in CALIPSO Products Based on U-Net Neural Network

作　　者：耿亿霖 臧琳 毛飞跃[1] 徐维维 龚威 Geng Yilin;Zang Lin;Mao Feiyue;Xu Weiwei;Gong Wei(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,Hubei,China;Chinese Antarctic Center of Surveying and Mapping,Wuhan University,Wuhan 430079,Hubei,China;Key Laboratory of Polar Environmental Monitoring and Public Governance(Wuhan University),Ministry of Education,Wuhan 430079,Hubei,China;Electronic Information School,Wuhan University,Wuhan 430079,Hubei,China)

机构地区：[1]武汉大学遥感信息工程学院,湖北武汉430079 [2]武汉大学中国南极测绘研究中心,湖北武汉430079 [3]极地环境监测与公共治理教育部重点实验室(武汉大学),湖北武汉430079 [4]武汉大学电子信息学院,湖北武汉430079

出　　处：《光学学报》2024年第24期89-98,共10页Acta Optica Sinica

基　　金：国家自然科学基金(42201344,42322109);中央高校基本科研业务费专项资金资助(2042024kf0037)。

摘　　要：首先利用二维假设检验(2DMHT)算法,实现了基于云气溶胶激光雷达和红外探测者卫星(CALIPSO)探测的高精度层次检测。然后基于UNet神经网络,以退偏比、色比、后向散射系数和纬度等为输入,构建云和气溶胶分类模型,对2DMHT算法多检测的大气层次,即CALIPSO官方产品漏检层次进行云和气溶胶分类。为了保证与CALIPSO官方产品分类的空间分布一致性,本研究以长期的CALIPSO官方分类产品为参考对模型进行训练。独立验证实验结果表明,本研究构建的分类模型与CALIPSO官方产品的分类总体相似度可达90%。将本研究分类后结果(即同时包含CALIPSO成功检测与漏检层次)与RadarLidar联合观测进行比较。结果表明,本研究可以有效识别由于信噪比低而被官方算法丢失的云层信息,陆地和海洋区域CALIPSO云底探测误差分别减少约21%和25%。Objective Clouds and aerosols play a crucial role in the Earth’s atmospheric system,significantly impacting the Earth’s radiation balance,water cycle,and air quality.Spaceborne lidar serves as a unique tool for the vertical simultaneous detection of aerosols and clouds,providing the advantage of allweather operation.The cloudaerosol lidar and infrared pathfinder satellite observations(CALIPSO)satellite represents the most notable example of this technology.However,due to its low signaltonoise ratio,traditional lidar layer detection algorithms based on slope and threshold often miss optically thin layers of clouds and aerosols.Therefore,we propose a UNet neural network classification model based on a twodimensional hypothesis testing layer detection algorithm(2DMHTUNet)to achieve highprecision detection and classification of these missed layers.Methods We initially employ a twodimensional hypothesis testing(2DMHT)algorithm for highprecision layer detection of CALIPSO observations.Subsequently,we construct a cloud and aerosol classification model based on the UNet neural network,using RGB inputs of optical signals such as depolarization ratio,color ratio,and backscatter coefficient.This model aims to categorize atmospheric layers detected by the 2DMHT but missed by official CALIPSO products.To ensure spatial consistency with CALIPSO products,we use longterm CALIPSO official classification products(VFM)as the training set,validating model performance with independent samples.Furthermore,we compare the combined classification results of 2DMHTUNet(including both successfully detected and missed layers by CALIPSO)with RadarLidar joint observation products for validation.Results and Discussions The model,trained using CALIPSO VFM official products as ground truth and validated for accuracy based on independent samples from one month,indicates a classification consistency of 89.4%(land)and 90.2%(sea),with accuracy above 88%for both day and night(Fig.2,Fig.3 and Table 2).Comparative results based on RadarLidar joint o

关 键 词：大气光学 层次分类 星载激光雷达 未检测层次 UNet神经网络 

分 类 号：P407.2[天文地球—大气科学及气象学]