基于深度语义分割的FY-2E遥感影像云检测方法  

A Cloud Detection Method for FY-2E Remote Sensing Imagery Based on Deep Semantic Segmentation

作　　者：高昂 肖萌[1] 唐世浩[1] 姜灵峰[1] 咸迪[1] 郑伟[1] GAO Ang;XIAO Meng;TANG Shihao;JIANG Lingfeng;XIAN Di;ZHENG Wei(National Satellite Meteorological Center,Beijing 100081)

机构地区：[1]国家卫星气象中心,北京100081

出　　处：《气象科技》2021年第5期671-680,共10页Meteorological Science and Technology

基　　金：国家重点研发计划重点专项(编号:2018YFC1506500)资助。

摘　　要：本文提出了一种基于深度语义分割技术的全自动云检测算法,可提高FY-2E遥感影像的云检测精度。首先,将FY-2EL1数据与精度较高的云检测结果进行匹配,获得用于训练和评估样本的数据集;其次,设计了深度语义分割网络,并针对训练集中正负样本严重失衡的问题,改进了损失函数,可以有效提取云的边界;最后,分别以FY-2E和MODIS数据作为训练和标签样本训练网络,得到了可用于FY-2EL1影像检测的四分类模型。试验结果表明,在四分类检测中,所提方法的准确率达到了75%,Kappa系数为0.53左右。与现有多通道阈值法相比,采用所提方法进行二分类检测可提高约90%样本的准确率,部分样本的准确率提升20%以上。此外,所提方法对云边缘、破碎云等细节识别能力较强,且具有一定的鲁棒性,受训练样本中的误判类别影响较小。未来通过扩充数据集并优化网络,可提高FY-2全圆盘影像的数据质量。A full-automatic cloud detection algorithm based on deep semantic segmentation is proposed to improve the accuracy of cloud detection for the remote sensing imagery of FY-2 Esatellites.Firstly,to train and evaluate,a sample data set is created by the data of FY-2 EL1 matched with the cloud detection results with high accuracy.Secondly,a deep semantic segmentation network is designed.A loss function is improved to extract the cloud’s boundary effectively for a severe imbalance between positive and negative samples in the train data set.Finally,FY-2 E and MODIS data,taken as train and label samples,respectively,are used for training networks,resulting in four classification models for detecting FY-2 EL1 imagery.The test results show that the proposed method’s accuracy and the Kappa coefficient are 75%and about 0.53 in four classification tests,respectively.Compared with the existing multi-channel threshold method in two classification tests,the proposed method can improve the accuracy of about 90%of the samples and the accuracy of some samples by more than 20%.In addition,the proposed method has a strong recognition ability for cloud edges,broken clouds and other details.It has a certain degree of robustness,which is less affected by the misclassification categories in train samples.Furthermore,by expanding the data set and optimizing the network,the proposed method will improve the data quality of the entire disk imagery of FY-2.

关 键 词：遥感 风云二号 云检测 深度学习 语义分割 

分 类 号：P407.8[天文地球—大气科学及气象学]