高分遥感影像云雪共存区轻量云高精度检测方法  

Accurate and lightweight cloud detection method based on cloud and snow coexistence region of high-resolution remote sensing images

作　　者：张广斌 高贤君[1,2] 冉树浩 杨元维 李丽珊 张妍 ZHANG Guangbin;GAO Xianjun;RAN Shuhao;YANG Yuanwei;LI Lishan;ZHANG Yan(School of Geosciences,Yangtze University,Wuhan 430100,China;Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake,Ministry of Natural Resources,Nanchang 330013,China;Hunan Provincial Key Laboratory of Geo-Information Engineering in Surveying,Mapping and Remote Sensing,Hunan University of Science and Technology,Xiangtan 411201,China;Beijing Key Laboratory of Urban Spatial Information Engineering,Beijing 100045,China)

机构地区：[1]长江大学地球科学学院,湖北武汉430100 [2]自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,江西南昌330013 [3]湖南科技大学测绘遥感信息工程湖南省重点实验室,湖南湘潭411201 [4]城市空间信息工程北京市重点实验室,北京100045

出　　处：《测绘学报》2023年第1期93-107,共15页Acta Geodaetica et Cartographica Sinica

基　　金：自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室开放基金(MEMI-2021-2022-08);湖南科技大学测绘遥感信息工程湖南省重点实验室开放基金(E22133,E22205);自然资源部地理国情监测重点实验室开放基金(2020NGCM07);城市空间信息工程北京市重点实验室经费资助项目(20210205);海南省地球观测重点实验室开放基金(2020LDE001)。

摘　　要：云检测是遥感图像预处理过程中的关键步骤,但是当场景的下垫面中存在雪时,常用的云检测方法易受到干扰而降低云检测精度。此外,现有云检测数据集多为中等分辨率,且并未强调探讨云雪共存区域。因此,本文创建发布了基于高分辨率云雪共存遥感影像的云检测数据集CloudS,并提出了一种面向高分辨率云雪共存场景的轻量云检测卷积神经网络RDC-Net。RDC-Net中包含可重构多尺度特征融合模块以用于多尺度云特征的提取;双重自适应特征融合模块以对有效云特征实现表征重建;可控深层梯度指导流模块进行网络梯度下降的无偏指导。受益于上述几个技术组件,该网络能进一步提升复杂区域云检测的稳健性并促进部署的轻量化。试验结果表明,本文方法对遥感影像中的薄云及雪域上空的云具有极佳的提取能力,同时对雪等高亮地物具有良好的抗干扰能力。此外,RDC-Net具有极少的参数量与前向推理浮点运算量,这也使得其适合于实际的工业生产部署。Cloud detection is a critical stage in remote sensing image preprocessing. However, when there is snow on the underlying surface of scenes, the general cloud detection methods wouldbe easily affected. As a result, the cloud detection accuracy of these methods would reduce.Furthermore, most available cloud detection datasets are of medium-resolution and do not focus on the cloud and snow coexistence study areas. As a result, a cloud detection dataset has been created and released based on high-resolution cloud-snow coexistence remote sensing images.Meanwhile, this study suggests a convolution neural network termed RDC-Net for cloud detection in high-resolution cloud and snow coexistence images. The RDC-Net contains the reconstructible multiscale feature fusion module for multiscale cloud feature extraction, the dual adaptive feature fusion module for effective cloud feature representation reconstruction, and the controllably deep gradient guidance flows module for unbiased network gradient descent guidance. Benefiting from the above technical components, the network can enhance the robustness of cloud detection in complicated regions and facilitate lightweight deployment of the network. The experimental results show that the RDC-Net has an excellent anti-interference capacity for highlighted ground objects and has outstanding detection performance for thin clouds and clouds over snow. Furthermore, the RDC-Net has fewer parameters and floating-point operations, making it acceptable for industrial production and application.

关 键 词：高分辨率遥感影像 云雪共存区域 云检测 卷积神经网络 高精度 轻量级 

分 类 号：P237[天文地球—摄影测量与遥感]