基于CNN的高分辨率遥感影像多级云检测  

Multi-level Cloud Detection in High-resolution Remote Sensing Images Based on CNN

作　　者：王智敏 谭海 郭正胜 魏旭 WANG Zhimin;TAN Hai;GUO Zhengsheng;WEI Xu(School of Geomatics,Liaoning Technical University,Fuxin 123000,China;Land Satellite Remote Sensing Application Center,Ministry of Natural Resources,Beijing 100048,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000 [2]自然资源部国土卫星遥感应用中心,北京100048

出　　处：《测绘与空间地理信息》2020年第9期139-142,145,151,共6页Geomatics & Spatial Information Technology

摘　　要：针对目前大多数云检测方法对薄-厚云之分检测较少以及对云边界检测性能不理想、准确率低等问题,本文通过结合深度学习的优势提出了基于CNN的遥感影像云检测方案。首先,将GF-1/2和ZY-3图像进行分割制作训练样本,用深层双分支卷积神经网络(CNN)在训练样本中提取云的多尺度特征;其次,利用改进的自适应简单线性迭代聚类方法(ASLIC)将待检测图像分割成超像素作为云检测的子区域。最后,通过网络模型对待检测子区域预测获得云区结果。选取多种高分辨率卫星图像进行云检测试验,并与单分支网络模型、经典SLIC方法进行对比。结果表明,该方法可以在不同成像平台很好地检测薄-厚云及其整个云区,且整个云区检测的平均精度在94%以上。Aiming at the problem that most existing cloud detection methods cannot discriminate thin and thick clouds and often have modest performance and low accuracy on cloud boundary detection,this paper proposes a remote sensing image cloud detection scheme based on the advantage of deep learning. Firstly,the GF-1/2 and ZY-3 images are segmented to produce training samples,and the deep convolutional neural network( CNN) is used to extract the multi-scale features of the cloud from the training samples. Secondly,using the improved adaptive simple linearity( ASLIC) method to segment the detected images into superpixels as a sub-region of the cloud detection. Finally,the cloud region results are obtained by the network model’s prediction of the detected sub-regions. A variety of high-resolution satellite images were selected for cloud detection experiments and results of proposed method are compared with single branch network model and classic SLIC method. The results show that the proposed method can detect thin and thick clouds and their entire cloud regions well on different imaging platforms,and the average accuracy of the entire cloud detection is above 94%.

关 键 词：ASLIC CNN ZY-3 GF-1/2 云检测 超像素 

分 类 号：P237[天文地球—摄影测量与遥感]