Abstract: Clouds in remote sensing imagery have an impact on its process and subsequent target recognition. Thus, automatic cloud extraction is essential to the application of high-resolution imagery. The complex shapes of the clouds in high-resolution imagery and the interference of cloud-like targets make it difficult to achieve a practical automatic cloud extraction. In this paper, we choose snow as the example of cloud-like target, and develop an algorithm which chooses shape, texture and edge as the key features to discriminate cloud from cloud-like targets. Firstly, the input image is preprocessed with Wallis filtering to enhance texture patterns at different scales. Then the input is segmented by a fast stable mean-shift segmentation. The first support vector machine classifier is built with gray and texture features, which divides all segmented parts into "cloud" and common ground targets. A second classifier is built with edge, shape and texture features to divide "cloud" areas into clouds and cloud-like targets. Finally, Grab-cut is applied to refine edges of cloud extraction results iteratively. Experiments achieve good results and demonstrate the algorithm's capability to extract clouds in high-resolution imagery precisely with the interference of cloud-like targets.

Key words: cloud detection, cloud-like target, SVM, mean shift