An Effective Cloud Removal Algorithm Based on Sparse Expression

Abstract

In order to remove the clouds cover in remote sensing images, this paper proposes a clouds removal algorithm based on spatiotemporal fusion and sparse expression. In this paper we select two images in the same area with a short time difference, one of them is covered by clouds, the other clear image is basically not. First, we use some indicators [1] to distinguish clouds and clear ground to get classified images, such as NDVI(Normalized Difference Vegetation Index), and EVI(Enhanced Vegetation Index). Then we select the cloudless regions in both images to train sparse expression dictionary. For the cloudy regions, we find the corresponding area in the cloudless area in the other image, in which we perform dictionary decomposition and get a sparse representation [2], [3]. We utilize the dictionary of cloudy image to map the sparse representation to the corresponding area in cloudy image, and replace the original cloudy area. The fusion image retains the original real information and predicts the type of ground surface under the clouds. The experimental results show that the algorithm has an excellent effect on the removal of thick clouds, and solves the problem of image distortion and grayscale mutations that may be caused by traditional clouds removal algorithms. Sentinel-2 satellite images were used to evaluate the proposed method, and it was compared with other related algorithms, for example, homomorphic filtering and LRMR(low-rank matrix recovery), the experimental results confirm that the proposed method is effective in correcting clouds contaminated images while preserving the true spectral information.