Cloud Removal and Satellite Image Reconstruction Using Deep Learning Based Image Inpainting Approaches

Abstract

The existence of clouds is one of the main factors that contribute incomplete information in optical remote sensing images, restricting their further monitoring applications of vegetation, land surfaces, water bodies, etc. Reconstruct the missing information caused by clouds can effectively improve the availability of remote sensing images. The problem can be broken into two parts detection & removal of clouds and reconstruction of the obscured data in satellite images. Several methods have been designed to solve this problem. Some require particular channels, while others are heavily dependent on the availability of temporally adjacent images. With the major breakthroughs in the field of deep learning and Convolution neural networks (CNN) attempts are being made for image reconstruction. In this work we use the image inpainting technique with two algorithms namely (Partial Convolutions and Generative Multi-column Convolutional Neural Networks) for the cloud removal in the satellite images. Image inpainting reconstructs an image with hole and fills the hole in such a way that it is not detectable. It maintains good global consistency with local fine textures. We evaluate our quality of result using SSIM and PSNR values and conclude that our approach is effective and efficient in removing small patches of thin clouds/haze.