Cloud Removal in Satellite Imagery using Adversarial Network and RGB-Optical Data Fusion

Abstract

Many earth observation activities rely on optical remote sensing data. The optical remote sensing imagery is exploited in various applications like farmland monitoring, land use, land cover, 3D city models, vegetation growth, and disaster mitigation. Despite all, cloud cover significantly impacts on spatial and temporal accessibility of the earth observation. Since the first observation, it has been one persistent difficulty for optical remote sensing. For decades, researchers have been studying to remove clouds from optical images. The procedure of clearing the clouds becomes more difficult as they thicken. In such instances, it is customary to reconstruct utilizing additional images such as synthetic aperture radar (SAR) or near-infrared (NIR). In this paper, we propose a two-stage architecture-based cloud removal framework. The first stage of our network translates SAR and optical cloudy images to synthetic optical (RGB) image using the conditional Generative Adversarial Network (cGAN) and the second stage reconstructs the cloud-free image by fusing the synthetic optical (RGB) and cloudy optical image. The network was tested on the real cloudy images and the proposed method was compared with the state-of-the-art models and showed better results for cloud removal.