Adversarial Learning with Limited Supervision for Atmospheric Cloud and Snow Identification in Satellite Imagery

Deep learning methods have made it much easier to find clouds and snow in remote sensing images. However, deep learning methods need a lot of labelled data, which takes up a lot of time and money. A lot of research has been done on weakly supervised methods to make annotation less work, but most of these methods only look for clouds and not snow very often. This article presents an innovative weakly supervised method for detecting clouds and snow. We create generative adversarial networks (GANs) to make cloud and snow images and pseudolabels for training detection networks, using the remote sensing imaging mechanism as a guide. The suggested method can make clouds in different states and copy the texture of snow. The cloud GAN model and the snow GAN model both use image-level annotation training supervision to make remote sensing images and their pseudolabels. They do this by producing both pixel-level cloud/snow reflectance and cloud opacity. Our method outperforms other weakly supervised methods when it comes to finding clouds and snow.