Deep Learning for Disaster Relief: Generating Synthetic High Resolution Images

The advancements in engineering and technologies have boosted the unprecedented development in the field of remote sensing. The amount of details modern satellites can capture carries immense value to a wide range of applications. Such high-resolution satellite imageries help us to generate a lot of information about the geography of the earth. Apart from applications in most earth science disciplines; those imageries also have applications in humanitarian works. As a part of the CERN Openlab project in collaboration with the UNITAR Operational Satellite Applications Programme (UNOSAT), this project aims at exploring the possibilities of using conditional progressive growth of GANs, where the conditioning is based on multiple encoders for image generation and image completion. Image completion is a very important problem in processing satellite images because of various reasons. Sometimes, the images taken by satellite under various impeding weather conditions might not be clear and missing parts need to be filled by extrapolation.

In addition, high-resolution satellite imageries are very often licensed in such a way that it can be difficult to share it across UNITAR, UN partners, and academic organizations. This reduces the amount of image data available to train deep learning models. Thus, to overcome this, the possibility of creating spectrally valid images using conditional progressive GANs with multiple encoders is explored in this project.