Super-resolution image generation from earth observation satellites using generative adversarial networks

The spatial resolution is one of the main criteria representing the level of detail in an image. The necessity for the high spatial resolution has increased with the development in satellite technologies. Using modern sensors and optics is an expensive way to improve image spatial resolution. Image super resolution is one of the most important computer vision research topics that aims to obtain higher spatial resolution image(s) from one or more lower spatial resolution ones. It is a cheaper and more effective way as it does not require any modification to the camera hardware. In this thesis, the Super-Resolution Generative Adversarial Networks (SRGAN) and the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) both trained with Google Earth were utilised for the super-resolution enhancement of Sentinel-2 and Göktürk-2 images. The results of the pre-trained deep learning models using the different data sources with multi-sensor and multi-temporal characteristics were analyzed and their super resolution performances were evaluated. The results show that the perceptual image quality of low spatial resolution satellite images can be improved by using SRGAN and ESRGAN methods.