A Framework for Realistic Paired Dataset Generation for Deep Learning Based Restoration of Satellite Images

Satellite images suffer from inevitable degradations such as blur, noise, and other artifacts making them less suitable for final applications unless they are restored using processing techniques. Modern methods based on deep learning have achieved impressive results in key tasks such as deblurring, denoising, super-resolution, etc. in the domain of terrestrial images but their applicability for satellite image restoration remains limited due to the unavailability of paired datasets for supervised learning. In this work, we propose a framework to address this problem by utilizing a Generative Adversarial Network along with conventional degradation estimation techniques to generate a paired dataset for deep learning. The paired dataset is generated from a set of images with desirable characteristics and hence acts as a surrogate for the actual clean data. We perform experiments to demonstrate that our strategy is able to employ recent deep learning models for standard datasets to achieve impressive restoration results on satellite images for which no paired datasets are available. Visual as well as quantitative results using 30 cm Worldview-3 and Cartosat-3 images show that our strategy is a simple, yet effective way of generating a realistic dataset for utilizing deep-learning solutions for satellite image restoration.