Super-resolution of satellite images: Feasibility of deep learning techniques

Abstract

The work is devoted to studying the feasibility of applying the convolutional neural networks with deep learning to the problems of super-resolution of satellite images. The main aim is to enhance the image details and delete the artifacts. The algorithms for resolution enhancement were studied. The training set of satellite images was prepared. The neural network was constructed and trained using the PyTorch library for the Python language and the NVIDIA Tesla K40m graphics processors. Comparison of constructed network with the classic interpolation algorithms was carried out for the reference satellite images. It was shown that the neural network gives a better quality of the images. © 2020 American Institute of Physics Inc.. All rights reserved.