A unified neural network-based single super-resolution method for heterogeneous digital earth remote sensing images

This paper is devoted to finding a unified solution for the practical problem of increasing the resolution of heterogeneous digital images of remote sensing of the Earth: panchromatic images, color images, multispectral and hyperspectral images, as well as images obtained by synthetic aperture radar (SAR). To obtain such a solution, an aggregated dataset was collected from the existing datasets, including both pairs of high- and low-resolution images and individual images acting as high-resolution images. For the collected database and a typical distortion model, an experimental comparison of various modern neural network models of resolution enhancement (including pre-tuned and retrained options) was carried out, which are listed in reviews as state-of-the-art: convolutional, generative adversarial and transform (based on the attention mechanism). Taking into account possible limitations of the models on the number of layers of space images, two options for image preprocessing were considered. The final result of the work is a single neural network model for increasing the resolution of digital space images, supplemented by adapters for their various types, which, when trained, surpasses or is not inferior to modern specialized state-of-the-art solutions.