Replication Data for: DEEP LEARNING FOR MULTIPLE-IMAGE SUPER-RESOLUTION OF SENTINEL-2 DATA
Michał Kawulok, Tomasz Tarasiewicz, Jakub Nalepa, Diana Tyrna, Daniel Kostrzewa · Harvard Dataverse · 2021
The dataset contains the data for training and validating deep neural networks that perform multiple-image super-resolution reconstruction. Each high-resolution (HR) patch extracted from 13 Sentinel-2 bands is associated with a set of 9 low-resolution (LR) images with different sub-pixel shifts. The low-resolution patches are simulated by downsampling the high-resolution patch by a factor of 3 (for details, see the referenced paper). The dimensions of the patches are as follows: - for the bands of 10 m GSD: 360x360 pixels (HR) and 120x120 pixels (LR), - for 20 m bands: 180x180 pixels (HR) and 60x60 pixels (LR), - for 60 m bands: 60x60 pixels (HR) and 20x20 pixels (LR).