SAR despeckling filters dataset

Published: 19 October 2023| Version 1 | DOI: 10.17632/2xf5v5pwkr.1
Ruben Vasquez


The dataset consists of images downloaded from, without taking georeferencing into account. Specifically, 10 SAR images were taken from Sentinel-1 in Level 1 Detected High-Res Dual-Pol (GRD-HD), polarization VV. With this set of images provided by ASF DATA SEARCH, a rescaling to 8-bit integer images was performed, all of them were registered (aligned) with respect to one reference image, and then a multitemporal fusion was performed by averaging all of them. The averaging process integrates all ten images, pixel by pixel in a single one, resulting in one single image. Finally, one of the 10 downloaded images (noisy) is taken and split into 1600 data with dimensions of 512x512 pixels, and then this data set is separated into 1500 images (training) and 100 (validation). This process is also performed for the image resulting from the fusion of 10 images (noiseless), generating another 1600 images. The root folder includes the "Main folder" which contains four subfolders, namely: Noisy: 1500 SAR noisy images for training Noisy_val: 100 SAR noisy images for validation GTruth: 1500 noiseless images for training GTruth_val: 100 noisless images for validation The images are in TIFF (Tagged Image File Format) with values from 0 to 255 uint8.



Earth-Surface Processes, Global Change, Synthetic Aperture Radar Images, Applied Machine Learning