SAR data for Land Cover Mapping using Deep Learning

Published: 05-02-2020| Version 1 | DOI: 10.17632/gzb3y9kmf2.1
Sanja Scepanovic


We provide preprocessed Sentinel-1 SAR images with corresponding CORINE labels that can be used for training and evaluating Deep Learning (DL) semantic segmentation models for land cover mapping. The data comes from 14 raw Sentinel-1 scenes with two polarisation channels (single, such as HH or VV, and cross-pol, such as HV and VH) that were multilooked, calibrated, terrain-flattened, and terrain-corrected. The Sentinel-1 scenes were split into ~7K 512x512 pixel imagelets. To create RGB images suitable for training DL models from the imagelets, each of the two SAR pols is used as one channel in the resulting RGB format, and a free Digital Elevation Model (DEM) layer is added as a third channel. To create labels, CORINE land cover map is simply split into pieces corresponding to the imagelets areas. We provide imagelets in the .geotiff format so that the georeference information is preserved. The folder structure is suitable for training and evaluating deep learning models: • test • test-labels • train • train-labels • val • val-labels


Steps to reproduce

Instructions and corresponding GitHub repository with the models: