FaultDeform-SAR database

This dataset of 5000 samples was generated using a synthetic physical model of realistic fault displacement maps (see [3]) warping patches of real SAR images in radar coordinate (range and azimuth). This database is similar to the database FaultDeform [1] , but for amplitude SAR images. It was generated to train the model GeoFlowNet-SAR [2], code available here: https://gricad-gitlab.univ-grenoble-alpes.fr/giffards/geoflownet-sar . 

It contains 5000 pairs of 1024x1024x2 matrices in npy format:

- input: 'pre_post_0001.npy' refers to the pre-earthquake and post-earthquake satellited images. They are Sentinel-1 acquisitions or the same region, spatial resolution of the data is defined by a range pixel size of 2.3 meters (range spacing measured along the radar line of sight) and an azimuth pixel size of 15.6 meters, supporting detailed
surface analysis. A pair is separated of  6 or 12 days, and was taken in ground stable areas during that time. They are taken from 4 different locations: California (June 25, 2023, to July 7, 2023), Arizona (August 4, 2023, to August 16, 2023), Colorado (January 12, 2023, to January 24, 2023), and Oregon (March 2, 2023, to March 14, 2023). We warp the second image with a synthetically generated deformation fields of earthquakes (see FaultDeform [1] and GeoFlowNet IEEE paper [3]), in this case with defined anisotropic deformation magnitudes in order to account for the specificity of SAR acquisitions: the size of the pixel is larger in azimuth than in range.

- target: 'dm_0001.npy' refers to the displacement map associated with the input, that was used to warp the synthetic post-earthquake image. It is also 1024x1024x2, with range (1st band) and azimuth (2nd band) displacement values, expressed in pixel.

[1] Montagnon, Tristan; Giffard-Roisin Sophie; Hollingsworth, James; Pathier Erwan; Dalla Mura Mauro; Marchandon Mathilde, 2025, "FaultDeform : Fast and Accurate Sub-pixel Displacement Estimation From Optical Satellite Images based on Deep Learning", https://doi.org/10.57745/G02ZXZ, Recherche Data Gouv, V1

[2] Junjie Wang, James Hollingsworth, Erwan Pathier, Tristan Montagnon, Wei Li, Mengmeng Zhang, Ran Tao, Jocelyn Chanussot, Sophie Giffard-Roisin. GeoFlowNet-SAR: Earthquake Displacement Estimation from Synthetic Aperture Radar Images
IEEE Transactions on Remote Sensing and Geosciences, 2025.

[3] Tristan Montagnon, James Hollingsworth, Erwan Pathier, Mathilde Marchandon, Mauro Dalla Mura, Sophie Giffard-Roisin, 2025. GeoFlowNet: Fast and Accurate Sub-pixel Displacement Estimation From Optical Satellite Images based on Deep Learning https://hal.science/hal-04782000v1