SEN12TS: A SAR and Multispectral Dataset for Land Cover Classification

The SEN12TS dataset contains Sentinel-1, Sentinel-2, and labeled land cover image triplets over six agro-ecologically diverse areas of interest: California, Iowa, Catalonia, Ethiopia, Uganda, and Sumatra. Using the Descartes Labs geospatial analytics platform, 246,400 triplets are produced at 10m resolution over 31,398 256-by-256-pixel unique spatial tiles for a total size of 1.69 TB. The image triplets include radiometric terrain corrected synthetic aperture radar (SAR) backscatter measurements; interferometric synthetic aperture radar (InSAR) coherence and phase layers; local incidence angle and ground slope values; multispectral optical imagery; and decameter-resolution land cover data. Moreover, sensed imagery is available in timeseries: Within an image triplet, radar-derived imagery is collected at four timesteps 12 days apart. For the same spatial extent, up to 16 image triplets are available across the calendar year of 2020.<br><br>The SEN12TS documentation demonstrates two initial use cases for the dataset. The first transforms radar imagery into enhanced vegetation indices by means of a generative adversarial network, and the second tests combinations of input imagery for cropland classification.