Landsat-8 and Sentinel-1 land-cover benchmark with a WorldCover–Dynamic World consensus reference: Hanoi, Thanh Hoa and Ho Chi Minh City (Vietnam), 2021, 30 m

Three sites of Vietnam (Hanoi, Thanh Hoa, Ho Chi Minh City), each a window of 1414 × 1414 pixels on the native 30 m UTM grid of Landsat-8 (EPSG:32648, Google Earth Engine), 5,998,188 pixels in total. Ten features per pixel from 2021 composites: Landsat-8 Collection 2 Level-2 surface reflectance B2–B5 and the NDVI, NDWI, MNDWI and BSI indices (median of the cloud-masked observations), and Sentinel-1 GRD VV and VH backscatter in dB (temporal median, mean of σ⁰ on the linear scale over each 30 m cell). Reference: a map of five land-cover classes (water, built-up, agriculture, perennial, forest) on which ESA WorldCover 2021 v200 and Dynamic World V1 (2021) agree; pixels without consensus stay unlabeled. The 30 m WorldCover and Dynamic World label rasters and the label draws of the paper (labels equal to 10 % of the pixels of each site, stratified over the reference pixels; seeds 42–51, with the training / scoring split used for parameter selection) and the list of the Landsat-8 scenes and Sentinel-1 images of the composites are included. Companion dataset of the paper on LP-SSCFC (label-penalized semi-supervised collaborative fuzzy clustering with GPU-batched particle swarm tuning). Landsat-8 imagery courtesy of the U.S. Geological Survey; contains modified Copernicus Sentinel data (2021); Dynamic World is produced for the Dynamic World Project by Google in partnership with National Geographic Society and the World Resources Institute (CC BY 4.0).