STRATA-S12 (D_lite): A Class-Balanced Subset of the Co-Registered Sentinel-1/Sentinel-2 Corpus for Self-Supervised Earth Observation

STRATA-S12 (D_lite) is a compact, class-balanced subset of the STRATA-S12 corpus, released so the data can be worked with directly without downloading the full multi-terabyte corpus. It is a curated sample of 20,916 pairs, of the same quality and format as the full corpus, intended as an accessible entry point rather than as a statistical mirror of it. Each pair is one Sentinel-2 optical tile and one Sentinel-1 radar tile over the same ground, written to the same grid: identical bounding box, identical CRS (the tile's local UTM zone) and 10 m pixels, so the two rasters correspond pixel for pixel with no resampling on the user's side. Each tile is 512x512 pixels at 10 m, covering 5.12 x 5.12 km. Sentinel-2 provides bands B4, B3 and B2 (true colour) and Sentinel-1 provides VV, VH and a precomputed VV-VH difference, all stored as float32; stacked, the pair forms a six-channel input. The data is stratified along three axes, giving 192 strata in total. The eight land-cover classes are the ESA WorldCover 2021 v200 classes present across all sampled regions - tree_cover, shrubland, grassland, cropland, built_up, bare_sparse_vegetation, permanent_water_bodies and herbaceous_wetland; the remaining three WorldCover classes are dropped because they are absent from at least one region and would leave structurally empty strata. The four seasons (spring, summer, autumn, winter) are mapped in a hemisphere-aware way, and the six regions are AFRICA, ASIA, AUSTRALIA_OCEANIA, EUROPE, NORTH_AMERICA and SOUTH_AMERICA. D_lite is drawn by sampling up to 112 pairs per stratum (class x season x region) with a fixed seed, so every stratum present in the corpus is represented up to that cap, and strata with fewer pairs contribute all they have. Because the cap deliberately flattens the natural class frequencies, D_lite is balanced across strata by design and is therefore not representative of the true class or geographic proportions of the full corpus - it is meant for convenient experimentation, not for estimating corpus statistics. Every pair ships with a STAC/PDS4-style JSON sidecar carrying its provenance and quality metadata (land-cover class, label purity, acquisition dates, geolocation and a per-pixel integrity checksum), and the directory path itself encodes the stratum, so the path acts as the index. This deposit contains the subset packaged as eight per-class archives (strata-s12_lite_<class>.tar), one per land-cover class, plus a SHA256SUMS.txt checksum file for verifying the downloads. A README.md, a load_example.py helper and an examples/ folder are included so the data can be opened and validated immediately; see the bundled README for the full layout, naming scheme and sidecar specification. The complete STRATA-S12 corpus (D) - 740,847 pairs - together with the pre-trained encoder, is archived separately; see the related identifiers for its record, where the corpus statistics are maintained. Source data: STRATA-S12 is derived from Copernicus Sentinel-1 and Sentinel-2 data accessed through Google Earth Engine, and from ESA WorldCover 2021 v200 and Natural Earth admin-0 boundaries; each is used and redistributed under its own licence. Contains modified Copernicus Sentinel data. Released under CC BY 4.0.