ViCoffee-S2: Sentinel-2 10 m time series and coffee labels for Dak Ha and Dak Mar, Kon Tum, Vietnam (2023–2025)
Description
This dataset contains all 226 Sentinel-2 Level-2A acquisitions of relative orbit 75 between 2 January 2023 and 27 December 2025 over Dak Ha and Dak Mar communes (Kon Tum, Central Highlands of Vietnam), at 10 m resolution with ten spectral bands (B02–B12) and the scene classification layer of every date. Cloudy acquisitions are kept. The data cover ten labelled sites of about 1.1 × 1.1 km and 120 unlabelled windows of 640 × 640 m located at least 500 m from the labelled sites. For the labelled sites, labels_10m.h5 gives the coffee area fraction of every pixel and three-class labels (coffee, non-coffee, ignored) for 2023, 2024 and 2025. Labels were derived from the coffee cultivation map of Kon Tum Province, corrected by visual interpretation of very-high-resolution imagery, and screened per year with dry-season NDVI. splits.json defines five spatially separated cross-validation folds and nested training subsets with 5, 10, 50 and 100% of the labelled blocks for three random seeds. The qa folder contains the per-polygon NDVI screening and the per-date co-registration offsets; 52 acquisitions were co-registered to the labelling imagery (median residual 0.4 m). Files are in HDF5, JSON and CSV formats. README.md describes the structure, and examples/quickstart.py shows how to read the data and draw training samples with Python (numpy, h5py, optional PyTorch). Contains modified Copernicus Sentinel data (2023–2025).