Parquet based Sentinel 2 L1C dataset for semantic forest segmentation

This Zenodo record provides a ~35 GB sample of the remote-sensing dataset, extracted from a larger national-scale Sentinel-2 Level-1C archive covering Romania (January-December 2021). The sample contains multi-spectral image patches and their corresponding semantic-segmentation labels for forest / non-forest mapping, built from Sentinel-2 acquisitions and aligned to ESA WorldCover 2021-derived ground truth in a parquet format. 

Each sample preserves the same data organization and metadata conventions as the full dataset: patches are grouped by Sentinel-2 MGRS tile_id (to support spatially clean splitting), include acquisition identifiers/timestamps, and provide the multi-band inputs used by the models (e.g., visible + NIR + red-edge channels, resampled to a common grid for learning), together with label masks in the same pixel alignment. The subset is intended to be fully runnable for reproduction of the paper’s preprocessing, training, validation, and inference workflows at smaller scale, enabling users to benchmark architectures, reconstruction strategies for patch-based inference, and thresholding/post-processing under the same assumptions as the full pipeline.

Due to its sample nature, it does not represent the complete spatial/seasonal variability of the full corpus; users should treat results obtained from this subset as illustrative rather than definitive, and cite the accompanying paper when using the data.

Additionally, we also provide two Sentinel 2 L1C scenes and 2 Deep learning segmentation models: Attention U-Net variant and DeepLabV3Plus in ONNX format. These can be used to test the inference capabilities of the models. Helper Scripts for loading and running infrences is also provided.

The complete dataset is 4.5 TB in size, covering the following Sentinel 2 L1C MGRS Tiles split into winter, spring, summer and fall subfolders: 

T34TET, T34TES, T34TER, T34TDS, T34TDR, T34TEQ, T34UFU, T34TFT, T34TFS, T34TFR, T34TFQ, T34TFP, T34UGU, T34TGT, T34TGS, T34TGR, T34TGQ, T34TGP, T35ULP, T35TLN, T35TLM, T35TLL, T35TLK, T35TLJ, T35UMP, T35TMN, T35TMM, T35TML, T35TMK, T35TMJ, T35UNP, T35TNN, T35TNM, T35TNL, T35TNK, T35TND, T35TPL, T35TPK, T35TPJ

The complete dataset is available by request. Please visit  https://eocube.ro/  for contact details, or contact us via Zeonod.