For each wildfire recorded in EFFIS (European Forest Fire Information System), this dataset provides two Sentinel-2 acquisitions — one taken shortly after the fire and one taken several months later — each paired with a 7-class segmentation mask identifying clear land, fresh burn, old burn, cloud, cloud shadow, water, and nodata.

~19,500 fires, each with up to two scene/mask pairs
Countries: Italy, France (south of 46°N), Spain, Greece
Date range: 2020–2026
~40 GB total

Classes

Mask value	Class	Definition
0	clear	No burn, cloud, shadow, or water detected
2	fresh burn	Burned area, ≤ 90 days since ignition
3	old burn	Burned area, > 90 days since ignition (up to 365 days)
4	cloud	Detected via OmniCloudMask
5	cloud shadow	Detected via OmniCloudMask
6	water	NDWI > 0.05, or Sentinel-2 SCL value 6
255	nodata	Outside the valid data mask

Dataset structure

acquisitions/
  <fire_id>/
    <timestamp>/
      <fire_id>_<timestamp>_s2.tif     # 13-band image, uint16
      <fire_id>_<timestamp>_mask.tif   # 1-band label mask, uint8

Each fire_id (e.g. effis_52171) may contain one or two timestamp subfolders — one from the fresh-burn acquisition window (5–60 days post-fire) and, when a usable scene was found, one from the old-burn window (120–300 days post-fire). These are independent acquisitions of the same location, not a single paired before/after image — see Limitations below.

Image format

Bands (13, in order): B01 B02 B03 B04 B05 B06 B07 B08 B8A B09 B10 B11 B12
Resolution: 10 m (coarser native bands resampled up to this common grid)
Tile size: 256 × 256 pixels (2.56 km × 2.56 km), centered on the fire
Encoding: uint16, scale factor 10,000 → divide by 10,000 for TOA reflectance
Processing level: Sentinel-2 L1C (top-of-atmosphere) — no atmospheric correction, no sensor simulation, no synthetic degradation

How to load

This dataset is distributed as georeferenced GeoTIFF pairs rather than a flat image-classification layout, so the standard datasets.load_dataset() image loader does not apply directly. Load with rasterio:

import rasterio
import numpy as np
from huggingface_hub import hf_hub_download

repo_id = "neet1797/burn2scar"

s2_path = hf_hub_download(repo_id, repo_type="dataset",
                          filename="acquisitions/effis_52171/2021-09-19T.../effis_52171_..._s2.tif")
mask_path = hf_hub_download(repo_id, repo_type="dataset",
                            filename="acquisitions/effis_52171/2021-09-19T.../effis_52171_..._mask.tif")

with rasterio.open(s2_path) as src:
    bands = src.read()          # shape: (13, 256, 256), uint16
    reflectance = bands.astype(np.float32) / 10000.0

with rasterio.open(mask_path) as src:
    mask = src.read(1)          # shape: (256, 256), uint8, values in {0,2,3,4,5,6,255}

To browse all files programmatically, use huggingface_hub.HfApi().list_repo_files() or snapshot_download() for a full local mirror.

Fire selection criteria

Filter	Value
Countries	IT, FR (south of 46°N), ES, EL
Date range	2020-01-01 to 2026-08-04
Area	5–2,000 hectares

Area bounds exclude both marginal micro-detections and mega-fires large enough to make a single tile 100% burn scar with no useful class boundary.

Source data

Imagery: Copernicus Sentinel-2, via the Copernicus Data Space Ecosystem
Burn labels: EFFIS (European Forest Fire Information System), Copernicus Emergency Management Service
Cloud / shadow: OmniCloudMask
Water: NDWI + Sentinel-2 Scene Classification Layer (SCL)

Limitations

No pre-fire baseline. Both acquisition windows are strictly post-fire; this dataset does not provide a true "before" image for change-detection-style pairing.
old_burn is comparatively rare. Because both windows search only forward in time from ignition, old_burn labels arise either from the tile's own fire (once enough time has passed) or from incidental overlap with a different, separately dated fire. Class balance should be checked empirically before training.
Cloud cover permitted elsewhere in-frame. Scene selection requires the burn scar itself to be minimally obscured (≤ 50%), but up to 70% cloud cover is permitted in the wider scene outside the scar.

Licensing

Sentinel-2 imagery: governed by the Copernicus Sentinel Data Legal Notice — free, full, and open, permitting reproduction, distribution, and adaptation, with attribution required ("Contains modified Copernicus Sentinel data").
EFFIS burn-perimeter labels: derived from Copernicus Emergency Management Service data. Redistribution terms for this specific product have not been independently verified beyond the general Copernicus data policy — consult EFFIS/CEMS directly before relying on this for redistribution outside your organization.

This is not legal advice; verify licensing terms independently for your use case.

Citation

If you use this dataset in your work, please cite:

@misc{thind2026burn2scar,
  author    = {Thind, Parampuneet Kaur},
  title     = {burn2scar: A Sentinel-2 Burn-Scar Segmentation Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/neet1797/burn2scar}
}

Please also retain attribution to the underlying source data, as required by its own license terms:

Contains modified Copernicus Sentinel data.
Burn labels derived from EFFIS (European Forest Fire Information System).

Author: Parampuneet Kaur Thind (Param)

Downloads last month: 2,101

Total file size:

10.9 GB