TMB-S3: Temporal Modelling for Burn Scars on Sentinel-3

TMB-S3 is a burned-area segmentation dataset of 246 European wildfire activations (2016–2025) of the Copernicus Emergency Management Service (CEMS), split into 770 bounding boxes of about 25 × 25 km. Each bounding box contains a Sentinel-3 OLCI time series (one pre-fire acquisition followed by the post-fire acquisitions), the CEMS burned-area delineation, the CEMS analysed area and the ESA WorldCover land cover.

• Paper: "Temporal Modelling for Burn Scars on Sentinel-3", Barco L., Arnaudo E., Bragagnolo A., Rossi C., Garza P., Application of Information and Communication Technologies (AICT) Conference 2026.
• Code: GitHub

Splits

	Train	Val	Test	Total
Events (EMSR activations)	146	50	50	246
Bounding boxes	517	135	118	770

Splits are made at event level: all bounding boxes of an activation are in the same split. The split of each bounding box is the split column of manifest.parquet.

Structure

README.md                   this card
manifest.parquet            one row per bounding box
data/
└── EMSR457-1-1.zarr/       one zarr (v3) store per bounding box, named <EMSR id>-<AOI>-<index>
    ├── zarr.json               attributes: bbox_id, event_time, delineation times, UTM CRS, geometry
    ├── s3olci/                 Sentinel-3 OLCI, 300 m
    │   ├── data                (T, 21, H, W) uint16, bands Oa01…Oa21, top-of-atmosphere reflectance
    │   ├── time                (T,) int64, acquisition time (Unix seconds, UTC)
    │   ├── prefix              (T,) uint8, 0 = pre-fire, 1 = post-fire
    │   └── cloudmask           (T, H, W) uint8, 1 = cloud (from the OLCI quality flags)
    ├── firemask/data           (2500, 2500) uint8, 10 m, CEMS burned area (1 = burned)
    ├── validity/data           (2500, 2500) uint8, 10 m, CEMS analysed area (1 = labelled)
    └── landcover/data          (2500, 2500) uint8, 10 m, ESA WorldCover class index (code / 10;
                                e.g. 1 = tree cover, 8 = permanent water)

Each group has its affine transform and crs_epsg in its attributes. OLCI images are about 83–90 × 83–90 pixels.

Frames within a sequence are sorted by time. Every bounding box has exactly one pre-fire acquisition and at least one post-fire acquisition (sequence length T: median 9, range 2–36). The last post-fire acquisition is the first clear acquisition after the CEMS delineation (median lag +1.5 days). Acquisitions with more than 20% invalid pixels were discarded.

Labels at the OLCI resolution

In the paper, the 10 m layers are resampled to the 300 m OLCI grid: a 300 m pixel is burned if any 10 m pixel inside it is burned, and pixels with no analysed 10 m pixel are ignored. The code repository implements this (ResampleMask).

Usage

hf download <org>/tmb-s3 --repo-type dataset --local-dir data/tmb-s3

import zarr

root = zarr.open("data/tmb-s3/data/EMSR457-1-1.zarr", mode="r")
olci = root["s3olci/data"][:]      # (T, 21, H, W)
prefix = root["s3olci/prefix"][:]  # 0 = pre-fire, 1 = post-fire
burned = root["firemask/data"][:]  # (2500, 2500), 10 m

With the code repository, place the dataset in data/tmb-s3/ and follow its README to reproduce the paper's results.

Sources and attribution

• Contains modified Copernicus Sentinel data (2016–2025), Sentinel-3 OLCI.
• Burned-area delineations: Copernicus Emergency Management Service (© European Union), rapid mapping activations listed by EMSR id in manifest.parquet.
• Land cover: ESA WorldCover 10 m (© ESA WorldCover project, CC BY 4.0).

License

CC BY 4.0. When using the dataset, please also credit the sources listed above.

Citation

TBD