# TMB-S3: Sentinel-3 OLCI Time Series for Burned Area Segmentation of 246 European Wildfires

## Facts

- Provider: LINKS Foundation and Politecnico di Torino (Barco, Arnaudo, Bragagnolo, Rossi, Garza)
- License: CC-BY-4.0 (https://huggingface.co/datasets/links-ads/tmb-s3)
- Size: 14,601 files
- Format: JSON, Parquet, Zarr (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/links-ads/tmb-s3
- Access: public, readable without an account
- Category: Earth Observation & Geospatial
- Source: https://github.com/links-ads/tmb-s3
- Page: https://scigantic.com/archive/c3c21521-3fe8-59b5-bd6b-15b265dd9391
- Keywords: Earth Observation & Geospatial, image segmentation, time series, benchmark, JSON, Parquet, Hugging Face

## About

TMB-S3 (Temporal Modelling for Burn Scars on Sentinel-3) is a burned-area segmentation benchmark built from 246 European wildfire activations of the Copernicus Emergency Management Service (CEMS) between 2016 and 2025, split into 770 bounding boxes of about 25 x 25 km. Each box holds a Sentinel-3 OLCI time series at 300 m (21 bands Oa01 to Oa21, top-of-atmosphere reflectance, uint16) with one pre-fire acquisition followed by post-fire acquisitions (median sequence length 9, range 2 to 36), a per-frame OLCI cloud mask, the CEMS burned-area delineation and analysed-area mask at 10 m (2500 x 2500), and ESA WorldCover 10 m land cover. Splits are made at event level (146/50/50 activations for train/validation/test) and recorded in manifest.parquet. Data are one Zarr v3 store per bounding box under data/, named - - .zarr, with groups s3olci/, firemask/, validity/ and landcover/ carrying affine transforms and EPSG codes; open them with zarr or xarray. Code to reproduce the paper (resampling 10 m labels to the OLCI grid) is at github.com/links-ads/tmb-s3. Licensed CC BY 4.0 by LINKS Foundation; credit Copernicus Sentinel data, CEMS and ESA WorldCover, and cite Barco et al., AICT 2026.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("links-ads/tmb-s3")
print(ds)
```

## Data structure

File types in sampled files: JSON, Parquet, Zarr

Example file: `manifest.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/c3c21521-3fe8-59b5-bd6b-15b265dd9391
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/c3c21521-3fe8-59b5-bd6b-15b265dd9391
- Over MCP: `get_schema_card` with id `c3c21521-3fe8-59b5-bd6b-15b265dd9391` at https://api.scigantic.com/mcp

## Related topics

- [Earth Observation & Geospatial datasets](https://scigantic.com/topics/earth-observation-and-geospatial)
- [Parquet datasets](https://scigantic.com/topics/parquet)
- [JSON datasets](https://scigantic.com/topics/json)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Image segmentation datasets](https://scigantic.com/topics/image-segmentation)
- [Time series datasets](https://scigantic.com/topics/time-series)
