---
license: apache-2.0
task_categories:
- image-segmentation
tags:
- wildfire
- burnscars
- segmentation
pretty_name: CEMS Burn Scars
size_categories:
- n<1K
---
# Dataset Card for CEMS Burn Scar
### Dataset Summary
This dataset provides Harmonized Landsat and Sentinel-2 (HLS) imagery of burn scars and associated masks for the years 2020-2024, primarily over Europe, with some non-European cases (e.g., Mexico, Guatemala). It includes 439 high-quality 512x512 scenes, designed for training geospatial machine learning models for burn scar segmentation.
The dataset follows the same principles and structure as the [HLS Burn Scars dataset](https://huggingface.co/datasets/ibm-nasa-geospatial/hls_burn_scars), ensuring compatibility and consistency.

## Dataset Structure
### TIFF Metadata
Each TIFF file is a 512x512 pixel scene.
Scenes contain eight bands, and masks have one band. 
Satellite scenes have bands in Digital Numbers (DN), as per Sentinel-Hub documentation (https://docs.sentinel-hub.com/api/latest/data/hls/#units).
$$ DN = 10000 \times REFLECTANCE $$

All images contain **metadata** about the event, including the title corresponding to the activation in CEMS, the date range and the area.


## Band Order
For scenes: 

| Channel | Name | HLS S30 Band number |
|---------|-----|---------------------|
|1| Blue| B02|
|2| Green| B03|
|3| Red| B04|
|4| NIR| B8A|
|5| SWIR 1| B11|
|6| SWIR 2| B12|
|7| dataMask| -|
|8| QA| -|

## Class Distribution
Masks are a single band with values:
- 1: Burn scar
- 0: Not burned

|Label | Representation|
|------|------|
|Burn Scar | 10 %|
|Not burned | 90 %|

## Data Splits
The 439 files are randomly split into training (70%), validation (12%), and test (18%) directories, each containing masks, scenes, and index files.

## Dataset Creation
Following the principles of the HLS Burn Scars dataset, we developed a sophisticated pipeline to transform Copernicus EMS (CEMS) burn scar geometries into high-quality image-label pairs. Key steps include:

- Extracting geometries from CEMS Rapid Mapping data and tiling into a UTM grid (512x512 at 30m resolution).
- Applying spatio-temporal clustering to group scars within a 30-day window, resolving overlaps and minimizing false negatives.
- Downloading HLS imagery via Sentinel-Hub, filtering out tiles with excessive cloud cover or invalid pixels. This process ensures robust data for training geospatial models.

## Source Data
Imagery is sourced from HLS V2 via Sentinel-Hub, accessible at https://www.sentinel-hub.com/. A full description of HLS is available at https://lpdaac.usgs.gov/documents/1698/HLS_User_Guide_V2.pdf.
Burn scar geometries are from CEMS Rapid Mapping (2020-2024), covering ~200 fire events, available at https://mapping.emergency.copernicus.eu/.

## Citation
If this dataset supports your research, please cite CEMS Burn Scars in your publications. Example BibTeX entry:
```
@dataset{cems_burnscars,
  author       = {Jose Moreno Ortega},
  title        = {CEMS BurnScars},
  year         = {2025},
  month        = {may},
  day          = {4},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.15335977},
  url          = {https://doi.org/10.5281/zenodo.15335977}
}
```