Datasets:

INSAIT-Institute
/

FireScope-Bench

license: cc-by-4.0
pretty_name: FireScope-Bench
tags:
  - wildfire
  - remote-sensing
  - geospatial
  - reasoning
  - CoT
task_categories:
  - image-text-to-image
  - image-to-image

FireScope-Bench dataset

This dataset accompanies the paper:

FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle
Accepted to CVPR 2026
Paper: https://arxiv.org/abs/2511.17171

File and Folder Description

`climate_data.json`

Contains climate data for various coordinates.
There is a correspondence between coordinates encoded in tile/raster filenames and dictionary keys in this file.

`usa/wildfire_rasters`

`images/`

Sentinel-2 satellite images corresponding to wildfire risk rasters. Images are fused in 2×2 tiles, resulting in ~2K resolution images.

`rasters/`

Wildfire risk rasters from the Wildfire Risk to Communities project.
Rasters are fused in 2×2 tiles, resulting in 682×682 resolution rasters.

`small_sample/`

Small sample subset used for the majority of the FireScope experiments.

images/ -- Sentinel-2 images for the small sample
rasters/ -- wildfire risk rasters for the small sample

`usa/wildfire_events`

`images/`

Sentinel-2 images for wildfire events, organized in subdirectories by year.

Images are pre-event images.
Therefore, images located in folder Y correspond to wildfire events in year Y+1 and are named accordingly to match the corresponding event mask.

`masks/`

Wildfire event masks from the MTBS (Monitoring Trends in Burn Severity) dataset.

`controls/`

Masks for control areas (zero-masks).

`controls_images/`

Sentinel-2 images corresponding to the control areas.

`europe/wildfire_events`

`images/`

Sentinel-2 images for wildfire events, organized in subdirectories by year.

Images are pre-event images.
Therefore, images located in folder Y correspond to wildfire events in year Y+1 and are named accordingly to match the corresponding event mask.

`masks/`

Wildfire event masks from the EFFIS Burnt Areas dataset.

`controls/`

Masks for control areas (zero-masks).

`controls_images/`

Sentinel-2 images corresponding to the control areas.

Licensing

This dataset aggregates data from multiple sources and is released under Creative Commons Attribution 4.0 (CC-BY-4.0).

Underlying data sources include:

Sentinel-2 satellite imagery (Copernicus programme)
Wildfire Risk to Communities rasters
MTBS (Monitoring Trends in Burn Severity) wildfire event data
EFFIS Burnt Areas dataset
NASA POWER climate data

Users should also comply with the licensing and attribution requirements of the original data providers.

Citation

If you find this work useful in your research, please cite:

@article{markov2025firescope,
  title={FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle},
  author={Markov, Mario and Ailuro, Stefan Maria and Van Gool, Luc and Schindler, Konrad and Paudel, Danda Pani},
  journal={arXiv preprint arXiv:2511.17171},
  year={2025}
}