AFSeq: A Multi-Region Benchmark of Daily Wildfire Spread Sequences

AFSeq is a benchmark of daily wildfire-spread sequences built from VIIRS active-fire detections (SNPP + NOAA-20) rasterized at 200 m resolution. It contains 1,194 fire events from 8 countries across 3 regions, spanning 2023-09-29 to 2026-06-02, with burned areas from 301 ha to 169,511 ha (median 1,035 ha). Compared with existing spread benchmarks, AFSeq is dominated by the short, small fires that those benchmarks under-represent.

It accompanies the WACV 2027 paper From Next Day Wildfire to Next Days Wildfire: A New Multi-Continent Dataset of Active-Fire Spread for Sequence-to-Sequence Forecasting.

• DOI: 10.57967/hf/11378
• Code (construction pipeline, splits, reference dataloader, training): https://github.com/joseg20/AFSeq
• Mirror: Harvard Dataverse, doi:10.7910/DVN/JIIMNE

Coverage

Region	Countries	Events
South America	Chile, Argentina, Bolivia	453
Southern Europe	Portugal, Spain, France, Italy	319
United States	United States	422
Total	8 countries	1,194

Events are split 70/15/15 by fire identity (835 train, 179 validation, 180 test), stratified by region. Three partitions with fixed seeds (5, 7, 23) are distributed with the code repository.

Files

• afseq.tar.zst.part-aa ... afseq.tar.zst.part-ai: the dataset archive (about 21.6 GB), split into 9 parts.
• SHA256SUMS.txt: SHA-256 of each part.
• croissant.json: machine-readable metadata (MLCommons Croissant 1.0, including Responsible AI fields).

Download and extraction

hf download josegcc20/AFSeq --repo-type dataset --local-dir AFSeq
cd AFSeq
shasum -a 256 -c SHA256SUMS.txt
cat afseq.tar.zst.part-* > afseq.tar.zst
shasum -a 256 afseq.tar.zst
# expected: e0c62cb677719c7c037b59e032e605987255a56e3410d6458e7e94982b320431
tar --zstd -xf afseq.tar.zst

Structure

afseq/
├── provenance.json          # sensor provenance of every fire day
├── excluded.json            # filtered events (33 duplicates, 74 persistent-signal sources)
└── id_XXXX/                 # one directory per fire event
    ├── info.geojson         # event metadata (area, dates, raster window)
    ├── contour.geojson      # event contour
    ├── dem.tif              # elevation, slope, aspect
    ├── vegetation.tif       # land cover, NDVI, EVI, ET
    └── YYYY-MM-DD/          # one sub-directory per active day
        ├── active-fire.tif  # band 1: binary detections; band 2: UTC hour of detection
        ├── climate.tif      # temperature, dew point, relative humidity, wind speed,
        │                    # wind direction, surface pressure, precipitation
        └── reflectance.tif  # VIIRS I1, I2, M11 (VNP09GA)

All rasters share a common 200 m grid per event. Missing observations are stored as NaN rather than filled with a constant.

Task

Next-day and multi-day fire-mask prediction from the event history: given the observed days up to t, predict the active-fire mask at t+1, ..., t+k.

Limitations

• Fire labels are VIIRS active-fire detections, not ground-surveyed burn perimeters. Detections can be missed under cloud cover, dense smoke or between overpasses, and low-intensity or sub-pixel fires may be omitted.
• The native detection resolution (375 m) is coarser than the 200 m grid. Temporal resolution is daily; the UTC detection-hour band only partially captures sub-daily dynamics.
• Events smaller than about 300 ha are excluded by construction.
• Coverage is limited to three regions and to the 2023-2026 fire seasons.

License and attribution

AFSeq is released under CC BY 4.0. It is derived from open data products: VIIRS active fires (SNPP, NOAA-20) and VNP09GA reflectance, MODIS (MCD12Q1, MOD13Q1, MOD16A2, MYD11A1), SRTM, CHIRPS and ERA5-Land, retrieved through Google Earth Engine.

Contains modified Copernicus Climate Change Service information (2026). Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains.

Citation

If you use AFSeq, please cite the paper:

@inproceedings{guillen2027afseq,
  title     = {From Next Day Wildfire to Next Days Wildfire: A New Multi-Continent Dataset of Active-Fire Spread for Sequence-to-Sequence Forecasting},
  author    = {Guillen, Jose and Zanca, Renzo and Barriere, Valentin},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}

To cite the dataset release itself:

@misc{guillen2026afseqdata,
  title     = {{AFSeq}: A Multi-Region Benchmark of Daily Wildfire Spread Sequences},
  author    = {Guillen, Jose and Zanca, Renzo and Barriere, Valentin},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/11378},
  url       = {https://huggingface.co/datasets/josegcc20/AFSeq}
}

Acknowledgments

This project has been partially funded by the grants "fAIrefighter: A wildfire early detection and spread prediction AI-driven decision support tool" ID24I10398 IDeA I+D ANID, "WWF – wildfire watch and fight: an AI platform and infrastructure for early detection of forest fires" PINC230011 ANID. This work was partially funded by the National Center for Artificial Intelligence CENIA FB210017, Basal ANID.