Kyoma001/IGNITE-fire-dataset

IGNITE: A Multimodal UAV-Collected Dataset for Wildfire Detection

IGNITE contains radiometric FLIR TIFF frames aligned with RGB video frames from four UAV-collected prescribed-fire sequences. The release includes 1,854 approved aligned samples. Processing code and release provenance are available in the companion GitHub repository.

Dataset Viewer

Each Viewer row is one aligned sample. The five image columns are:

• thermal: display rendering of the radiometric TIFF
• thermal_tiff: original 16-bit radiometric TIFF
• rgb: aligned RGB video frame
• overlay: thermal/RGB visual overlay
• mask_80c: binary mask for temperatures greater than or equal to 80 °C

thermal_tiff is decoded losslessly as a Pillow I;16 image. Convert it to a NumPy array to access the calibrated uint16 digital numbers; an ordinary browser preview may not preserve their intensity semantics. The backward-compatible thermal_tiff_path column identifies the processed TIFF, while thermal_source_path points to the corresponding TIFF in the raw sequence.

The default train view contains all 1,854 approved samples. sequence_id identifies the source sequence (0001–0004); these acquisition sequences are not train/test partitions.

import numpy as np
from datasets import load_dataset

dataset = load_dataset(
    "Kyoma001/IGNITE-fire-dataset",
    split="train",
)

sample = dataset[0]
sample["thermal"], sample["thermal_tiff"], sample["rgb"], sample["overlay"], sample["mask_80c"]

thermal_dn = np.asarray(sample["thermal_tiff"], dtype=np.uint16)
temperature_c = thermal_dn * 0.04 - 273.15

In addition to the five image columns, each row includes alignment identity and timing, RGB frame index, visual alignment score, fire-mask statistics, and paths to the processed and source radiometric TIFFs.

Quick start for complete download

git clone https://github.com/Ding-Research-Lab/IGNITE.git
cd IGNITE

hf download Kyoma001/IGNITE-fire-dataset \
  0001/ 0002/ 0003/ 0004/ metadata.jsonl catalog.csv checksums.sha256 statistics/ \
  --repo-type dataset \
  --local-dir data

• The complete data/ directory is about 26.1 GB.
• For offline visual review, open data/<sequence_id>/processed/index.html in a browser.
• data/<sequence_id>/raw/ contains the source RGB video and TIFF sequence.
• data/<sequence_id>/processed/ contains aligned samples, manifests and the browser index.
• data/<sequence_id>/metadata/ contains path-portable configuration and calibration provenance.

Dataset summary

Sequence ID	Alignment strategy	Raw TIFF	Aligned samples	Anchor time(s)
0001	dual anchor	482	439	01:23.233, 08:13.233
0002	dual anchor	468	451	04:52.500, 12:00.567
0003	single anchor*	651	460	00:29.533 (take-off anchor)
0004	single anchor*	535	504	09:27.833 (landing anchor)
Total		2,136	1,854

* Sequence 0003 does not include RGB video of the landing, while 0004 does not include thermal imagery of the takeoff. Both were therefore aligned with a single-anchor strategy.

All sequences use a FLIR Vue Pro R 640 13mm radiometric TIFF (640×512, 16-bit, one frame per second) and H.264 RGB video (3840×2160, 30 FPS). The thermal calibration stored in the TIFF XMP metadata is:

temperature_C = DN * 0.04 - 273.15

The mask threshold is 80 °C, equivalently DN >= 8829. Each aligned sample contains:

thermal.tiff   original radiometric uint16 TIFF
thermal.png    display rendering of the TIFF
video.png      RGB frame after the approved crop/resize transform
overlay.png    thermal/RGB visual overlay
mask_80c.png   indexed 0/1 mask for temperature >= 80 °C

Notebook and validation

notebooks/inspect_aligned_data.ipynb displays a radiometric TIFF with a Celsius color bar together with its RGB frame, overlay and 80 °C mask. A small sample is bundled in examples/sample_aligned/ so the notebook can be previewed without downloading the complete dataset.

uv sync
uv run python code/validate_dataset.py --root data

Processing flow

raw TIFF + MP4
      │
      ├─ timestamp scan and initial time model
      ├─ landing-pad circle detection + manual confirmation
      ├─ seeded RGB circle tracking
      ├─ 90-frame anchor-window crop scoring
      ├─ Rank 1 transform selection (single or dual anchor)
      └─ full-range aligned sample export + 80 °C mask

Acknowledgment

The authors thank the prescribed-fire personnel and land-management partners who made the data collection possible, especially Dr. Deborah Landau, Mr. Gabriel Cahalan, and Mr. Chase McLean from The Nature Conservancy, and Mr. Miles Roy of American University. This research is based upon work supported in part by the NSF (#2536664) and NASA ESTO Program (80NSSC25K7777). The views and conclusions contained herein should not be interpreted as representing the official policies of the U.S. Government.

Citation

If you find this work useful for your research, please cite our paper:

@article{wang2026ignite,
  title  = {IGNITE: A Multimodal UAV-Collected Dataset for Wildfire Detection},
  author = {Yiding Wang and Zhaoxi Zhang and Chenzhi Zhao and Zhangyu Guan and Dong L. Wu and Leah Ding},
  year   = {2026}
}