WildfireVLM

YOLO-format imagery and bounding-box annotations for wildfire/smoke detection, prepared as the dataset companion to WildfireVLM: AI-powered Analysis for Early Wildfire Detection and Risk Assessment Using Satellite Imagery.

Dataset description

The release contains 3,776 JPEG images and matching YOLO text annotations. It preserves the supplied train, valid, and test directory structure:

Split	Images
train	3,118
valid	57
test	601

Each non-empty label row uses the YOLO bounding-box convention:

class_id x_center y_center width height

Coordinates are normalized to [0, 1]. The supplied annotations contain 9,640 objects with class ID 0 and 237 with class ID 1.

Repository layout

data.yaml
train/images/
train/labels/
valid/images/
valid/labels/
test/images/
test/labels/

Important label note

The original data.yaml is retained unchanged for reproducibility. Its two class-name entries are export-generated strings rather than human-readable semantic labels. Consumers should confirm the intended mapping for class IDs 0 and 1 with the dataset authors before reporting per-class results or renaming them.

Intended use

This dataset is intended for research and development of computer-vision systems for detecting wildfire-related regions in Earth-observation imagery. It supports the detection component described in the associated paper; it does not itself contain the language-based risk assessments discussed there.

Provenance and license

The supplied export identifies its Roboflow source as video-dikld/wildfire-blgsv, version 7, and reports the license as CC BY 4.0. This repository preserves that attribution. Users must comply with the applicable upstream license and independently verify that their intended use is permitted.

Citation

If you use this dataset, please cite:

@article{ayanzadeh2026wildfirevlm,
  title={WildfireVLM: AI-powered Analysis for Early Wildfire Detection and Risk Assessment Using Satellite Imagery},
  author={Ayanzadeh, Aydin and Dixit, Prakhar and Kamal, Sadia and Halem, Milton},
  journal={arXiv preprint arXiv:2602.13305},
  year={2026},
  url={https://arxiv.org/abs/2602.13305}
}

Limitations

The dataset is an existing labeled export. It may reflect geographic, sensor, temporal, annotation, and sampling biases. Validate performance on the deployment domain, especially for faint smoke, cloud cover, haze, and small or early fire events.