BDF-18K Dataset(ForestFireDataset)

BDF-18K Dataset (BoWFire - DFS - Forest Fire Dataset) The BDF-18K dataset is designed to support forest fire detection tasks by integrating both public datasets and self-collected data. It combines images from the BoWFire and DFS public datasets with a proprietary dataset collected in forest areas of Shifeng District, Zhuzhou City, Hunan Province, China. The dataset covers diverse forest fire scenarios and characteristics. Citation Requirement If you use the BDF-18K dataset in your research, please cite both the dataset and the following publication: Q. Yu et al.,“FFDNet: An end-to-end lightweight real-time fire detection model for forest fire environments,”Physics and Chemistry of the Earth, 2026.https://doi.org/10.1016/j.pce.2026.104325 @article{YU2026104325,title = {FFDNet: An end-to-end lightweight real-time fire detection model for forest fire environments},journal = {Physics and Chemistry of the Earth, Parts A/B/C},volume = {144},pages = {104325},year = {2026},issn = {1474-7065},doi = {https://doi.org/10.1016/j.pce.2026.104325},url = {https://www.sciencedirect.com/science/article/pii/S1474706526000586}} Failure to cite the associated publication may constitute misuse of the dataset. Dataset Composition Public Datasets BoWFire: Contains 122 images, mainly covering building fires, industrial fires, and similar scenarios. DFS Dataset: Contains 9,462 images, including flame, smoke, and background scenes. Self-Collected Dataset The forest fire dataset was collected using UAV platforms and handheld devices, focusing on dynamic scenes, multi-view perspectives, and complex backgrounds. It is designed to simulate real-world forest fire rescue scenarios. Acquisition Equipment DJI Mavic 2 Enterprise Dual (low-altitude acquisition with varying viewpoints and illumination conditions) DJI Matrice 300 RTK with Zenmuse H20T camera (high-altitude acquisition supporting both infrared and visible modalities) Moto Edge X30 smartphone (simulating firefighter perspectives for close-range fire details) Acquisition Location Typical forest regions in Zhuzhou City, Hunan Province, China, representing the natural environment and climate conditions of southern China. Dataset Content The dataset includes: Fire (small-scale and large-scale) Smoke (light smoke and dense smoke) Firefighters and volunteers Environmental background (vegetation, bare ground, etc.) Image Specifications Resolution: 1920 × 1080 Frame extraction rate: 30 FPS Annotations The dataset uses the “fire” label to annotate fire regions, making it suitable for training object detection models such as YOLO. Dataset Split The dataset is divided into three subsets: 70% training set 20% validation set 10% test set The class distribution is balanced across all subsets to ensure fair training and evaluation. Public Dataset Access BoWFire and DFS datasets are publicly available and can be downloaded separately. Download links can be found through their respective official sources. BoWFire Dataset: Includes building fires, industrial fires, and non-fire images (e.g., sunset or fire-like objects), suitable for fire detection tasks. DFS Dataset: Contains flame, smoke, and background images for fire and smoke detection.