Boreal Forest Fire: UAV-collected Wildfire Detection and Smoke Segmentation Dataset.

Area of Science:

• Computer Vision
• Forestry Science
• Remote Sensing

Background:

• Automated wildfire detection systems require extensive, annotated datasets, which are currently lacking.
• Existing datasets often do not capture the specific conditions of boreal forest environments from an uncrewed aerial vehicle (UAV) perspective.

Purpose of the Study:

• To address the data scarcity in automated wildfire detection by creating an open-access, annotated dataset.
• To provide valuable data for training and validating wildfire detection models, particularly in boreal forest settings.

Main Methods:

• Collected images and videos from multiple prescribed burning events in Finnish boreal forests using UAVs.
• Co-annotated the visual data using both human experts and computer vision foundation models.
• Structured the dataset into three sections: images with bounding boxes, labeled video clips, and images with segmentation masks.

Main Results:

• Successfully trained wildfire detection models using the developed dataset in prior research, demonstrating its efficacy.
• The dataset includes diverse annotations crucial for robust model development.
• Released accompanying code to facilitate data utilization and model implementation.

Conclusions:

• The Boreal Forest Fire dataset effectively bridges the gap in open-access, annotated data for wildfire detection.
• The dataset's quality and comprehensiveness support the advancement of automated wildfire detection technologies in challenging environments.
• The availability of data and code promotes further research and development in the field.