Firespot-seg: a segmentation database for early-stage wildfire smoke detection with demonstration of yolov8-seg model

Wildfires are among the most destructive natural disasters, causing extensive dam- age to ecosystems, communities, and economies. In recent years, computer vision tech- niques have emerged to improve detection accuracy and eciency. One of the interesting approaches is the segmentation method, which is particularly suitable for this task due to the complex behavior of wildfire smoke, which varies in volume, density, and color based on factors like wind speed, humidity, and temperature. However, there is a scarcity of seg- mentation datasets for wildfire smoke. To address this gap, we created a dataset in YOLO segmentation format and binary mask format containing 2,913 images of wildfire smoke. This paper presents a new wildfire smoke detection segmentation dataset derived from the FireSpot dataset. Our implementation of the YOLOv8s-Seg model used as a baseline for future development, achieved 94.4% accuracy in bounding box detection and 92.8% in seg- mentation detection at a 0.3 IoU threshold. These results demonstrate the model’s perfor- mance and present the use case of this dataset for researchers developing new segmentation models.