Long-Distance Wildfire & Smoke Detection Dataset

Wildfires are won or lost in the first hour. The earlier a plume is spotted, the smaller the response — a single engine instead of an air tanker, a contained acre instead of an evacuated valley. This dataset is purpose-built for that exact moment: the faint grey wisp on a ridgeline that almost nobody sees in time. Every image is photorealistic and rendered from the perspective of an elevated fire-watch camera or lookout tower, with smoke plumes appearing as small, low-contrast objects against mountainous terrain, forest canopies, valleys, and hazy skylines. The dataset spans diverse biomes (boreal forest, chaparral, eucalyptus, mixed conifer, alpine), times of day (dawn, midday, golden hour, dusk), and atmospheric conditions (clear, hazy, partly cloudy, smoky horizon) so models generalize beyond a single geography. Annotated in YOLO format with two classes — smoke and wildfire — including pixel-perfect bounding boxes for plumes occupying as little as 0.1% of the frame. The open-source sample includes 240 images on Kaggle; the full package contains 1,500 images covering the full progression from incipient ignition to established plume.