Early detection and rapid response are critical in forest fire scenarios to minimize environmental and economic damage. This work explores the use of the YOLO (You Only Look Once) version 11 neural network for detecting three key classes: dispersed smoke, dense smoke, and fire. These classes were defined to support a tracking strategy aimed at identifying the fire source. In a typical forest fire, dispersed smoke is often visible from a distance, followed by dense smoke as one approaches the origin, and eventually, the flames themselves. A dataset composed of images from the Atlantic Forest was collected for training. The best model achieved an 85% mask accuracy with a recall of 76%. The network successfully segmented the target classes with few false positives, even on data not seen during training.