Brazil suffers from fires, especially in the seasons of greatest drought and heat. The occurrence and poor control of active fires can lead to significant environmental, economic, and safety issues for the population. These aspects can be impactful when they occur close to factories, farms, preserved areas, or even cities. This article presents a comparison of different techniques for segmenting fire hotspots using satellite images, with Heuristic and Deep Learning approaches. The results are compared on two satellite data, with different spatial resolutions and revisit times. The multispectral images come from S-NPP/VIIRS and Sentinel-2 sensors, exploring bands correlated to active fires. The results obtained through the Deep Learning approach showed that the model trained on Sentinel-2 images performed a better role in the segmentation task than the model trained with S-NPP/VIIRS images. When comparing the two approaches, heuristic segmentations proved to be more faithful to fire hotspots under certain conditions that are covered in this paper. Finally, a discussion of the different results obtained is presented.
