Deep Learning-Based Active Fire Detection Using Satellite Imagery

The detection of active fire from satellite images is of extreme importance for the implementation of environmental conservation policies and fostering decision-making. In recent years, deep learning models have enjoyed enormous success in various fields, but their use for active fire detection is relatively new. In this manner, this paper addresses analyzing the feasibility of using deep learning for active fire detection. This study identifies, through a CNN of type U-Net, active fire segmentation in Sentinel-2 and Landsat-8 satellites images using the masks produced by methods known in the literature and commonly used for active fire detection. More specifically, this study comprises 100 scenes from the Sentinel-2 satellite and 23 from the Landsat-8 satellite in the period between January 2021 and June 2023. The validation set produced promising results, with an accuracy of 97.98%, although with five misclassifications among the fifteen runs. The accuracy for Sentinel-2 images stood at 97.73%. For Landsat-8 image tests, the accuracy reached 90.22%, while the ensemble model achieved accuracy of 99.70%. These results contribute to the advancement of fire detection expertise, providing support for environmental preservation and related policy implementations.