Segmentação de Fogo Ativo em Imagens Landsat-8 Utilizando U-Net

Abstract

Wildfires represent a growing threat to biodiversity, human health, and global cli mate stability, driven by climate change and uncontrolled land use. In Brazil, the burned area increased by 79% in 2024, highlighting the urgent need for effective technological systems for environmental monitoring. This study investigates the use of the U-Net ar chitecture for active fire segmentation in multispectral images from the Landsat-8 satellite, focusing on highly imbalanced scenarios where burned areas account for a small fraction of the image. The model was trained with masks generated by traditional algorithms and evaluated using expert-annotated data, employing metrics such as F1-Score, IoU, preci sion, and recall. The results indicate high sensitivity (0.89) and precision (0.88), an IoU of 0.79, and a Matthews Correlation Coefficient of 0.88, demonstrating the robustness of the model even in challenging scenarios. It is concluded that U-Net is a promising approach for burned area segmentation, especially when combined with specific training strategies and appropriate handling of class imbalance, making it applicable to large-scale environmental monitoring and early warning systems.