Segmentação Semântica de Incêndios Florestais com U-Net e Encoders Pré-Treinados: Uma Abordagem com Função de Perda Composta e Balanceamento de Classes

Abstract

emantic segmentation of forest fires in satellite imagery is a relevant problem for environmental monitoring, particularly due to severe class imbalance and the sparse spa- tial distribution of fire events. In this work, the study is conducted using multispectral Sentinel-2 images with 20 m spatial resolution, focusing on the accurate identification of the fire class. The objective of this dissertation is to propose, implement, and eva- luate a semantic segmentation pipeline for forest fire detection, investigating the impact of different backbones, loss functions, and class imbalance mitigation strategies. The methodology is structured into two experimental cycles. The first, exploratory in nature, evaluates a U-Net architecture with VGG16 and EfficientNetB3 encoders using conventi- onal loss functions. The second cycle adopts an optimized pipeline incorporating selective undersampling, data augmentation, loss functions tailored to imbalanced data (Dice, Fo- cal, and Dice+Focal), a two-stage training strategy with encoder fine-tuning, and decision threshold calibration through validation-based sweeping to maximize the F1-score. The results show substantial improvements in the second cycle, with increased stability and performance, particularly for the U-Net with the VGG16 backbone, which achieved high F1 and IoU scores on the test set. Qualitative analysis further confirmed improved spatial coherence of the predicted masks. These findings indicate that combining class balancing strategies, appropriate loss functions, and threshold calibration is essential for effective forest fire segmentation in multispectral Sentinel-2 imagery.