Evaluating Deep Segmentation Models for Wildfire Mapping Using Sentinel-2 Satellite Imagery

Wildfires pose serious threats to ecosystems, human health, and urban infrastructure, making accurate and efficient mapping of burned areas essential for mitigation and recovery. In this context, the use of computer vision techniques has shown highly effective results for rapid and accurate delineation of burned areas. However, these techniques require considerable computational cost, and their performance is sensitive to model selection and hyperparameter adjustments. Therefore, this work evaluates deep learning models for the semantic segmentation of regions affected by wildfires using multispectral images from the CEMS-Wildfire dataset, obtained from the Sentinel-2 satellite. We evaluated the combination of three traditional architectures: U-Net, Feature Pyramid Network (FPN), and SegFormer, combined with different encoders (ResNet-50, EfficientNet B2, and MiT-B0), under different data augmentation strategies (none, mild, strong) and loss functions (Binary cross-entropy, Dice loss, and hybrid BCE + Dice loss). U-Net achieved the best performance with the MiT B0 encoder, trained using Dice loss and moderate augmentation, resulting in an IoU of 0.7790 and an F1-score of 0.8758.