Green AI aplicada à detecção de incêndios florestais: comparação entre desempenho preditivo e eficiência energética de redes neurais convolucionais

Wildfire detection through remote sensing is essential for reducing environmental impacts. This work compares the ResNet50, VGG19, and MobileNetV3 architectures for wildfire detection using Landsat-8 and Landsat-9 imagery, incorporating Green AI metrics. ResNet50 achieved the best predictive performance, reaching 85% accuracy, an AUC of 0.95, and zero false negatives. VGG19 presented higher energy consumption and lower sensitivity to the wildfire class. MobileNetV3 stood out for its lower carbon emissions, reduced energy consumption, and shorter inference time, demonstrating its suitability for largescale and resource-constrained applications.