Transfer learning in convolutional neural networks for fire scene classification in the Pantanal Biome (Brazil)

During 2024, wildfires in the Pantanal, Brazil, posed a significant threat to the world’s largest wetland. This study presents a scene classification approach using Sentinel-2 imagery from fire-affected and non-affected areas, based on deep learning techniques. First, a model was experimentally trained using the EuroSAT dataset to classify grasslands and herbaceous vegetation. Subsequently, a transfer learning technique was applied to images of the fire-affected Pantanal. Finally, the VGG-19 architecture was trained from scratch, without using pretrained parameters. Considering these three experiments, training with the EuroSAT dataset achieved a loss of 0.202 after 50 epochs. The model with transfer learning, using the weights learned during the experimental training, obtained a validation loss of 0.15 and an accuracy of 95%, whereas training from scratch reached a validation loss of 0.07 and an accuracy of 97%. The results indicate similar performance between the two strategies, demonstrating that in-domain training can be considered an effective method for training CNNs applied to satellite imagery, without compromising classification accuracy, even in the presence of significant differences in the contextual information of the images. This approach is presented as a methodological alternative for environmental monitoring in complex tropical regions, characterized by high spatial and spectral heterogeneity.