Transfer Learning in Convolutional Neural Networks for Fire Scene Classification in the Pantanal Biome (Brazil)

During 2024, fires in the Pantanal, Brazil, posed a significant threat to the world’s largest wetland. This study presents a scene classification approach using Sentinel-2 images of fire-affected and unaffected areas, based on deep learning. First, a model was experimentally trained with the EuroSAT dataset to classify pasture and herbaceous vegetation. Then, transfer learning was applied to the Pantanal fire affected images. Finally, VGG-19 was trained from scratch without pre-trained parameters. Results show that the transfer learning model achieved 95% accuracy, while VGG-19 from scratch reached 97%. These findings indicate that intra-domain transfer learning is effective in remote sensing, even with spectral and spatial differences, and can support environmental monitoring in ecologically complex regions such as the Pantanal. This method allowed us to determine that intra-domain transfer learning is effective in remote sensing, even in the presence of spectral and spatial differences, and can be used for monitoring tropical biomes with high spatial and spectral heterogeneity.