Convolutional neural networks for semantic segmentation of aerial images in land cover mapping of environmental protection areas

Abstract

Environmental Protection Areas (EPAs) in Brazil are key legal instruments for conserving biodiversity, ensuring the sustainable use of natural resources, and supporting the socioeconomic development of local communities. Remote Sensing (RS) has emerged as a practical and cost-efficient alternative for monitoring these regions. In this context, advanced computational techniques - particularly Convolutional Neural Networks (CNNs) — have demonstrated strong performance in the semantic segmentation of RS imagery. This study proposes a low-cost and scalable methodology for land use and land cover mapping in the EPA-Petrópolis region (Rio de Janeiro), based on RGB images from Google Earth and CNN models. This work stands out by offering a low-cost and scalable methodology using RGB imagery from Google Earth, and by introducing the EpaPetroBR dataset–one of the first annotated semantic segmentation datasets focused on the Atlantic Forest biome. The SegNet and U-Net architectures were evaluated across four experimental scenarios. The best overall accuracy (0.87) was obtained in scenario 4, which employed U-Net with the Focal Loss function. Scenario 3, using U-Net with the cross-entropy loss function, achieved comparable accuracy (0.87) and the highest Jaccard Index (IoU) score (0.72). Despite these promising results, some classes — such as Exposed Soil — remained challenging, with F1-scores ranging from 0.31 to 0.52. The comparative analysis of loss functions indicated limited influence on overall performance, reinforcing the robustness of the U-Net architecture. The results highlight the potential of combining CNNs with freely available high-resolution imagery for environmental monitoring in tropical forest regions.