Redes neurais convolucionais na segmentação semântica de imagens aéreas para o mapeamento da cobertura do solo em áreas de proteção ambiental

• Ano de defesa: 2023
• Autor(a) principal: Bizotto, Fabricio
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Ciência da Computação; UTFPR
• Palavras-chave em Português: Redes neurais (Computação); Imagens de sensoriamento remoto; Computação semântica; Conservação da natureza; Neural networks (Computer science); Remote-sensing images; Semantic computing; Nature conservation; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/33330

In Brazil, the Environmental Protection Area (EPA) plays a key role in environmental conservation and in balancing the sustainable use of natural resources with local socio-economic development. Effective management of EPA activities requires comprehensive monitoring of land use and land cover. Remote sensing has emerged as an efficient and economical alternative for monitoring these areas. In this context, advanced computational techniques, such as convolutional neural networks (CNNs), are emerging as promising tools. This study proposes a low-cost methodology for semantic segmentation, using remote sensing images acquired on the Google Earth platform to monitor land use and land cover in the EPA-Petrópolis, Rio de Janeiro, using RNCs. The SegNet and U-Net architectures were employed for semantic segmentation, along with the development of an aerial image database for the EPA-Petrópolis region, used to training and testing of the models. Four training and testing scenarios with different settings were presented and discussed. The analysis results indicate that scenario 4, using the U-Net with the Focal Loss function, achieved the highest overall accuracy (0.87). However, scenario 3, employing the U-Net with the cross-entropy loss function, showed comparable results (0.87). In terms of the Jaccard Index (IoU), scenario 3 stood out with the best value (0.72), while Scenario 4 was close (0.71). Challenging classes, such as Exposed Soil, showed low f1-score indices (0.31 to 0.52). The variation in the loss function between scenarios had limited impact. The results highlight the effectiveness of the U-Net and suggest the need for continuous refinements, especially in more challenging classes.