Título:	Deforestation detection in sar images using deep neural networks
Autor(es):	Correia, Igor Bispo de Moraes Coelho
Orientador(es):	Farias, Mylène Christine Queiroz de
Assunto:	Sensoriamento remoto Redes neurais convolucionais (Computação) Radar de Abertura Sintética (SAR) Desmatamento
Data de publicação:	14-Nov-2024
Data de defesa:	3-Jun-2024
Referência:	CORREIA, Igor Bispo de Moraes Coelho. Deforestation detection in sar images using deep neural networks. 2024. 86 f., il. Dissertação (Mestrado em Engenharia Elétrica) — Universidade de Brasília, Brasília, 2024.
Abstract:	Deforestation has broad and significant impacts, becoming a major global environmental threat. This problem endangers biodiversity, contributes to climate change, and compromises the sustainability of natural ecosystems. Therefore, monitoring and detecting deforested areas is a critical issue. Although deforestation affects many regions, the Amazon rainforest is one of the most prominent and frequently discussed cases. Optical satellites are extremely powerful and important tools in remote sensing, prevention, and mitigation of deforestation. However, this type of sensor is not robust to climatic variations and is sensitive to cloud occlusion. On the other hand, synthetic aperture radars (SARs) stand out for their resilience to adverse weather conditions, as they are active sensors that operate in microwave bands which penetrate water particles. Nonetheless, accurately recognizing deforested areas in SAR images is challenging due to the amount of speckle noise and the variability in the appearance of objects between different captures. In this study, we conducted an online experiment with voluntary participants who identified deforested areas in SAR images. For this purpose, we developed software that allows participants to annotate SAR images, delineating deforested areas. With the results of this experiment, it was possible to analyze the relationship between the participants’ self-reported experience level and the accuracy in detecting deforested areas. We also compared human performance with the performance obtained from an automatic model based on the UNet architecture. The results show that greater knowledge in remote sensing or SAR does not guarantee quality annotations. Moreover, the UNet’s performance surpasses human performance in the task. To explore SAR image segmentation in greater depth, a second experiment was conducted using state-of-the-art models for segmentation in fused SAR and optical data. This second part of the study showed that the most modern models, despite having a smaller number of trainable parameters, can outperform a heavier model. The study reinforces the potential of deep learning in deforestation detection, emphasizing the need for continuous improvements in architectures and specialist training.
Unidade Acadêmica:	Faculdade de Tecnologia (FT) Departamento de Engenharia Elétrica (FT ENE)
Informações adicionais:	Dissertação (mestrado) — Universidade de Brasília, Faculdade de Tecnologia, Departamento de Engenharia Elétrica, 2024.
Programa de pós-graduação:	Programa de Pós-Graduação em Engenharia Elétrica
Licença:	A concessão da licença deste item refere-se ao termo de autorização impresso assinado pelo autor com as seguintes condições: Na qualidade de titular dos direitos de autor da publicação, autorizo a Universidade de Brasília e o IBICT a disponibilizar por meio dos sites www.bce.unb.br, www.ibict.br, http://hercules.vtls.com/cgi-bin/ndltd/chameleon?lng=pt&skin=ndltd sem ressarcimento dos direitos autorais, de acordo com a Lei nº 9610/98, o texto integral da obra disponibilizada, conforme permissões assinaladas, para fins de leitura, impressão e/ou download, a título de divulgação da produção científica brasileira, a partir desta data.
Aparece nas coleções:	Teses, dissertações e produtos pós-doutorado

Os itens no repositório estão protegidos por copyright, com todos os direitos reservados, salvo quando é indicado o contrário.