Mapeamento de áreas permeáveis na Bacia Hidrográfica do Córrego Bandeira usando aprendizado profundo

• Ano de defesa: 2022
• Autor(a) principal: Aurimar da Costa Lima Filho
• Orientador(a): Jamil Alexandre Ayach Anache
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Deep learning, sensoriamento remoto, cobertura do solo, áreas permeáveis, classificação de imagens de satélite
• Link de acesso: https://repositorio.ufms.br/handle/123456789/4739

Deep Learning is a sub-area of machine learning, applied to solve tasks involving image processing, computer vision, signal processing and natural language processing. It consists of a collection of algorithms as the machine learning subset, specializing in hierarchical learning of big data concepts. The objective of this work is to identify the permeable areas of the Córrego Bandeira Watershed, using aerial and orbital images and Deep Learning methods. In recent years, many studies are using this tool to process remote sensing images, as it has shown to be efficient in processing both optical (hyperspectral and multispectral) and radar images, presenting accurate results in mapping different types of land cover. Identifying and mapping permeable areas from and/or inserted in urban areas, using remote sensing, is relevant for hydrological and microclimate studies, as well as for their protection, monitoring and management. Using the U-Net network and the ResNet-34 backbone in the aerial image with a spatial resolution of 10 cm, the architecture achieved the classification with 55 epochs (epochs), an accuracy of 0.84 and an F1-Score of 0.88, such as the orbital image, with a spatial resolution of 50 cm, using the same architectural models, reached the classification with 27 epochs (epochs), accuracy of 0.796 and F1-Score of 0.86. Comparing the classifications, it is observed that the classification using image of spatial resolution 10 cm, there was segmentation, labeling and significant identification. While the classification using 50 cm spatial resolution image, there was only segmentation, labeling and identification of large-scale areas. Note that the spatial resolutions interfered in the training and classification results. As much as both are conceptualized as high resolution images, the orbital image, with a resolution of 50 cm, proved to be unsuitable for this type of study, in which it achieved an unsatisfactory rating.