Análise orientada a objetos de imagens de satélite para mapeamento de áreas de preservação em reservatório hidrelétrico

• Ano de defesa: 2015
• Autor(a) principal: Soares, Jefferson Francisco
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: UNIVERSIDADE FEDERAL DE LAVRAS; DEG - Departamento de Engenharia; UFLA; BRASIL
• Palavras-chave em Português: Sensoriamento remoto; Quickbird; Matas ciliares; Algoritmo watersheds by immersion; Índice kappa; Exatidão global; Remote sensing; Riparian woods; Watershed by immersion algorithm; Kappa index; Global accuracy; CNPQ_NÃO_INFORMADO
• Link de acesso: http://repositorio.ufla.br/jspui/handle/1/9467

Considered one of the vegetative mitigation practices for water resource degradation, the maintenance of riparian woods is recommended and demanded by law. However, in Brazil, these areas are still uncharacterized. In light of this reality, it becomes necessary to widen researches that allow us to characterize these areas in an integrated manner, generating efficient and quick results with low cost. Remote sensing is the option that demonstrates great application potential. Thus, in this work, we aimed at mapping and characterizing soil use and occupation in permanent preservation areas at the Funil Hydroelectric Power Plant (Funil HEP) reservoir, using high spatial resolution satellite imaging – Quickbird – in true composition (RGB-321) allied to object-oriented analysis techniques. For image segmentation, based on the watersheds by immersion algorithm, we used the Envi EX® 4.8 software. In order to classify the image, we used the algorithms K-nearest neighbor, Support vector machine and Maximum Likelihood. We analyzed the accuracy of the mappings comparing the results obtained to the map generated with the visual classification of the image of the study area (reference map). With the results, we concluded that the K-nearest neighbor algorithm was the best for mapping soil use and occupation in the study area, with kappa index of 0.88 and global accuracy of 91.40%.