Métodos de aprendizado de máquina para o mapeamento de florestas de Pinus utilizando dados de alta resolução espacial e elaboração de um mapa de risco de incêndio

• Ano de defesa: 2023
• Autor(a) principal: Ferla, Andressa Kossmann
• Orientador(a): Breunig, Fábio Marcelo
• Banca de defesa: Silva, Ricardo Dal’Agnol da, Pereira, Gabriel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; UFSM Frederico Westphalen
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência e Tecnologia Ambiental
• Departamento: Ciências Ambientais
• País: Brasil
• Palavras-chave em Português: Sensoriamento remoto; Classificador; Aprendizado de máquina; Mapa de risco de incêndio
• Palavras-chave em Inglês: AHP; Remote sensing; Classifier; Machine learning; Fire risk map
• Área do conhecimento CNPq: CNPQ::OUTROS::CIENCIAS
• Link de acesso: http://repositorio.ufsm.br/handle/1/29137

Native and planted forests are essential across the planet and provide many benefits to life on earth. However, the magnitude of its benefits depends on ongoing environmental changes. In this sense, forest fires cause economic losses and changes in the environment. In view of this, environmental monitoring is indispensable, which has made use of remote sensing techniques. This work aims to identify the use of remote sensing techniques for mapping pine plantations and then develop a fire risk map. The study was divided into two stages: a) first, it verified which machine learning classifier (Radom Forest or Support Vector Machine) was more effective in classifying land use and land cover and which is the best time of year for mapping forests of Pinus spp. in an area located in the municipality of São José do Norte in Rio Grande do Sul, for this purpose PlanetScope high resolution image was used; b) preparation of a forest fire risk map. The results showed that the two classifiers obtained good results, however Random Forest was more efficient in the spring and summer seasons. The work achieved adequate accuracy and can be reliably used for monitoring and managing land cover in the study region. The results of the second stage used remote sensing techniques and the Analytic Hierarchy Process – AHP method to create a fire risk map in the same region to identify areas close to forests that are prone to the occurrence of forest fires. For the creation of the risk map, the map of land cover and use, topographical, climatic and anthropogenic variables was used as a variable. Finally, the map was validated through fire outbreaks detected by the INPE Burn Program where it was identified that these outbreaks were close to areas classified as High, Very High and Extreme Risk in the risk map, in this way the map can be used in the management and monitoring to prevent or minimize the impacts caused by forest fires.