Uso de técnicas de aprendizado de máquina para a elaboração de modelos de suscetibilidade à ocorrência de incêndios florestais nas áreas de proteção ambiental do Mosaico de Áreas Protegidas Sertão Veredas-Peruaçu, MG

• Ano de defesa: 2021
• Autor(a) principal: Dias, Lívia Caroline César
• Orientador(a): Moschini, Luiz Eduardo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciências Ambientais - PPGCAm
• Palavras-chave em Português: Unidades de conservação; Carbono; Incêndios florestais; Sensoriamento remoto; Veredas
• Palavras-chave em Inglês: Machine learning; Conservation units; Carbon; Wildfires; Remote sensing
• Área do conhecimento CNPq: CIENCIAS AGRARIAS::RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL::CONSERVACAO DA NATUREZA::CONSERVACAO DE AREAS SILVESTRES
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/15635

The Cerrado has been under strong pressure from the agricultural frontier in recent decades, with protected areas being almost the only remaining areas with some degree of conservation of this biome in Brazil. However, the imaginary lines that separate these conserved areas from anthropized areas have not been sufficient and as a consequence, wildfires have increasingly devastated these once protected environments. Geoprocessing tools have played an important role in enabling the mapping of these areas and in identifying more vulnerable areas, while machine learning techniques have made it possible to create models to predict the occurrence of wildfires in the near future in certain regions. In this work we seek to associate the geoprocessing tools with the machine learning techniques to be able to create for the Environmental Protection Areas of the Sertão Veredas Peruaçu Mosaic a forest fire predictive models. For this, we consider climatic variables (temperature, relative humidity, wind speed and precipitation), environmental (altitude, slope, NDVI, Hydrographic Distance), anthropogenic (Distance from Roads, Population Density, Distance of Ocupations and Land Use and Land Cover) and soil fuel moisture index (fine fuel moisture code, Duff moisture code and drought index) in order to explain when, where and why fire occurs in the region. In addition, these areas are composed of veredas environments of great relevance to the Cerrado and with an important role in the carbon stock, and for this reason we evaluate the quantification of organic carbon in these veredas. The choice of the best forest fire occurrence model by the logistic regression method was based on the Akaike Information Criterion. For each Environmental Protection Area (EPA), a final model was generated with a set of different variables, but all models showed high performance in predicting forest fires. The results of the final models showed that the Area Under the Curve (AUC), which is the model's hit rate, was greater than 95% for each of the three APA's. Another parameter used to assess the effectiveness of the models was the R² value, which presented values close to 1 in the final models, confirming that these models were adjusted. As for the concentration of organic carbon in the vereda soil, the results showed that this peat soil demonstrates the capacity to store carbon. The concentration of organic carbon in the arboreal area of the vereda is up to 20 times higher than in the cerrado strictu senso area.