Previsão de variáveis ​​climáticas e classificação de risco de incêndios florestais no Pantanal brasileiro

• Ano de defesa: 2023
• Autor(a) principal: Zamith Santos, Bruna
• Orientador(a): Cerri, Ricardo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Aprendizado de máquina supervisionado; Classificação; Previsão de séries temporais; Algoritmos genéticos; Risco de incêndio florestal; Supervised machine learning
• Palavras-chave em Inglês: Classification; Time series forecasting; Genetic algorithms; Forest fire risk
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/19084

The Pantanal biome is of inestimable ecological importance, mainly due to its fauna and flora. However, this biome is constantly threatened by the occurrence and recurrence of forest fires, which are strongly associated with the regions climatic conditions. Thus, methods for forecasting and classifying forest fire risk are essential for forest fire prevention and firefighting planning in the Brazilian Pantanal region. The main fire risk indexes known in the literature have limitations, such as (1) not adjusting to the characteristics of each biome; (2) being limited to specific climatic variables; (3) not being able to predict forest fire risk for a given number of days. This last aspect, in particular, is of utmost relevance. Addressing it allows for coordinated planning and action by environmental authorities with adequate anticipation. Aiming to solve this problem, this study developed a software capable of: (1) Climatic variables forecasting for a given number of days; and (2) Forest fire risk classification in the Brazilian Pantanal. For the first objective, different time series prediction algorithms based on Machine Learning (ML) were tested for climactic variables forecasting. This prediction is used as input for the second objective, for which different classification algorithms, also based on ML, were tested. Such software was then improved from a exhaustive hyperparameters search approach to two different Genetic Algorithms (GAs) approaches: Traditional and NSGA-II. Results for both software versions were evaluated based on the average correlation between forest fire risk classes and hotspots' observation. The exhaustive search version demonstrated that the software can outperform the main statistical forest fire indexes regarding "Null", "Low", "High" and "Very High" classes. When it comes to the GA version, the software was competitive to the forest fire indexes, still with the advantage of being able to predict the forest fire risk for a given number of days.