Predi??o de escorregamentos de encostas baseada em aprendizado de m?quina

• Ano de defesa: 2022
• Autor(a) principal: Pedreira, Laedson Silva
• Orientador(a): Calumby, Rodrigo Tripodi
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Feira de Santana
• Programa de Pós-Graduação: Programa de P?s-Gradua??o em Ci?ncia da Computa??o
• Departamento: DEPARTAMENTO DE TECNOLOGIA
• País: Brasil
• Palavras-chave em Português: Escorregamento; Aprendizado de m?quina; Minera??o de dados; Predi??o; Deslizamento de terra; Data Mining; Prediction
• Palavras-chave em Inglês: Random forest; LightGBM; Landslide; Machine Learning
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; METODOLOGIA E TECNICAS DA COMPUTACAO::SISTEMAS DE INFORMACAO; GEOTECNICA::MECANICAS DOS SOLOS
• Link de acesso: http://tede2.uefs.br:8080/handle/tede/1508

Landslides are among the main phenomena that cause natural disasters across the planet. Every year landslides have caused numerous material damages and claimed a large number of fatalities. In order to understand and describe the phenomenon of landslides, in addition to preventing or minimizing the problems caused by them, many studies have been carried out on their dynamics. However, considering the complexity of the problem and the scarcity of integrated and large-scale data, specific studies of individualized predictive models and with a temporal relationship, for monitoring and indicating risks are challenging. Despite this, the application of predictive models based on machine learning has great potential to contribute with effective and efficient tools, capable of assisting in the monitoring and prevention of damages arising from such events. In this context, this work proposes and experimentally evaluates data mining and machine learning techniques for the construction of a database from multiple sources, its pre-processing and the prediction of landslides individually, in time and in space. In addition, in order to verify the impact on the predictive capacity of the classifiers, the implications of two methods of generating non-slip samples, the number of days of accumulated rainfall considered and the lead time of prediction were analyzed. With the application of the methodology proposed here, it was possible to predict landslides in a promising way, with F1-score values greater than 0,929?0,002 and AUC greater than 0.930?0.002. The results presented also suggest that the use of these predictive models can contribute to a better decision-making by the competent about the regarding the monitoring and prevention of damage caused by landslides induced by rain.