Avaliação do desempenho estatístico dos algoritmos de classificação Random Forest (RF), Decision Tree (DT) e Support Vector Machine (SVM) para imagens de satélite com diferentes cubos de dados: estudo de caso no bioma cerrado

• Ano de defesa: 2021
• Autor(a) principal: Carlos Eduardo Fernandes de Holanda
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; IGC - DEPARTAMENTO DE GEOGRAFIA; Programa de Pós-Graduação em Geografia; UFMG
• Palavras-chave em Português: Aprendizado de Máquina; Uso e cobertura da terra; Dados altimétricos; Sensoriamento Remoto; Aprendizado do computador; Solo – Uso; Mapeamento da cobertura do solo; Sensoriamento remoto
• Link de acesso: http://hdl.handle.net/1843/43270

The main objective of this work is to analyze and compare the performance of three algorithms based on machine learning: Random Forest (RF), Decision Tree (DT) and Support Vector Machine (SVM) in the classification of land use and land cover satellite images using altimetric information in different data cubes. In this sense, the results found in the statistical indices of validation and agreement between classes in the confusion matrix were interpreted, which can help students and researchers in the field of remote sensing to choose the classification algorithm that best suits their research. For this study, images from the Sentinel 2 satellite and another image that stacks those from Sentinel 2 with images derived from the Shuttle Radar Topography Mission (SRTM) were used in order to identify and interpret the differences in the classifications of these images. The classifier that obtained the best performance was the SVM, both for images with altimetric data and for images without this additional information. Although the DT presented satisfactory results, but inferior to the SVM, the DT classifier had much lower processing time for the classification of images than the SVM. Given this data, it is observed that, if the research area is very extensive or has several small areas combined with a computer with an unsuitable processor, the SVM classifier would not be the best option. The RF algorithm obtained practically the same processing time as the DT, but reached the lowest statistical indexes among the three classifiers. Furthermore, it was found that the use of the data cube with the SRTM image and its derivations such as slope and roughness in the classification of land use and land cover for the three algorithms showed superior results compared to the images without these altimetric data.