Seleção de Co-Variáveis para Classificação de Imagens de Satélite Através do Algoritmo Random Forest

This papers aims to present a method to select co-variables using the Random Forest classifier to classify satellite images, by the land use classification of Lagoa Formosa, a county in Minas Gerais state. The data used was Landsat 8 images for the dry and wet seasons, cartography data of the county obtained from IBGE, such as geomorphology, vegetation, boundary, localities and roads at the 1:250.000 scale, and a geologic map of the area at 1:100.000 scale and a soils map at 1:250.00 scale. The SRTM was used to obtain the digital elevation model and other topography co-variables. With these in hand, 98 co- variables were obtained, being 26 spectral co-variables, 45 topography co-variables, 5 geomorphological co-variables, 2 geological co-variables, 3 soils co-variables, 3 vegetation co-variables and 14 Euclidian distance co-variables. The numeric co-variables were analyzed by non-linear correlation, and the categorical co-variables were analyzed by dissimilarity to eliminate those statistically similar. From the 58 co- variables analyzed by the Random Forest classifier, a list with the models with 20 to 5 co-variables was generated with the Kappa index. The selected model was that with the highest value for the Kappa index, the one with 16 co-variables, to obtain the simplest model that better explain the final classification of land use. However, it is still necessary to develop other techniques that help decrease the researcher subjectivity to choose the final model.