Mapeamento da cobertura da terra com aprendizado de máquina e aplicações no monitoramento de lagos artificiais

Abstract

Land use and land cover mapping has several applications and great importance for environmental and hydrological issues. In the State of Mato Grosso, hydroelectric power plants were built along the Teles Pires River, which led to significant changes in land use and land cover, and to provide an accurate mapping of changes in this type of area, this study assessed different satellite image classification models. Multispectral images from the Landsat-8 and Sentinel-2 satellites were used, and classifiers based on Machine Learning such as SVM, RF, KNN and LDA. Image preprocessing was done using Google Earth Engine. The samples were created using QGIS software, the classifications were executed in R software.The performance evaluation considered the balanced accuracy and the F-Score index for each model. The best models were Sentinel-2 SVM and Landsat-8 RF. In addition to providing information for classifying land use in similar areas and serving as a framework for future studies in the region, the results showed a worrying scenario in the Colíder HPP area due to the expansion of the flooded area and loss of vegetation, and a more positive scenario in the Sinop HPP, with recovery of vegetation on the banks of the reservoir.