Técnicas Machine Learning y Deep Learning en la clasificación de coberturas del suelo: Estudio de caso municipio de Facatativá usando imágenes Sentinel-2

Abstract

In recent years, Geographic Information Systems (GIS) have made significant progress in the development of supervised classification methods, as they have been implementing computational approaches that allow for the automatic, fast, accurate, and cost-effective classification of land cover in satellite images. In this work, two supervised classification methods are presented and compared, with the aim of evaluating their efficiency in generating land cover maps from Sentinel-2 imagery of the municipality of Facatativá. The first model developed was the Support Vector Machine (SVM), which is considered a classical Machine Learning method. The second model implemented belongs to the field of Deep Learning and is known as U-Net. To establish the accuracy and reliability of the classification obtained by each model, different evaluation metrics were calculated, such as overall accuracy, per-class accuracy, recall, and the Kappa index. The results obtained from these evaluation metrics showed that the model that achieved the best land cover classification was the Support Vector Machine (SVM), with an overall accuracy of 94.76% and a Kappa index of 0.94.