DIGITAL MAPPING OF SOIL CLASSES COMPARING MACHINE LEARNING ALGORITHMS IN INTERIOR TABLELANDS

Land use management requires knowledge about the properties and attributes that characterize. The objective of this work was to test the use of machine learning in soil prediction, aiming to compare the respective performances. The selected area is located in the region of Tabuleiros Interioranos of Recôncavo da Bahia, Brazil. The adopted method used for training the observation points and legacy data models, the variables were generated through maps of source material, land use, geomorphometric data and satellite images. In the R software, prediction calculations were performed. The tested algorithms were: Random Forest (RF), Decision Trees (C5), K-Nearest Neighbors (KNN), Gradient Boosting Machines (GBM) and Support-vector machine (SVM). The accuracy of digital mapping was measured using the index Kappa and general accuracy. The C5 model showed the highest kappa and accuracy index.