Mapeamento pedológico digital e uso de algoritmos de aprendizado de máquina em Tracuateua, Pará.

Abstract

The need for soil mapping is continuous. However, the traditional method does not allow for a quick and economical execution. Therefore, Digital Soil Mapping (DSM) is able to improve this process by spatializing the soil using models that quantify spatial variability. In this perspective, the objective of this study was to carry out the DSM of the city of Tracuateua, in the North of Brazil, evaluating the performance of machine learning algorithms, using a set of attributes derived from the Digital Elevation Model (DEM), satellite images and legacy data as input parameters. 244 pixels were randomly selected, distributed in nine Mapping Units (MU), where field prospecting was carried out, in order to avoid transition zones and confirm training data. The data were organized and processed in QGIS software, which is a Geographic Information System (SIG) and R programming language. The DEM derived the predictive covariates and subsequently selected the significant ones by the Recursive Feature Elimination (RFE), function of the caret package. The evaluation of the performance of the algorithms was performed using the confusion matrix, the Kappa index and global accuracy. The Random Forest, Ranger and C5.0 algorithms were considered moderate when mapping the soils of the study area, with Ranger having the best performance in pedological mapping in Tracuateua, in the northeast of Pará; both in the modeling stage with a Kappa index of 0.71 and global accuracy of 0.74, and in the comparison with the conventional map (reference), with results of a global Kappa index of 0.49 and global accuracy of 0.56. The Artificial Neural Networks (ANN) algorithm presented the lowest results in the modeling, with a Kappa index of 0.14 and accuracy of 0.26 and in comparison, with the reference map, with a global Kappa index of 0.35 and global accuracy of 0.48. Decision tree algorithms (Ranger, RF and C5.0) showed moderate potential for digital soil mapping at a scale of 1:100,000 in the municipality of Tracuateua, in northeastern Pará.