Prediction of soil classes in a complex landscape in Southern Brazil

Abstract

The objective of this work was to evaluate the use of covariate selection by expert knowledge on the performance of soil class predictive models in a complex landscape, in order to identify the best predictive model for digital soil mapping in the Southern region of Brazil. A total of 164 points were sampled in the field using the conditioned Latin hypercube, considering the covariates elevation, slope, and aspect. From the digital elevation model, environmental covariates were extracted, composing three sets, made up of: 21 covariates, covariates after the exclusion of the multicollinear ones, and covariates chosen by expert knowledge. Prediction was performed with the following models: decision tree, random forest, multiple logistic regression, and support vector machine. The accuracy of the models was evaluated by the kappa index (K), general accuracy (GA), and class accuracy. The prediction models were sensitive to the disproportionate sampling of soil classes. The best predicted map achieved a GA of 71% and K of 0.59. The use of the covariate set chosen by expert knowledge improves model performance in predicting soil classes in a complex landscape, and random forest is the best model for the spatial prediction of soil classes.