Spatial prediction of soil properties in two contrasting physiographic regions in Brazil

Abstract

This study compared the performance of ordinary kriging (OK) and regression kriging (RK) to predict soil physical-chemical properties in topsoil (0-15 cm). Mean prediction of error and root mean square of prediction error were used to assess the prediction methods. Two watersheds with contrasting soil-landscape features were studied, for which the prediction methods were performed differently. A multiple linear stepwise regression model was performed with RK using digital terrain models (DTMs) and remote sensing images in order to choose the best auxiliary covariates. Different pedogenic factors and land uses control soil property distributions in each watershed, and soil properties often display contrasting scales of variability. Environmental covariables and predictive methods can be useful in one site study, but inappropriate in another one. A better linear correlation was found at Lavrinha Creek Watershed, suggesting a relationship between contemporaneous landforms and soil properties, and RK outperformed OK. In most cases, RK did not outperform OK at the Marcela Creek Watershed due to lack of linear correlation between covariates and soil properties. Since alternatives of simple OK have been sought, other prediction methods should also be tested, considering not only the linear relationships between covariate and soil properties, but also the systematic pattern of soil property distributions over that landscape.