Digital mapping of sand, clay, and soil carbon by Random Forest models under different spatial resolutions

Abstract

The objective of this work was to evaluate the effect of the digital elevation model spatial resolution and of the efficiency of Random Forest models on the prediction of sand, clay, and organic carbon contents, using few soil samples. The study was carried out in a Cerrado area with lithological diversity, in the state of Mato Grosso do Sul, Brazil, using morphometric attributes, TM Landsat 5 sensor data, and lithology as
predictive covariates. The surface layer data (0.0–0.2 m) of 175 soil profiles (0,009 profiles km-2) and of 26 predictor covariates were used with 30 (set 1) and 90-m (set 2) spatial resolutions. The performed analysis by Random Forest models showed that channel base level, elevation, and lithology were the most important ones to explain the variability. The validation of the models showed similarity among sets for the prediction of sand,
clay, and organic carbon contents, which explains the following values of spatial variability, respectively: 44, 40, and 33%, for the spatial resolution of 30 m; and 45, 46, and 33%, for the spatial resolution of 90 m. The spatial resolution of the predictive covariates has little effect on attribute predictions, and the Random Forest approach has potential use for estimating soil properties.