Optimizing soil texture spatial prediction in the Brazilian Cerrado: Insights from random forest and spectral data

Abstract

Soil texture is crucial for assessing soil quality, crop suitability, and land management. However, precise large-scale soil texture mapping remains challenging. This study integrated a Synthetic Soil Image (SySI) with standard environmental covariates in a digital soil mapping framework to map soil particle size distribution in Brazil's Cerrado biome. Four random forest model arrangements were explored for soil texture modeling. Using an extensive legacy dataset of Cerrado topsoil (0–20 cm), the most accurate model explained approximately 83 % of clay, 86 % of sand, and 74 % of silt variance, with RMSE values of 89 g kg⁻¹ (clay), 102 g kg⁻¹ (sand), and 53 g kg⁻¹ (silt). The findings revealed that elevation and bioclimate have strong predictive capacities, especially when bare soil spectra data are available. Elevation was the only relevant terrain derivative for predicting soil texture. This approach improved model interpretability and provided high-resolution, accurate soil texture maps, aiding users and public policies. Since 65 % of Cerrado soil classes (Ferralsols and Arenosols) do not significantly increase clay content with depth, this work adds value to agricultural soil mapping.