Geological and terrain attributes for predicting soil classes using pixel- and geographic object-based image analysis in the Brazilian Cerrado

Abstract

One of the great challenges in the geosciences is mapping targets and phenomena that are not fully in direct range of imaging instruments such as remote sensors. Thus, soil maps are still scarce, especially at the detail level for large regions, with the few official maps being restricted to small spatial cutouts, resulting in an increasing demand for pedological information. The aim of this work was to analyze the use of covariates related to geological and terrain attributes on the performance of soil class predictive models for digital soil mapping with pixel- and geographic object-based image analysis (GEOBIA) approaches in the context of sub-humid tropical environments of the Cerrado biome, in central-western Brazil, to reproduce a conventional map. A reference soil map at the scale 1:50,000 with twelve soil classes was used. Randomly stratified samples and the Random Forest (RF) classifier was performed with seventeen digital elevation model derivatives and the geology categorical data (lithostratigraphic units) as covariates for the prediction of soil classes in both approaches. The main results demonstrated that GEOBIA is a promising method for mapping soil classes, with overall accuracy of 0.70, which is 5% higher than pixel-based. The most important covariates for soil class mapping were related to Geology and terrain attributes: Altitude, Vertical Distance to Channel Network, Vector Terrain Ruggedness, Relative Slope Position, Topographic Height, Profile Curvature, Roughness, Slope, and Valley Depth. The integration of the geological covariate improved the models in both approaches. Integrated evaluation of the geological units and terrain attributes along with the use of machine learning techniques, such as the RF algorithm, is a promising and relevant strategy for digital predictive mapping of soil classes that can contribute toward the greater reproducibility of conventional soil maps.