Evaluating Digital Soil Mapping Approaches for 3D Mapping of Soil Organic Carbon

Whileland resource management needs detailed and accurate information about soil properties and distribution, this kind of data is limited in Iran. In this research, we tested performance of three digital soil mapping (DSM) approaches including Multiple Linear Regression (MLR), Cubist (CU) and Random Forest (RF) to map the spatial 3D distribution of soil organic carbon (SOC) in Saadat Shahr plain in Fars Province. Latin hypercube sampling (LHS) was used to determine locations of soil profiles in the field. The soil profiles were sampled and SOC was measured. Different environmental covariates including terrain attributes, remote sensing auxiliary variables, and maps of soil, geoform and distance from rivers were used in this research as auxiliary data. According to the link of the environmental covariates and soil organic carbon contents in the framework of each model in combination with equal-area spline algorithm, soil organic carbon maps were produced at five standard depths of soils in the whole study area. Model performance was evaluated by root-mean-square error (RMSE), mean error (ME) and normalized root-mean-square error (NRMSE). Among the used models, RF model showed the highest performance to predict organic carbon in depths of 0-5 and 60-100 cm. Meanwhile, MLR and CU had the lowest error for prediction in depths of 5-15 and 15-30 cm, respectively. In spite of these results, RF model was considered as the best model for its power to explain the spatial distribution of soil organic carbon in all soil depths in the study area.