Digital Soil Subgroup Mapping in Semi-Arid Mountainous Regions Using Multi-Source Environmental Covariates

Digital Soil Mapping (DSM) provides a modern framework for assessing soil variability in complex landscapes. This study evaluated the performance of Random Forest (RF) and Support Vector Machine (SVM) classifiers for predicting soil subgroups in a semi-arid mountainous area of 3,000 ha, based on 81 soil profiles and multiple environmental covariates. Two predictor sets were tested: (i) topographic indices and spectral data, and (ii) an extended set including subsurface soil properties. Results showed that both models benefited from the integration of soil profile variables, with RF outperforming SVM. The most influential predictors were soil depth, clay content of second horizon, and a sentinel carbonate index derived from satellite data, which capture key pedogenic processes in semi-arid conditions. The final soil map indicated that Typic Calcixerepts and Lithic Xerorthents were the most widespread subgroups. Overall, the study demonstrates that combining pedological information with environmental covariates significantly improves DSM performance, offering more reliable soil maps for land evaluation and sustainable management in mountainous semi-arid regions.