Recognition of Soil Profile Stratification Using Machine Learning and Remote Sensing Methods

This study presents a novel machine learning–based framework for recognizing soil profile stratification and estimating key soil properties in arid and semi-arid regions of Kazakhstan. A multi-task stacking ensemble integrating Random Forest, Gradient Boosting, and XGBoost models was developed to jointly predict quantitative indicators (clay and sand content, bulk density) and categorical variables (textural classes and soil horizon classes) down to 200 cm depth. The predictor space combines Sentinel-2 multispectral reflectance, ERA5-Land climate parameters, Digital Elevation Model (DEM)-derived topographic attributes, and static soil–environmental covariates. Independent ground-truth validation across the Aral, Bozaigyr, and Atyrau regions demonstrates high predictive reliability (R² = 0.86–0.97 for quantitative properties and 94–98 % accuracy for classification tasks). SHAP analysis revealed that topography, SWIR bands, and hydrological gradients are the dominant predictors governing vertical soil differentiation. The integration of topographic and spectral features within the SCORPAN framework ensures both methodological transparency and pedological interpretability. The proposed approach advances digital soil mapping by providing a scalable and transferable solution for stratified soil profile recognition in data-sparse environments.