ТОПЫРАҚ ПРОФИЛІНІҢ 0–200 СМ ТЕРЕҢДІКТЕГІ СТРАТИФИКАЦИЯСЫН КӨПДЕҢГЕЙЛІ СТЕКИНГ-МОДЕЛІМЕН АНЫҚТАУ

Abstract

In contemporary agroecological research and sustainable land management, accurate characterization of the vertical structure of soil profiles remains a critical task. Conventional approaches based on field drilling and laboratory analysis are time-consuming, costly, and spatially limited, which restricts their applicability at larger scales. This study proposes an automated soil stratification framework that integrates Sentinel-2 multispectral imagery, ERA5-Land climatic variables, and OpenLandMap static soil datasets. A multi-task stacking ensemble was implemented to jointly predict quantitative soil properties such as clay and sand content and bulk density as well as categorical variables including texture and land cover classes. The modeling framework combined Random Forest, Gradient Boosting, and XGBoost as base learners, while linear and logistic regression models were used at the meta-learning stage. The experimental evaluation conducted in the Bozaigyr Lake Valley (Kazakhstan) demonstrated strong predictive performance. For clay and sand content, the coefficient of determination reached R² = 0.999–1.000, with mean absolute errors of approximately 1.0–1.2 %. Bulk density predictions yielded R² values between 0.985 and 0.996. Overall classification accuracy ranged from 97.4 % to 99.7 % for texture classes and was close to 99 % for soil_horizon_class. Misclassifications were primarily observed between spectrally similar categories. The results indicate that a stacking-based ensemble integrating multispectral, climatic, and static soil information can provide an efficient and scalable solution for digital soil mapping, particularly in arid and semi-arid environments.