Akademik Veri Yönetim Sistemi
Advances in Space Research, 2026 (SCI-Expanded, Scopus)
Accurate prediction of soil textural properties is essential for understanding soil functions and supporting digital soil mapping (DSM), particularly in semi-arid and topographically complex landscapes. This study aimed to predict sand, silt, and clay fractions, as well as USDA soil texture classes, at surface (0–30 cm) and subsurface (30–60 cm) depths by integrating visible–near infrared (vis–NIR) spectroscopy with environmental covariates and machine learning (ML) models. A total of 214 soil samples were collected in the Eastern Mediterranean region of Türkiye using a conditional Latin Hypercube Sampling (c-LHS) strategy. Environmental variables in the c-LHS sampling were digital elevation model (DEM), Slope, Aspect, Mean Annual Precipitation (MAP), Mean Annual Temperature (MAT), Landform, Parent Material and Land Use/Land Cover (LULC). Spectral data were pre-processed and used in conjunction with environmental covariates derived from Sentinel-2 imagery and DEM. Regression models (XGBoost (XGB), Random Forest (RF), and Support Vector Regression (SVR)) and classification models (Multinomial Logistic Regression (MNLR), C5 Decision Tree (C5DT), and RF) were evaluated. Among the regression approaches, XGB achieved the highest predictive performance, particularly for silt and clay fractions, with higher R2 and RPIQ values and lower bias. SHAP analysis identified soil organic carbon, land use, and soil depth as the dominant drivers of particle-size variability. Texture class maps generated by RF were more spatially coherent and pedologically consistent, although overall classification accuracy was constrained by class imbalance and complex horizon transitions. The results demonstrate that combining vis–NIR spectroscopy with ML substantially improves the prediction of soil particle-size distribution, providing a reliable and scalable framework for DSM applications in data-scarce semi-arid Mediterranean environments.