Accurate mapping of subsurface soil texture is important for sustainable land management and precision agriculture in semi-arid environments, where soil observations are often spatially sparse. This study evaluates a multi-temporal Digital Soil Mapping framework combining wet-season Sentinel-1 Synthetic Aperture Radar (SAR) and dry-season Sentinel-2 optical observations for mapping clay, silt, and sand fractions at 30–40 cm depth in the Tal Kaif district of northern Iraq. Four hyperparameter-optimized machine-learning algorithms—Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Artificial Neural Network (ANN/MLP)—were benchmarked using strict five-fold out-of-fold validation. Pure machine-learning models showed limited independent predictive performance for all three fractions, with negative R² values and RPD values below 1.0. The best pure-ML models achieved RMSE values of 9.726% for clay, 9.201% for silt, and 6.890% for sand. Residual variograms demonstrated moderate to strong spatial dependence, supporting the incorporation of spatially structured residuals through Regression Kriging (RK). Hybrid RK produced fraction-specific effects: RMSE decreased slightly for clay (9.726% to 9.689%) and more clearly for sand (6.890% to 6.702%), whereas silt performance deteriorated slightly (9.201% to 9.334%). The results therefore provide partial support for the hypothesis that geostatistical modelling can complement pure machine learning, but do not demonstrate a universal improvement from RK. The principal limitation is the indirect relationship between surface remote-sensing observations and subsurface texture at 30–40 cm depth, together with the sparse sampling network. Nevertheless, the resulting spatially continuous maps identified a predominance of fine-textured classes, particularly Silty Clay Loam and Silty Clay, providing spatial information relevant to irrigation planning, precision agriculture, and soil-management decisions. The study contributes evidence that integrating environmental feature-space modelling with spatial residual structure can provide a more spatially informed framework for subsurface soil-texture mapping in data-sparse semi-arid environments, while highlighting the need for fraction-specific evaluation and independent validation in future studies.