High-resolution soil depth maps are valuable for environmental modelling, yet reliable data remains scarce in the tropics. This study evaluates the feasibility of mapping depth to bedrock in Sri Lanka using a legacy dataset (n = 88) and global environmental covariates (n = 247). A machine-learning workflow was used, including feature selection, hyperparameter tuning, and a stacked ensemble of Random Forest, XGBoost, Cubist, and a support vector machine. Under spatial cross-validation the ensemble reached R² = 0.197 and RMSE = 35.4 cm. That is modest, and it outperforms existing global products. Global covariates at about 1 km capture regional trends, and they do not resolve local variation below 50 m. Further gains need higher-resolution local data rather than a more complex model.