This study compared the prediction accuracy of random forest, quantile regression forest, support vector machine and ensemble learning in mapping soil thickness taken as a continuous variable, where the machine learning models were weighted as individual models. Furthermore, a feature-ensemble learning algorithm was proposed for mapping soil thickness, in which soil thicknesses was classified as a new categorical variable, and the discrete predictions were further weighted with the predicted continuous soil thicknesses. The results showed that soil thicknesses in Sichuan Province were characterized with high spatial variation, of which the dominated drivers included multiresolution index of valley bottom flatness, elevation and topographic wetness index. The overall performance of prediction models in terms of coefficients of determinations and root mean square errors were 0.32-0.47 and 0.28-0.41 m, respectively. For the prediction of continuous soil thickness, ensemble models had low errors than those of individual models. For soil thickness types, the proposed feature-ensemble learning algorithm achieved higher robustness than other considered models by reducing the variance of prediction.