Soil thickness (ST) plays an important role in regulating soil processes, vegetation growth and land suitability. Therefore, it has been listed as one of twelve basic soil properties to be delivered in GlobalSoilMap project. However, ST prediction has been reported with poor performance in previous studies. Our case study is located in the intensive agriculture Beauce area, central France. In this region, the ST mainly depends on the thickness of loess (TOL) deposits over a calcareous bedrock. We attempted to test the TOL prediction by coupling a large soil dataset (10978 sampling sites) and 117 environmental covariates. After variable selection by recursive feature elimination, quantile regression forests (QRF) was employed for spatial modelling, as it was able to directly provide the 90% prediction intervals (PIs). Averaging a total of 50 models, generated by repeated stratified random sampling, showed a substantial model performance with mean R2 of 0.33, RMSE of 30.48 cm and bias of −1.20 cm. The prediction interval coverage percentage showed that 86.70% of the validation samples fall within the predefined 90% PIs, which also indicated the prediction uncertainty produced by QRF was reasonable. The relative variable importance indicated the importance of airborne gamma-ray radiometric data and Sentinel 2 products in TOL prediction. The produced TOL map with 90% PIs makes sense from a soil science and physiographic point of view. The final product can guide evidence-based decision making for agricultural land management, especially for irrigation in our case study.