Generating soil thickness maps by means of geomorphological-empirical approach and random forest algorithm in Wanzhou County, Three Gorges Reservoir

[Display omitted] •A state-of-the-art geomorphological model is customized to get distributed soil thickness map.•A hybrid version (GIST-RF) is developed combined with a random forest algorithm.•The two models are compared in a complex setting with very thick soils.•GIST-RF model showed promising results. Soil thickness, intended as depth to bedrock, is a key input parameter for many environmental models. Nevertheless, it is often difficult to obtain a reliable spatially exhaustive soil thickness map in wide-area applications, and existing prediction models have been extensively applied only to test sites with shallow soil depths. This study addresses this limitation by showing the results of an application to a section of Wanzhou County (Three Gorges Reservoir Area, China), where soil thickness varies from 0 to ∼40 m. Two different approaches were used to derive soil thickness maps: a modified version of the geomorphologically indexed soil thickness (GIST) model, purposely customized to better account for the peculiar setting of the test site, and a regression performed with a machine learning algorithm, i.e., the random forest, combined with the geomorphological parameters of GIST (GIST-RF). Additionally, the errors of the two models were quantified, and validation with geophysical data was carried out. The results showed that the GIST model could not fully contend with the high spatial variability of soil thickness in the study area: the mean absolute error was 10.68 m with the root-mean-square error (RMSE) of 12.61 m, and the frequency distribution residuals showed a tendency toward underestimation. In contrast, GIST-RF returned a better performance with the mean absolute error of 3.52 m and RMSE of 4.56 m. The derived soil thickness map could be considered a critical fundamental input parameter for further analyses.