Modelling soil thickness using environmental attributes in karst watersheds

Soil thickness exerts significant influence on numerous earth surface processes and it can be estimated using various methods. However, it is unclear which environmental variables can best predict soil thickness, especially in the karst region where soil distribution is highly heterogeneous. In this study, three variable selection methods (partial least-squares regression (PLSR), random forest method (RF), and generalized linear model (GLM)) were used to identify the most important variables to explain variation in soil thickness among 19 quantitative and 4 qualitative environmental variables. Random forest (RF), artificial neural networks (ANNs), and support vector regression (SVR) models were then used to model soil thickness based on the three variable selection methods, and model performance was compared. Results showed that the models (RF, ANNs and SVR) based on the selected variables could explain similar or slightly higher (0.63 < R2 < 0.73) variation in soil thickness compared with the selection of all variables (0.63 < R2 < 0.71). This demonstrated that the variable selection methods are effective ways to select the most representative variables to predict soil thickness, and that they resulted in comparable prediction performance when explaining the variation of soil thickness. Results also indicated that the SVR had slightly higher prediction accuracy than RF and ANNs among the three selection methods. In general, the proposed variable selection methods coupled with a suitable model may be a feasible and effective framework to estimate soil thickness in karst regions where soil is highly heterogeneous.