Terrain attributes probably only explain horizon thickness and particularly soil depth to a limited extent. The vertical development of the soil profile is less influenced by surface processes. Bauer (2010) limited downslope subsurface flow within the research area to the topsoil. Although our earlier analysis related the occurrence of stagnic properties to terrain parameters, the extent of the stagnic horizon could not be explained. Park and Vlek (2002) reported that soil attributes whose vertical distribution is strongly determined by vertical pedogenesis or unknown factors are poorly modelled by environmental variables.
Another aspect to be considered is the spatial map resolution. The soils within the research area change within a few metres radius as typical for tropical soils. Accordingly, we used the highest possible DEM resolution, and small scale soil variability was included within the models, which would be neglected while working on a larger scale. We conclude that the size of the dataset we applied was not enough to represent the investigated soil-landscape at this high precision. However, due to the lack of any better option, we still applied the overall best model to predict the depth of the failure plane and discuss the influence of terrain parameters, keeping in mind that they can only explain it partially.
The overall best model was used to predict soil depth (best model: RFnn), Ah (best model: RFm) and stagnic (best model: RFm) horizon thickness. Fig. 4a displays the terrain variable importance in constructing the RFnn model to predict soil depth, and Fig. 4b shows it regarding the RFm models to predict Ah and stagnic horizon thickness. The variable importance measure indicates how much the mean square error (MSE) would increase if the respective predictor would be suspended from the model ( Prasad et al., 2006).