Bauer (2010) describes similar findings. However, he refers to values between 1.0 ± 0.1 and 1.4 ± 0.5 g cm−3 for the topsoil of Stagnosols and values of only 0.9 ± 0.1 to 1.2 ± 0.2 g cm−3 for the topsoil horizons of landslides. We assume that Ah horizons within the research area rather developed on accumulated landslide material, whereas soils developing on in-situ material include a stagnic soil layer, which is usually found at the soil surface. We further assume that the complete research area would be covered by stagnic soils without the influence of landslides that lead to lower soil bulk densities in their accumulation zones.
Horizon-wise model development to predict bulk density from terrain parameters was not possible. We therefore predicted the bulk density within the first soil horizon (H1) regardless of its characteristic, expecting topsoil bulk density to be much more dependent on surface morphology. On the other hand, we also had to consider subsoil bulk density. From Fig. 5 it can be observed that Bw mean bulk density is the same as in H1 and a bit higher than that of Ah. However, Bw median is the lowest of all horizons. Correspondingly, we used the predicted H1 bulk density also for the subsoil.
Fig. 6 shows the histograms of the 100 Pearson's rxy from external cross validation to compare RT and RF models to predict bulk density. RF models where again constructed with the optimal mtry size = 2 and 500 trees. According to the rxy distribution mean, RFnn was the best model with mean rxy = 0.28 and maximum rxy = 0.7. Reasons for the poor performance of some parts of the dataset are similar to those discussed for the depth of the failure plane models. In addition to the low rxy values, we also encountered a high rxy standard deviation (0.2) which is probably due to the small dataset. Bulk density was only measured in 56 soil profiles. Using a smaller test dataset, e.g. 5%, to leave the major part of the data for model development, might improve rxy. However, splitting of the dataset was only done to compare model performance. Spatial prediction was based on the complete dataset.