Uncertainty in the spatial prediction of soil texture: Comparison of regression tree and Random Forest models

Within the southern Ecuadorian Andes, landslides have an impact on landscape development. Landslide risk estimation as well as hydrological modelling requires physical soil data. Statistical models were adapted to predict the spatial distribution of soil texture from terrain parameters. For this purpose, 56 soil profiles were analysed horizon-wise by pipette and laser method. Results by pipette compared to laser method showed the expected shift to higher silt and lower clay contents. Linear regression equations were adapted. The performance of regression tree (RT) and Random Forest (RF) models was compared by hundredfold model runs on random Jackknife partitions. Digital soil maps of sand, silt and clay percentage mean and standard deviation indicate model variability and prediction uncertainty.RF models performed better than RT models. All terrain factors considered in the analysis influenced soil texture of the surface horizon, but altitude a.s.l. was assigned the highest variable importance during model construction. Shallow subsurface flow is considered responsible for increasing sand/clay ratios with increasing altitude, on steep slopes and with overland flow distance to the channel network by removing clay particles downslope. Deeper soil layers are not influenced by this process and therefore, did not show the same texture properties. However, the influence of parent material and landslides on the spatial distribution of soil texture cannot be neglected. Model performance, most probably, could be improved by a bigger dataset.

► Soil texture is predicted on a landscape scale.
► Regression Trees and Random Forest models are compared in their performance.
► The digital soil maps include model variability and prediction uncertainty.
► Model dependence on the dataset is addressed by model runs with various data subsets.
► Surface processes are distinguished from the influence of parent material.