Estimation of soil thickness in karst landforms using a quantile regression forests approach

Abstract

Soil thickness is a basic feature of the earth’s land surface. Accurately representing the spatial distribution of soil thickness is important for various models of earth surface processes. However, mapping soil thickness in karst landforms is highly uncertain. To address this challenge, this study analyzed the correlation between 906 soil thickness measurements and 376 environmental characteristics in a typical karst landscape covering 54,000 km². We employed a quantile regression forests (QRF) approach to estimate soil thickness and evaluate the associated uncertainty in the predicted results. We found that, like other regional scale soil mapping models, climate and topographic data were key factors influencing soil thickness. Specific for karst landscapes, we found that the characteristics of karst rocky desertification play a key role in predicting soil thickness. The rocky desertification information indexes (RIs), which use exposure rate of bedrock to represent the degree of karst rocky desertification, showed relatively high importance in the variable importance assessment. The developed model explained 40 % of the spatial variability of soil thickness across the study area, with an RMSE (37.3 cm) of 50–60 % of the mean thickness. This indicates that the model, and environmental factors evaluated within, explained a little less than half of the spatial variability. The prediction results reveal the distribution pattern of soil thickness at both regional and local scales within karst landforms. Thick soil was commonly found in low-lying landscape features like depressions and foothills, whereas areas with steep slopes, ridges, and peaks tended to have thin soil, following a typical toposequence. In areas with relatively deep soil or severe rocky desertification, the uncertainty of predicting soil thickness is relatively high. The results of uncertainty analysis, as a supplement to the prediction results, have improved the usability of the predictions to a certain extent. This study has, to some extent, addressed the challenges of predicting soil thickness in karst areas and has also provided transferable methods for other complex regions.