Predicting soil depth in a large and complex area using machine learning and environmental correlations

Soil depth is critical for eco-hydrological modeling, carbon storage calculation and land evaluation. However, its spatial variation is poorly understood and rarely mapped. With a limited number of sparse samples, how to predict soil depth in a large area of complex landscapes is still an issue. This study constructed an ensemble machine learning model, i.e., quantile regression forest, to quantify the relationship between soil depth and environmental conditions. The model was then combined with a rich set of environmental covariates to predict spatial variation of soil depth and straightforwardly estimate the associated predictive uncertainty in the 140 000 km² Heihe River basin of northwestern China. A total of 275 soil depth observation points and 26 covariates were used. The results showed a model predictive accuracy with coefficient of determination (R²) of 0.587 and root mean square error (RMSE) of 2.98 cm (square root scale), i.e., almost 60% of soil depth variation explained. The resulting soil depth map clearly exhibited regional patterns as well as local details. Relatively deep soils occurred in low lying landscape positions such as valley bottoms and plains while shallow soils occurred in high and steep landscape positions such as hillslopes, ridges and terraces. The oases had much deeper soils than outside semi-desert areas, the middle of an alluvial plain had deeper soils than its margins, and the middle of a lacustrine plain had shallower soils than its margins. Large predictive uncertainty mainly occurred in areas with a lack of soil survey points. Both pedogenic and geomorphic processes contributed to the shaping of soil depth pattern of this basin but the latter was dominant. This findings may be applicable to other similar basins in cold and arid regions around the world.