High‐Resolution Prediction of Spatial Distribution of Soil Thickness in Areas With Heterogeneous Soil Parent Materials: Machine Learning Methods and Variable Optimisation

Abstract

Soil thickness is an important indicator for evaluating soil production and ecosystem support functions. However, traditional field surveys cannot characterise the spatial distribution of soil thickness with high resolution. Although machine learning models that integrate multiple environmental variables offer a promising approach to predict the high‐resolution spatial distribution of soil thickness, their predictive accuracy is often constrained in areas with heterogeneous soil parent materials. In this study, the Zhangbei area with heterogeneous soil parent materials was selected as a case to optimise the model. In detail, soil parent materials were used as one of environmental variables incorporated into the machine‐learning model for soil thickness prediction, and their contributions were evaluated. Based on 97 soil thicknesses of surveyed profiles in the Zhangbei area, four machine learning models were developed and compared, including support vector machine (SVM), artificial neural network (ANN), random forest (RF) and extreme gradient boosting (XGB). Results showed that all models performed better than the simple linear baseline model, indicating the advantage of nonlinear modelling for soil thickness prediction. Especially, the XGB model incorporating soil parent material explained 81% of the spatial variation in soil thickness and outperformed models of SVM, ANN and RF. Based on high‐resolution data of soil parent materials, the soil thickness with high‐resolution spatial distribution across the study area was precisely predicted. Overall, our study provides a high‐precision framework for predicting soil thickness in areas with heterogeneous soil parent materials, and its potential applications are expected for ecological restoration and land management.