Comparison of Machine Learning and Geostatistical Methods on Mapping Soil Organic Carbon Density in Regional Croplands and Visualizing Its Location-Specific Dominators via Interpretable Model

Abstract

High-precision soil organic carbon density (SOCD) map is significant for understanding ecosystem carbon cycles and estimating soil organic carbon storage. However, the current mapping methods are difficult to balance accuracy and interpretability, which brings great challenges to the mapping of SOCD. In the present research, a total of 6223 soil samples were collected, along with data pertaining to 30 environmental covariates, from agricultural land located in the Poyang Lake Plain of Jiangxi Province, southern China. Furthermore, ordinary kriging (OK), geographically weighted regression (GWR), random forest (RF), and empirical Bayesian kriging (EBK), along with three hybrid models (RF-OK, RF-EBK, RF-GWR), were constructed. These models were used to map the SOCD (soil organic carbon density) in the study region with a high resolution of 30 m. After that, shapley additive explanations (SHAP) were used to quantify the global contribution and spatially identify the dominant factors that influence SOCD variation. The study outcomes suggested that compared to the single geostatistics model and hybrid model, the RF method emerged as the most effective predictive model, showcasing superior performance (coefficient of determination (R²) = 0.44, root mean squared error (RMSE) = 0.61 kg m⁻², Lin's concordance coefficient (LCCC) = 0.58). Using the SHAP, we found that soil properties contributed the most to the prediction of global SOCD (81.67%). At the pixel level, total nitrogen dominated 50.33% of the farmland, followed by parent material (8.11%), available silicon (8.00%), and mean annual precipitation (5.71%), and the remaining variables accounted for less than 5.50%. In summary, our study offered valuable enlightenment toward achieving a balance between accuracy and interpretability of digital soil mapping, and deepened our understanding of the spatial variation of farmland SOCD.