Spatial prediction of soil aggregate stability and soil organic carbon in aggregate fractions using machine learning algorithms and environmental variables

Knowledge about the spatial variability of soil aggregate stability indices, soil organic carbon (SOC) in various aggregate sizes, and aggregation across the landscape is crucial for sustainable land use planning and management practices. Direct traditional measurements for the target variables, as mentioned above, are timeconsuming and expensive. Thus, this study attempts to spatially predict the soil aggregate stability indices, including mean weight diameter-MWD, geometric mean diameter-GMD, water-stable aggregates-WSA, and SOC in various aggregate fractions using digital soil mapping and machine learning models using the environmental covariates as the time and cost-effective approaches. Thus, a total of 100 soil surface samples (0-10 cm depth) were collected from the natural forest, tea plantation, and paddy rice field land uses, and soil aggregate stability indices were determined following laboratory analyses. The machine learning models, including random forest (RF), k-nearest neighbors (kNN), support vector machine (SVM), artificial neural network (ANN), and the ensemble of four single models, were trained using the repeated 10-fold cross-validation method. The models were evaluated by the root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), and normalized RMSE (nRMSE). The modeling results demonstrated that the RF model outperformed for MWD (R2 = 0.74, nRMSE = 24.28), GMD (R2 = 0.75, nRMSE = 12.72), and WSA (R2 = 0.58, nRMSE = 10.40), while kNN and SVM models resulted in the best prediction of SOC in (meso and micro-aggregates (RMSE = 1.03 and 0.88)) and macroaggregates (RMSE = 1.49), respectively. However, the ensemble model increased the prediction accuracies for all soil targets (RI >= 15.78%). Moreover, the variable importance analysis showed that soil properties such as soil organic matter (SOM) and remote sense-data mainly explained the variation of soil aggregate stability indices and SOC in various aggregate fractions. Overall, the results revealed that the machine learning-based models could accurately predict the soil aggregate stability and associated SOC, and the produced maps can be a baseline map for land use planning and decision making.

Automated summary

This study spatially predicted soil aggregate stability indices and SOC in various aggregate sizes across the landscape using digital soil mapping and machine learning models. The random forest model performed best for MWD, GMD, and WSA, while kNN and SVM models showed the best prediction for SOC in different aggregate fractions. The ensemble model increased prediction accuracies for all soil targets, highlighting the importance of machine learning-based models for land use planning and decision making.