Evaluation of Machine Learning Methods in Digital Mapping of Soil Organic Carbon (part of Khorramabad Plain)

Knowledge about the spatial distribution of soil organic carbon (SOC) is one of the practical tools in determining sustainable land management strategies. During the last two decades, the utilization of data mining approaches in spatial modeling of SOC using machine learning algorithms have been widely taken into consideration. The essential step in applying these methods is to determine the environmental predictors of SOC optimally. This research was carried out for modeling and digital mapping of surface SOC aided by soil properties ie., silt, clay, sand, calcium carbonate equivalent percentage, mean weight diameter (MWD) of aggregate, and pH by machine learning methods. In order to evaluate the accuracy of random forest (RF), cubist, partial least squares regression, multivariate linear regression, and ordinary kriging models for predicting surface SOC in 141 selected samples from 0-30 cm in 680 hectares of agricultural land in Khorramabad plain. The sensitivity analysis showed that silt (%), calcium carbonate equivalent, and MWD are the most important driving factors on spatial variability of SOC, respectively. Also, the comparison of different SOC prediction models, demonstrated that the RF model with a coefficient of determination (R2) and root mean square error (RMSE) of 0.75 and 0.25%, respectively, had the best performance rather than other models in the study area. Generally, nonlinear models rather than linear ones showed higher accuracy in modeling the spatial variability of SOC.