Comparison of machine learning for predicting and mapping soil organic carbon in cultivated land in a subtropical complex geomorphic region

Soil organic carbon (SOC) is a key indicator of soil quality and ecosystem health. At present, machine learning (ML) models for predicting soil properties based on environmental variables are increasingly popular; however, the performance of different ML algorithms in predicting and mapping SOC, especially at high spatial resolutions, have not been compared. This study aimed to develop, evaluate, and compare the performance of Support Vector Machine (SVM), Random Forest (RF), and Ordinary Kriging (OK) models for predicting and mapping the SOC contents in the northeast of Fujian Province. Remote sensing vegetation indices were derived from Sentinel-2 image data with a spatial resolution of 10 m. These vegetation indices, along with selected terrain and climate factors, were adopted as environmental variables to map SOC using the SVM and RF models. The results showed that the performance of the RF model (RMSEroot-mean-square error=2.004, r=0.897) was better than that of the OK model (RMSE=4.571, r=0.623) and explained most of the SOC spatial heterogeneity. The SVM model had the poorest prediction accuracy (RMSE=5.190, r=0.431). SOC mapped from the three models had similar spatial patterns, with an increasing SOC gradient from east to west and from south to north of the study area. SOC in the farmlands predicted with the RF model varied in the range of 15.33±4.07 g·kg^-1. Elevation and rainfall were the most important variables for the RF and SVM models, respectively, whereas the remote sensing vegetation indices were less important than elevation.