Abstract
Soil organic carbon (SOC) is the largest carbon reservoir in the terrestrial ecosystem and plays an important role in the global carbon cycle. Consequently, even a slight change in SOC content due to land use, soil management, or rates of soil erosion can considerably increase atmospheric CO2 concentrations. The main purpose of this study is to predict and map SOC content in small area by applying machine learning (ML) algorithms using field measurements and remote sensing data. We used to three different algorithms such as Random Forest (RF), Extreme Gradient Boosting (eXGB), and Gradient Boosted Regression (GBR) of ML. According to field work, 123 soil samples (0–30 cm) were collected from Bayanzurkh soum in Khuvsgul, and 26 variables were used to predict SOC content. As shown the prediction results, the GBR algorithm demonstrated the highest performance, explaining 78% of the variation in soil SOC content, with an RMSE of 42.9 g/kg and an MAE of 33.1 g/kg. The ranking of model performance in terms of prediction accuracy was GBR > eXGB > RF. Therefore, we found a strong relationship (R² = 0.94) between the predicted and measured values based on linear regression analysis. The most influential predictor variables were SILT (13.6%), CLAY (7.8%), NDVI (7.3%), and SOLAR RADIATION (6.3%). These results demonstrate that SOC content can be effectively predicted using machine learning algorithms. However, it is advisable to compare the performance of multiple algorithms and select the most suitable approach for the small area.