Spatial distribution prediction of soil carbon content and estimation of carbon stock based on heterogeneous ensemble learning model

Objective Accurate estimation of soil carbon stocks and sequestration potential is conducive to the development of precision agriculture and ensuring food security. It also provides references for natural resource management and utilization.

Methods In this study, 512 soil samples were collected from Wuchang and Shuangcheng districts of Harbin City, Heilongjiang Province. The spatial distribution of soil carbon content was predicted using six types of environmental variables and 15 factors through the Random Forest (RF) model, AdaBoost model, Support Vector Regression (SVR) model, and Stacking model. Moreover, the carbon stock and sequestration potential were estimated.

Results The results indicate that the soil carbon content is relatively lower in cultivated land located in the northwest plains and higher in forest areas in the southeast hills. The Stacking model shows the best prediction performance (R² = 0.49), suggesting that heterogeneous integrated learning models are more effective than single learning models in predicting soil carbon content. The carbon stock in the surface layer (0~30 cm) is 260.99 Mt. Dark brown soil has the highest carbon stock, accounting for 35.32% of the total study area, followed by black soil and meadow soil.

Conclusions The surface layer (0 ~30 cm) has strong carbon sink capacity with a carbon sequestration potential of 247.63 Mt. Black soil, dark brown soil, and meadow soil also possess relatively strong carbon sink capacity. The carbon sequestration potential of cultivated land is higher than that of forest in the hills. Therefore, it is necessary to further develop a reasonable cultivation method to increase the carbon sequestration capacity of cultivated land. This study is expected to provide a reference for the accurate assessment of carbon resources and land use management.