Estimation and spatio-temporal analysis of soil organic carbon stock in China using machine learning algorithms from 2000 to 2014

Soil is the largest carbon pool in terrestrial ecosystems, storing up to 2 or 3 times the amount of carbon present in the atmosphere, and a small change in soil carbon stock could have profound effects on atmospheric CO₂ and climate change. However, an accurate estimate of soil organic carbon (SOC) stock is still challenging. Previous studies on SOC stock prediction across China were mainly from biogeochemical models and national soil inventories, and large uncertainties still remained. In this study, we predicted SOC stock at 0 – 20 cm and 0 – 100 cm with 3419 and 2479field observations using artificial neural network (ANN), extreme gradient boosting (XGBoost), random forest (RF), and gradient boosting regression trees (GBRT) across China with the linkage of climate, vegetation and soil variables. Results showed that RF performed best among the four machine learning approaches with model efficiency of 0.61 for 0 – 20 cm and 0.52 for 0 – 100 cm. The trained RF model was used to predicted the temporal and spatial patterns of SOC stock at a spatial resolution of 1 km from 2000 to 2014 across China. Temporally, SOC stock at 0 – 20 cm (p = 0.07) and 0 – 100 cm (p = 0.3) did not change significantly. However, SOC density showed strong spatial patterns, the mean value of SOC density at 0-20 cm and 0-100 cm increased firstly, then decreased and then increased with the increase of latitude, and the minimum density was 39.83° and 41.59°, respectively. The total SOC stocks across China were 33.68 and 95.01 Pg C for 0 – 20 cm and 0 – 100 cm, respectively. The developed SOC stock could serve as an independent dataset that could be used for decision-making and help with baseline assessments for inventory and monitoring SOC stocks for global biogeochemical models in China.