Pedoclimatic zone-based three-dimensional soil organic carbon mapping in China

Up-to-date maps of soil organic carbon (SOC) concentrations can provide vital information for monitoring global or regional soil C changes and soil quality. In this study, a national soil dataset collected in the 2010 s was applied to produce SOC maps of mainland China at soil depths of 0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm, 60–100 cm and 100–200 cm. A stacking ensemble learning framework was utilized to take advantage of the optimal predictions from individual models. A voting-based ensemble learning model (VELM) was proposed with consideration of pedoclimatic zones. In this model, three machine learning models were separately trained for every pedoclimatic zone, and their predictions were selectively merged together. A weighted ensemble learning model (WELM), in which the parameterization considered all zones (i.e., the whole study area) simultaneously, was also trained for comparison. The overall R2 values of these two methods ranged from 0.16 to 0.57 and decreased with depth. Based on the independent validation, the R2 values ranged from 0.41 to 0.57 in the topsoil (0–5 cm, 5–15 cm and 15–30 cm). Overall accuracy metrics implied that the VELM and WELM yielded nearly the same prediction performances. However, model validation in the pedoclimatic zones showed that the VELM obviously outperformed the WELM, with the VELM generally improving the accuracy by 12.6%. Based on the independent validation, we also compared our predictions with other soil map products. Although the spatial patterns were similar, the predicted SOC maps outperformed two other products. The comparison of the two ensemble models should serve as a reminder that if new national or regional soil maps are generated, validation based on pedoclimatic zones or other soil-landscape units may be necessary before applying these maps.