More Reliable Baseline Estimates of Soil Organic Carbon Content Using Reinforced Density of Legacy Soil Profiles at a Provincial Scale

Revealing the temporal dynamics of soil organic carbon (SOC) requires a high-precision and reliability baseline estimate of SOC, but the accuracy of the SOC estimates is largely dominated by the model-calibrating approaches and the sample size used to training the models. In this study, we trained the random forest (RF) models for SOC content prediction in two ways. Firstly, based on a dataset containing 3348 legacy soil profile observations and 34 of environmental covariates, a RF model was calibrated for the whole study region (so called RM approaches) to predict the SOC content in the 5 standard depth intervals (0–5, 5–15, 15–30, 30–60 and 60–100 cm) . Secondly, the study region was divided into 6 sub-regions by the differences in the climate/terrain/soil combinations, then the 6 RF models were separately calibrated at the sub-regional level. Finally the overall prediction for the whole study region was generated by combining the outputs of the 6 RF models locally calibrated for sub-regions (so-called LM approach). The results showed that, using the same training sample size, the models calibrated respectively by RM and LM approaches had similar accuracy and comparable performance in predicting SOC content at every depth intervals of the soil profiles in the study region. The achieved R 2 indicated that RM and LM calibrated models could explain 46% and 48% of the SOC content variation at the 0-5cm soil layer, and 59% and 57% of the SOC content variation at the 5-15cm soil layer of soil profiles, respectively. Comparatively, the spatial prediction capacity of the models calibrated by LM approach in the plain area was significantly better than that by RM approach. Besides, the uncertainty analysis revealed that the spatial distribution of SOC content predicted by the LM approach was more reliable than that predicted by the RM approach, except in the 30-100cm depth of the western mountainous area. Compared with the national and global scale SOC estimates and soil map-based data, the prediction results obtained in this study had the highest accuracy and the lowest uncertainty. The R 2 for depth-weighted prediction of SOC content in the 0-30 cm soil layer by the LM approach was as high as 0.67