Spatial Prediction and Uncertainty Assessment of Soil Organic Carbon in Hebei Province, China

Abstract

To quantify the spatial distribution of SOC in Hebei Province five models were compared: multiple linear regression (MLR), universal kriging (UK), regression-kriging (RK), artificial neural network combined with kriging (ANN-kriging), and regression tree (RT). The modelling was supported by 359 SOC density (total SOC by volume, SOCD) data points, as well as relief parameters derived from a 100m × 100m resolution DEM, and NDVI calculated from NOAA AVHRR data to map SOCD (to a depth of 1m) spatial distributions. Only 19.5% of the total SOCD variation can be explained by MLR method, the UK method resulted in a wider range of SOCD compared with MLR method. The UK method and RK method explain 53 and 65% of the total variation, respectively, and the local variation of lower SOCD in the southeast of the province was detected. The ANN-kriging and RT mapping both explained 67% of the total variation. Compared to ANN-kriging, the RT method has lower root mean square prediction error. The sequential indicator simulation (SIS) was applied for assessing topsoil SOCD (0–20 cm) uncertainty at unsampled locations. The conditional variance of 1,000 realizations generated by SIS was greater in mountainous areas where SOCD fluctuated the most, and the uncertainty was less on the plain area where SOCD was consistently low. The RT model is of best performance for mapping the spatial distribution of SOCD, and the SIS technique can quantitatively assess the local and spatial uncertainty of SOCD being greater than a given threshold.