Abstract: To get a suitable method for spatial prediction of soil microbial diversity, multiple linear regression (MLR), ordinary kriging (OK), regression kriging (RK), and empirical Bayesian kriging regression (EBKR) methods were compared. A total of 35 soil samples were collected from grasslands in Chifeng, central Inner Mongolia. Soil bacterial diversity, measured by Faith’s phylogenetic diversity (PD) index, was examined by high-throughput sequencing. Auxiliary environmental data related to climate, vegetation and elevation were acquired from relevant websites as the inputs of the MLR, RK, and EBKR predictions. Leaveoneout crossvalidation, mean error (ME), and root mean square error (RMSE) were used to evaluate the prediction accuracy. Results showed that aridity index, mean annual precipitation, and net photosynthesis were the best combination of environmental variables explaining the variation of soil bacterial diversity. The spatial distribution patterns of the diversity predicted by the MLR, OK, RK, and EBKR methods were quite similar. The diversity in the southeast part of the study area was higher, whereas that in the northwest part of the study area was relatively lower. The three methods involving regression analyses could better describe the local variation of the diversity than the OK method. Furthermore, for the predictions of MLR, OK, RK, and EBKR, the coefficients of determination of leaveoneout crossvalidations were 0.408, 0.439, 0.476, and 0.638, the values of ME were -0.065, 0.033, 0.017, and -0.009, and the values of RMSE were 5.23, 5.04, 4.95, and 4.05, respectively. These results suggested that the OK method had a slightly higher prediction accuracy than the MLR method, while the RK and EBKR methods, which combined the auxiliary environmental variables, further improved the prediction accuracy. Specifically, the EBKR method better predicted the spatial distribution of the diversity because it could overcome the limitation of the RK method, which usually generalizes the spatial structure of all location data with a single semivariogram function. In conclusion, the geostatistical methods combining auxiliary environmental variables and considering local differences in spatial variability have great potential in predicting the spatial distribution of soil microbial diversity.

Key words: soil bacterial diversity, multiple linear regression, ordinary kriging, regression kriging, empirical Bayesian kriging regression prediction.