Objective The study was to make full use of the multi-source data with uncertainty or prior distribution in improving spatial prediction accuracy of soil organic carbon density(SOCD).

Method The typical agricultural landscape unit of subtropical red soil hills were selected as the research area, and the environmental factors as auxiliary variables. Three methods were used to calculate and compare the results of spatial prediction for SOCD, including geographically weighted regression model(GWR), Bayesian maximum entropy combined with geographically weighted regression model(BME-GWR), and Bayesian maximum entropy combined with geographically weighted regression model estimated by land use type(BME-GWRL).

Results BME-GWR and BME-GWRL model had stronger ability to explain the spatial heterogeneity of SOCD. The leave-one-out cross validation results of determination coefficient(R2)of BME-GWR and BME-GWRL models were 0.81 and 0.79, the root mean square errors(RMSE)of BME-GWR and BME-GWRL models were 0.35 and 0.33, and the mean absolute fitting errors(MAE)of BME-GWR and BME-GWRL models were 0.19 and 0.21, respectively. These two methods had a higher fitting accuracy than GWR model, which could better reflect the spatial local characteristics of SOCD with auxiliary variables as soft data. In particular, BME-GWRL model used the soft data extracted from the prediction of SOCD under various land use types, and the prediction result was more accurate than that of BME-GWR model, of which soft data was directly simulated in the whole study area without considering land use types.

Conclusion BME-GWRL considers the uncertainty of the estimation unit of soft data, which can provide an effective method for improving the accuracy of spatial prediction with rational utilization of multi-source auxiliary data.