Can the spatial prediction of soil organic matter contents at various sampling scales be improved by using regression kriging with auxiliary information?

The data quality of soil properties, such as the soil organic matter (SOM) content, can be improved and the spatial sampling intensities may be reduced by incorporating secondary information, such as those derived from topographic (TOPO) and remote sensing (RS) data to enhance their spatial estimates. This study adopted a generic framework for spatial interpolation using regression kriging (RK) developed by Hengl et al. (2004) to evaluate RK's capability in improving SOM spatial interpolation using internal secondary variables (sampling coordinates) and external auxiliary information, such as soil map (SOIL), vegetation indices (VIs) derived from a Landsat 5 TM image, and several terrain attributes (elevation, slope, convergence and wetness indices, and plan and profile curvatures). Meanwhile, the SOM spatial distribution was also interpolated by using ordinary kriging (OK) and universal kriging (UK) methods for comparison purposes. The results of this study showed that the prediction accuracy of SOM by using RK was unimproved with the inclusion of more auxiliary information in the regression models, but in contrast it significantly declined when TOPO, VI and SOIL information were combined, particularly the last one. It was also observed that with the increase of the minimum sampling distances from 25 to 500 m or with the decrease of the sampling densities from 0.42 to 0.26 # km−2, the RK techniques did not outperform OK and UK in improving the SOM prediction accuracy at coarse sampling resolutions. Interestingly, the highest accuracy of the SOM prediction by all these interpolation methods was achieved at the minimum sampling distance of 250 m. The suitability of RK implementation in the spatial interpolation was therefore discussed by considering the minimum sampling distance, the sampling density and the compatibility of spatial resolutions of target variables and auxiliary information or the spatial scales.

Research Highlights
► RK did not outperform OK and UK in improving the SOM prediction accuracy.
► Best SOM predictions by all methods were at the minimum sampling distance of 250 m.
► Data quality was regarded more important than prediction methods.