| Article ID | Journal | Published Year | Pages | File Type |
|---|---|---|---|---|
| 4573983 | Geoderma | 2012 | 8 Pages |
Regression-kriging (RK) is becoming an important tool in geostatistics because of the availability of many covariates at high spatial resolution with the advancement in proximal and remote-sensing with positioning technologies. This paper presents the application of a new local RK algorithm for prediction of soil properties. The algorithm was tested using 985 observations for prediction of soil pH, clay content and carbon content in the lower Hunter Valley of New South Wales in Australia. Environmental covariates for the area were compiled. First, the covariates used in regression analysis for each of the soil properties were obtained through a step-wise regression analysis. Secondly, different spatial prediction methods were examined. Finally, the prediction efficiency of various techniques was tested at validation sites using the standardised squared deviation as a measure of the goodness of theoretical estimates. The results from validation showed that the local RK method does not always present the best predictions, but for specific cases it may be highly accurate. We conclude that local RK performance depends on the actual soil and environmental factor relationships, and in general performs no worse than global RK. Furthermore, the advantage of local RK is that it can provide an approach to understand how much the regression models and variogram models change across a region. We have developed a software programme named RKGuider to carry out the local RK steps automatically. Further investigation and numerous datasets are required to verify the algorithm.
Research highlights► A new local regression-kriging algorithm was developed for prediction of soil properties. ► Local regression-kriging method does not always present the best prediction, the performance depends on the actual soil and environmental factor relationships. ► Local regression-kriging can provide an overview of how much the regression relationships and variogram parameters vary across a region.
