Using a process model and regression kriging to improve predictions of nitrous oxide emissions from soil

In this paper we use regression kriging to improve predictions of a hard-to-measure soil variable based on an established process model. In our case study the target variable is the rate of nitrous oxide (N2O) emission from soil. We used three different process models that each have the same soil properties as input variables, but make different assumptions about the kinetics of denitrification with respect to the concentration of nitrate. The data were divided into a prediction and a validation subset. Cross validation at prediction sites by regression kriging was used to determine which of the three process models was the most plausible. The regression model was fitted using Residual Maximum Likelihood (REML), which also estimates the variance model for kriging. The selected REML fitted model was then used to predict the N2O emission rates at the validation sites and the results compared with those achieved with the fitted model alone and estimates made by disjunctive kriging from the measurements at prediction sites. The regression kriging predictions were substantially better than those from the model alone or disjunctive kriging.