Spectral tempering to model non-stationary covariance of nitrous oxide emissions from soil using continuous or categorical explanatory variables at a landscape scale

The rate of nitrous oxide emissions was measured from 276 soil cores on a 7.5-km transect, and then a subset of these data was used to compute geostatistical models in which land categories (land-use and soil type) were fixed effects. In one model the random effects were assumed to be second-order stationary. In the other models non-stationary random variation was modelled independently for the autocorrelation and variance of the spatially correlated component of emission rate, and for the nugget variance. This was done with the method of spectral tempering. Non-stationary variance parameters were modelled as functions of discrete or continuous auxiliary variables. Models in which spectral tempering was applied using quadratic functions of soil pH fitted the data significantly better than a stationary model and gave better estimates of the prediction error variances. A significantly better fit was also obtained using splines on location to model non-stationarity, but mapped soil associations did not provide a basis for a significantly better variance model. Computational difficulties with spectral tempering are identified and strategies to overcome them are discussed.

Research Highlights
► Non-stationary variance models fit landscape-scale emission data best.
► Soil pH was the best ancilliary variable to measure this non-stationarity.
► There are further computational challenges for spectral tempering models.