Spatial prediction using kriging ensemble

It is surprising how a commonly used concept in temporal prediction—combining forecasts, or rather combining predictions—has not really been brought forward in spatial prediction. Analogous to forecasting, where forecasts made using models such as exponential smoothing or neural networks are combined through regressions, the various prediction combination methods are herein transferred to spatial prediction problems. Through a series of empirical studies, the advantage and potential of kriging ensemble, or more generally, spatial-interpolator ensemble, are demonstrated. Both geostatistical and lattice data (solar irradiance) are considered. Although in theory, the improvement in predictive performance is not guaranteed, just like how we cannot guarantee that ensemble improves forecasts, in practice, a validated ensemble performs at least as good as the best component model, just like how the ensembles in forecasting would behave.