Keywords: spatial, fixed rank kriging, reduced rank spatial model, knot selection
Modern advances have enabled the collection of massive spatial and spatio-temporal datasets (upwards of tens of thousands of observations). Reduced rank models are typically required for computational feasibility when analyzing such datasets. Multiple strategies exist for specifying reduced rank spatial models (RRSMs) for both spatial and spatio-temporal models. However, a common feature is to define a reduced-dimension latent process over a selected number of knot locations across the spatial domain. Most work on RRSMs focuses on estimation of parameters or specification of appropriate prior distributions. Thus far the selection of the knot locations has received relatively little attention. The methods which do address knot selection often rely on computationally intensive Bayesian methods. In this paper we propose an efficient, non-Bayesian, data-driven approach to knot selection for reduced rank spatial models. We illustrate our method through simulations and demonstrate it on data obtained from the Deepwater Horizon disaster.