Computationally simple anisotropic lattice covariograms

Abstract

When working with contemporary spatial ecological datasets, statistical modellers are often confronted by two major challenges: (I) the need for covariance models with the flexibility to accommodate directional patterns of anisotropy; and (II) the computational effort demanded by high-dimensional inverse and determinant problems involving the covariance matrix $$\vec {V}$$ . In the case of rectangular lattice data, the spatially separable covariogram is a longstanding but underused model that can reduce arithmetic complexity by orders of magnitude. We examine a class of covariograms for stationary data that extends the separable model through affine coordinate transformations, providing a far greater flexibility for handling anisotropy than that offered by the standard approach of using geometric anisotropy to extend an isotropic model. This motivates our development of an extremely fast estimator of the orientation of the axes of range anisotropy on spatial lattice data, and a powerful visual diagnostic for nonstationarity. In a case study, we demonstrate how these tools can be used to analyze and predict forest damage patterns caused by outbreaks of the mountain pine beetle.