Non-stationary spatial covariance structure estimation in oversampled domains by cluster differences scaling with spatial constraints

In the analysis of spatiotemporal processes underlying environmental studies, the estimation of the non-stationary spatial covariance structure is a well known issue in which multidimensional scaling (MDS) provides an important methodological approach (Sampson and Guttorp in J Am Stat Assoc 87:108–119, 1992). It is also well known that approximating dispersion by a non-metric MDS procedure offers, in general, low precision when accurate differences in spatial dispersion are needed for interpolation purposes, specially if a low dimensional configuration is employed besides a high number of stations in oversampled domains. This paper presents a modification, consisting of including geographical spatial constraints, of Heiser and Groenen’s (Psychometrika 62:63–83, 1997) cluster differences scaling algorithm by which not the original stations but the cluster centres can be represented, while the stations and clusters retain their spatial relationships. A decomposition of the sum of squared dissimilarities into contributions from several sources of variation can be employed for an exploratory diagnosis of the model. Real data are analyzed and differences between several cluster-MDS strategies are discussed.