In environmental studies, measurements of interest are often taken on multiple variables. The results of spatial data analyses can be substantially affected by the spatial configuration of the sites where measurements are taken. Hence, optimal designs which result in data guaranteeing efficient statistical inferences need to be studied.
We study optimal designs on two large classes of spatial regions with respect to three design criteria, which were prediction, covariance parameter estimation, and empirical prediction. The first class of regions includes those in the plane, where Euclidean distance is used. The performance of the optimal designs is compared to that of randomly chosen designs. Optimal designs for a small example and a relatively large example are obtained. For the small example, complete enumeration of all possible designs is computationally feasible. For the large example, the computational difficulty in searching for the optimal spatial sampling design is overcome by a simulated annealing algorithm.
The second class of spatial regions includes streams and rivers, where the distance is defined as distance along the stream network. A moving average construction is used to establish valid covariance and cross-covariance models using stream distance. Optimal designs for small and large examples are obtained. An application of our methodology to a real stream network is included.
We discuss the impact of asymmetry in the cross covariance function on the spatial multivariate design. The relationship between multivariate optimal design and univariate optimal design if the multivariate design is restricted to be completely collocated is studied. The efficiency lost if we consider the design that is optimal within the class of collocated designs is discussed.