A regionalization method based on a cluster probability model

作者：Cowpertwait Paul S P^*

来源：Water Resources Research, 2011, 47: W11525.

DOI：10.1029/2011WR011084

[1] A regionalization method based on a cluster probability model (a mixed multivariate Gaussian model) is proposed for grouping sites into nonoverlapping contiguous homogeneous regions defined by a Voronoi tessellation (Theisson polygons). The cluster probability model is applied to second-order standardized annual sample properties (mean, coefficient of variation, and autocorrelation) evaluated at the daily level of aggregation taken from each of 234 daily rainfall records with positions located in the Basque Country, Spain. Using the Bayesian information criterion, four clusters of sites are identified (which do not fall into contiguous regions). The distances between all neighboring pairs of sites connected by edges from the Delaunay planar graph are found. The probability that a site belongs to each of the four clusters is extracted from the fitted Gaussian model and multiplied into the probability that the neighboring site belongs to the same cluster. These products are divided by the squared distance between the sites and are summed to give an overall measure of a site belonging to a cluster that takes into account the classification of neighboring sites. Regions from the Voronoi tessellation of the points are classed on the basis of this measure and according to whether they are spatially isolated from other regions of the same class. Points that have the least influence on the variance of residual errors of the fitted model are found using a criterion based on Wilks' lambda for multivariate analysis of variance, and the classes of the least influential points are adjusted to ensure the overall regions are contiguous.

• 出版日期2011-11-29