This paper deals with three related problems in a geostatistical context. First, some data are available for given areas of the space, rather than for some specic locations, which creates specic problems of multiscale areal data. Second, some uncertainties rely both on the input locations and on measured quantities at these locations, which creates specic uncertainty propagation problems. Third, multidimensional outputs can be observed, with sometimes missing data. These three problems are addressed simultaneously here by considering mixtures of multivariate random elds, and by adapting standard Kriging methodology to this context. While the usual Gaussian setting is lost, we show that conditional mean, variance and covariances can be derived from this specic setting. A numerical illustration on simulated data is given.