“…Following the comprehensive reviews given by Zimmerman and Zimmerman (1991), and by Cressie (1993), these can be divided into least squares based techniques in the space of the semi-variogram (Journel and Huijbregts, 1978;Cressie and Hawkins, 1980;Cressie, 1985); least squares techniques in the space of observations defined in the form of generalised covariance expressed as a linear function of parameters (Delfiner, 1976;Kitanidis, 1983); Maximum Likelihood in the space of residuals from a linear trend (Mardia and Marshall, 1984); Maximum Likelihood in the space of cross-validation errors (Samper and Newman, 1989); Maximum Likelihood in the space of "error contrasts" (Kitanidis, 1983), using what is known as Restricted Maximum Likelihood (REML) (Patterson and Thompson, 1971;1974). All these techniques have pros and cons that will be addressed briefly in the sequel, but this paper will focus only on the ML and REML type estimators, since they allow for the derivation of the covariance matrix of the parameters, which can be computed as the inverse of the Fisher information matrix, and can be used for investigating the effect of parameter uncertainty over the Kriging estimates, as advocated by Kitanidis (1983) and proposed by Todini and Ferraresi (1996).…”