yh) = 5S 2Cw, +h) ~ 2Gx,) 1? h i (N, = number of pairs, h apart) is biased upward by the quantity (1/2N,) > im(x, +h) - m(x,)1*, while the variogram of residuals, i computed after removal of an estimated drift, is biased downward. This inference problem may be approached by another method that cannot be presented here (see Matheron, 1973; Delfiner, 1975). It will suffice to say that this method allows automatic identification of the optimum local drift and variogram models within a prespecified class. This task can be performed by BLUEPACK (see Delfiner et al., 1976) a kriging program package, now available in Las Vegas on the computer at the Nevada Operations Office of the U.S. Department of Energy. STRUCTURAL ANALYSES In order to apply the kriging procedure, we need to specify the structural parameters: type of drift, if any, and variogram model. Since our goal is to estimate Pu concentrations in cells defined by the 100~-foot FIDLER grid, it is preferable to consider only the data located in that area (to guard against heterogeneities). The F(x) Field As noted above, the field of FIDLER counts is certainly not stationary, even in the log scale. The automatic structure identification module of BLUEPACK found that the best local model for the 100-foot grid F(x) data is: 1. a linear drift (i.e, in IR?: m(x,y) = bg + bix = boy) 2. a linear variogram, with a "nugget effect" (a discontinuity at the origin), i.e. y(O) 0 y(h) 0.0127 + 1.53 x 107*|h| . (|h| > 0, in feet) (10) As an indication of what is meant by "local," the program outputs a rough estimate of the maximum radius of the circular moving neighborhood within which the model is valid: here 270 feet. Also, the selected model is cross-validated by reestimation of known values as if they were unknown, Another way of checking the above variogram model is to plot it together with the raw directional variograms (Figure 6). At first glance, it may 379