k = 1, 2, 3, where k = 1 refers to the variance at h = 0, i.e. the nugget variance; the
components PC were computed. The data were then rotated to obtain scores that were kriged
separately avoiding the necessity of modeling the cross-variograms required by cokriging. There-
fore, variography was used to study the spatial correlation of both the most relevant measured
variables and the principal components. Omnidi- rectional sample variograms were calculated using
the version 2.0 of the software package GSLIB Deutsch and Journel, 1998.
To represent the variogram of each variable studied and of the principal components, a double
exponential model with a nugget was fitted, by using a semi-empirical least-squares method im-
plemented in the NLIN procedure in SASSTAT SAS, 1998. The model in its isotropic form is:
g h = C
+ C
1
{1 − e
− ha
1
} + C
2
{1 − e
− ha
2
} for
h \ 0 g
0 = 0 1
where, h is the lag distance, C is the nugget
variance, C
1
is the sill of the short-range variance, C
2
is that of the long-range variance, a
1
and a
2
are the distance parameters of the short- and long-
range structures, respectively. The distances 3a
1
and 3a
2
are the effective ranges, that is the h-val- ues, where the variogram is approximately 95 of
its sill. However, as the statistically independent PC
scores are rarely spatially orthogonal, we pre- ferred to incorporate PC technique in geostatistics
using factorial kriging Vargas-Guzman et al., 1999a,b. Multivariate geostatistical analysis was
performed using only six soil properties SAND, SILT, P01, P1, P15, OC; the variables CLAY,
P05, P2 and P10 were eliminated from coregional- ization analysis because they were redundant. All
the variables were standardized to zero mean and unit variance, so that the covariance matrix equals
the correlation matrix between the attributes. In this case, principal component analysis computed
from the covariance matrix produces the same results as that from the correlation matrix.
The classical approach to the principal compo- nent analysis PCA neglects the spatial relation-
ships among variables, whereas factorial kriging analysis FKA recognizes the correlation struc-
tures of measured soil properties. The theory un- derlying FKA has been described in several
publications Matheron, 1982; Goovaerts, 1992, 1997; Goovaerts and Webster, 1994; Wacker-
nagel, 1994; Vargas-Guzman et al., 1999a,b. Here we describe the major steps in coregionalization
analysis as applied to our data.
According to the linear model of coregionaliza- tion LMC, we assumed that all the auto and
cross-variograms are modeled as linear combina- tions of the same set of three basic variogram
functions g
k