Variogram

It is impossible to know a specific property of an aquifer in each spatial point and at any time of the
studied domain. This due to the fact that there are several uncertainties that arise from the lack of
knowledge of parameters that describe the studied region, the measurement scale, the initial and
boundary conditions, the discretization that has been made in the models and that take into account
some hypothesis that affect the estimations.

As a result, is better to understand hydrogeological properties as random variables defined in space

In this section we obtain the experimental variogram of the logarithm of hydraulic conductivity from
measured data in different points located on Selby region.

To do this we use SGeMS software which allow us to obtain the experimental variogram by
introducing different parameters specific to the variograms, it is: lag, the interval number on which
we want to estimate the experimental variogram; lag separation, interval amplitude (the distance
between the points in which we are going to evaluate the variogram); lag tolerance, tolerance of the
lag separation.

In order to define these parameters, we look at the extension of our domain – from the provided
experimental data – and we realize that the distance from the minimum and maximum coordinate
on 𝑥 and 𝑦 directions are 10350 𝑚 and 11725 𝑚 respectively. Consequently, we define the lag as
4, the lag interval 3000 𝑚 and the lag tolerance 1500 𝑚 so we ensure that all the studied area is
covered. The variogram obtained is:

Figure 8.1
We emphasise that on the software we set a tolerance of 360 as we want the variogram is
calculated in all possible directions from the different points in which is estimated. Table 8.1 also
shows values obtained through SGeMS:

X [m] Υ (exp) Pairs

3000 0.342041 33
6000 0.430746 58
9000 0.421046 79
12000 0.426477 59

Table 8.1

Now we can fit the experimental values to different variogram models that are represented through
the equations below:

ℎ ℎ 3
𝛾(ℎ) = 𝜎 2 [1.5 ( ) − 0.5 ( ) ] 𝑓𝑜𝑟 ℎ < 𝑎 ; 𝛾(ℎ) = 𝜎 2 𝑓𝑜𝑟 ℎ > 𝑎 𝐸𝑞. 8.1
𝑎 𝑎
ℎ
𝛾(ℎ) = 𝜎 2 [1 − exp (−3 )] 𝐸𝑞. 8.2
𝑎

ℎ2
𝛾(ℎ) = 𝜎 2 [1 − exp (−3 )] 𝐸𝑞. 8.3
𝑎2

Where 𝐸𝑞. 8.1 represents the Spherical model, 𝐸𝑞. 8.2 the exponential model and 𝐸𝑞. 8.3 the
Gaussian model, 𝜎 2 is called sill and represents the variance, 𝑎 is the range and represents the
distance in which two different points are not correlated anymore, ℎ is the length of the vector that
represents the distance between two points. We note that variograms also can include the nugget
effect which is a displacement that it is made from the origin on the vertical direction, and it Is
included when there are measurement errors or data was not collected in small enough intervals;
in this case we are not considering this effect.

To fit the data with these models we first estimate values of sill and range from SGeMS for each
model, then we calculate the variogram for the four ℎ (𝑥 on table 8.1) values that we have decided
(table 8.1), we calculate the square residual and then we use solver tool of Excel to minimize them
using as parameters the sill and the range. Results are shown below:
Spherical Exponential Gaussian
X [m]
Υ (ana) RS(X) Υ (ana) Υ (ana) RS(X) Υ (ana) RS(X)
3000 0.2419 1.00.E-02 0.2419 0.3473 2.77E-05 0.3422 2.48E-08
6000 0.4267 1.65.E-05 0.4267 0.4149 2.52E-04 0.4256 2.63E-05
9000 0.4972 5.80.E-03 0.4972 0.4280 4.87E-05 0.4263 2.72E-05
12000 0.3964 9.04.E-04 0.4973 0.4306 1.69E-05 0.4263 4.74E-08
Σ 1.68.E-02 Σ 3.45E-04 Σ 5.36E-05

Table 8.2
 Spherical Exponential Gaussian
σ2 0.50 0.43 0.43
a 8901.18 5498.01 4078.07

Table 8.3

We notice that the last value of the spherical model is lower than the estimated sill, so we set that
value as the same of the sill. We also notice that the model that has the lowest sum of square
residual is the Gaussian one, this is relevant because we will use these values for different
methodologies explained in the next sections, however there are some problems with this model
that will be discussed. Finally, we observe that the estimated sill is essentially the same in
exponential and Gaussian models, and it is higher in the spherical one. Same thing happens with the
range, in which spherical model almost double the values of the two remaining models, this means
that spherical model considers a high distance in which the values of the logarithm of the hydraulic
conductivity is correlated. The variograms estimated are shown in the figure below:

Figure 8.2