Professional Documents
Culture Documents
ASSIGNMENT 2 23/09/2013
BRUNA KHARYN DE ASSUNCAO BARBOSA
STUDENT NUMBER 43282812
ANSWERS
QUESTION 1
Omnidirectional Experimental Variograms:
GC (2D)
GC
Model type
Nugget
Sill
Range
Major/ semi-major
Major/ minor
Spherical
6.3600
Structure 1: 14.0000
Structure 1: 93.0000
2.5993
1.2360
Variogram Maps:
Discussion I
Guibal (Guibal, 2001) states that: "variography is the calculation of experimental variograms
and the subsequent fitting of appropriate variogram models. (...) We wish to fit the
experimental variogram with a model that captures the main features of spatial variability, i.e.
the attributes that are important for the estimation of grades in space. (...) Because variography
reflects the actual spatial distribution of grades, it can be a powerful exploratory tool for the
geologist."
In order to guarantee the selection and modelling of a proper variogram, we should firstly
analyse and select certain model/kriging parameters such as lag value (h), number of lags, lag
tolerance, isotropy or anisotropy, direction and angular tolerance etc.
The first parameter to be chosen for this exercise is the lag value. The general rule for this
selection is "the lag distance should be at least equal to the sample spacing" (Coombes, 1996).
That is because h values that are too low or too high do not truly represent the variability of the
samples. Low lag values will only provide appropriate information at every number of lags,
since "variograms at intermediate lags are based on relatively low numbers of
sample pairs" (Coombes, 1996). The result of effect of low lag values can be seen in the first
variograms, especially of figures 1 and 2, where they are too erratic or "noisy".
On the other hand, variograms calculated with too high h values end up neglecting important
points and masking real variability information. That makes variograms too "smooth", such as
can be seen in figures 7 and 8.
In the studied case we have a 20m X 20m grid. So the variogram of h= 25 seems appropriate
enough, showing a nice curve that is smooth, has low nugget and long range, which will be
good for modelling.
Another point to be analysed in this exercise was the matter of directional behaviour. When
choosing the lag value, the experimental variograms were calculated assuming the body was
isotropic, that is, had the same variation in all directions. However, by plotting and looking at
variogram maps for gold and thickness (figures 11 and 12), it is clear that there is a change in
variability according to direction. Isotropic variogram maps should not show any important
changes in colours. In this case, though, there is obvious variation, indicating anisotropy.
Thus, the model presented in figure 9 is not valid. New, directional variograms needed to be
calculated and modelled to express the true variability of the deposit.
Model type
Nugget
Sill
Range
Major/ semi-major
Major/ minor
GC
Spherical
7.0194
Structure 1: 2.8762
Structure 2: 10.1268
Structure 1: 53.3500
Structure 2: 100.2200
4.2194
6.4061
Thickness
Spherical
0.5433
0.4869
73.2770
2.7784
1.0000
Discussion II
Figures 13 to 16 show the final directional variograms that were selected for grade and
thickness as the most appropriate from a number of variograms that were tested. Table 2
summarises the parameters used in the modelling process.
The resulting variograms both have low nugget effect, that is, their variability at short distances
is not too high. The range, which is the distance at which the variance stabilises and the samples
are no longer correlated, is long. That means that samples separated by a large distance are
likely to be somewhat similar to each other.
These variograms indicate a reasonably continuous distribution of the variables, with a gradual
variation up to a maximum.
As to the modelling process, we can see that for some cases the model has a higher sill than the
expected (green line). That is because real variability does not necessarily follow mathematical
expectations. The variogram calculations indicate that the sill should always be achieved at the
variance value. Nevertheless, reality is messy and the modelling of the true sill depends much
on the modellers sensibility.
QUESTION 2
The block model is a geological representation of a deposit that evaluates grades and other
variables from the drill hole data by interpolating the drilling data, limited within a certain
wire frame model, using geostatistical techniques. The purpose of the block model is to
associate grades with the volume model. Two of the main features of the block model are the
size of the blocks and the model rotation. The former is defined according to the spacing/
drilling patterns, being as small as possible; the latter is used to adjust the block axis to be
parallel to the strike of the ore body, to ensure a good intersection between the model and
the ore body and exclusion of excessive waste.
The blocks extents and attributes are summarised in Table 3 and 4.
Type
Minimum Coordinates
Maximum Coordinates
User Block Size
Min. Block Size
Rotation
Total Blocks
Y
7320
7620
10
10
0.000
X
1570
1970
10
10
0.000
1200
Z
200
201
1
1
0.000
Attribute Name
gold_id
gold_nn
gold_ok
gold_thk
zok_ads
zok_cbs
zok_dns
zok_ke
zok_kv
zok_ns
Type
Real
Real
Real
Real
Real
Real
Real
Real
Real
Integer
Decimals
2
2
2
2
2
2
2
2
2
-
Background
0
0
0
0
0
0
0
0
0
0
Description
Gold value estimated by inverse distance
Gold value estimated by nearest neighbour
Gold value estimated by kriging
Estimated mineralisation thickness
Average distance to samples
Conditional bias slope
Distance to nearest sample
Kriging efficiency
Kriging variance
Number of informing samples
QUESTION 3
For any technique applied in an estimation, it is important to choose certain parameters that
define a search strategy, which controls the samples that really contribute for the estimation
(informing samples).
The first parameter is the search type that could be either ellipsoid or octant. The octant, or
the quadrant for 2D, defines a maximum of N points in each of the eight octants, or four
quadrants, to be used in the interpolation calculations showing high performance with
clustered data. For data that shows anisotropy, such as in this case, the ellipse method is
better because it controls the shape of the search space that surrounds the interpolation
point and uses only a subset of the scatter points in the vicinity of the target for the estimation
calculations. That is why we use the ellipse for this data.
The minimum and maximum number of samples control how many of the points found in the
search radius are actually used in the kriging calculations. If fewer than the minimum value
are found, a default value is used. If greater than the maximum, the closest points are used.
The maximum search ellipse is used in conjunction with the minimum and maximum number
of samples to select the appropriate samples for kriging, and it should generally be a value
slightly above the range of the variogram of the major axis.
The search anisotropy ratio is associated with the sizes of major and minor ellipse axis, whose
lengths are based on the spatial continuity defined in the variography analysis.
The search parameters used are presented in Table 5.
Parameter
Search Type
Min. Number of Samples
Max. Number of
Samples
Max. Search Ellipse
Search Anisotropy
Ratios
Feature
Ellipse
3
12
100
2.83
Table 5: Search parameters
QUESTION 4
Discussion
Nearest Neighbour:
Assigns values to blocks in the model by assigning the values of the nearest sample point to
the block of interest. It is a very simple technique and it is effective if training data is large.
It is, however, of computational complexity, has problems with memory limitations and is
easily compromised by irrelevant data.
Inverse Distance Estimation
It is one of the earliest estimation methods, using interpolation based on an empirical
observation that the weight of each sample is proportional to an inverse power of the
distance.
The degree of smoothing, or variance reduction may be controlled by changing the weighting
power and search parameters. A lower power results in the estimation being more
continuous. Higher weighting powers on the other hand result in estimations that seem more
erratic. The appropriate parameters may only be found through trial and error.
Indicator Kriging:
Indicator kriging and probability kriging are related methods that are used to improve
estimation when ore zones are erratic and grade distributions are highly variable and
complex. Advantages of indicator kriging include less smoothing of estimated grades than
ordinary kriging and robustness in handling nonstandard grade distributions.
The first step in indicator kriging is to set one or more cut-offs with which to define indicator
variables. Given a cut-off gc, the indicator variable is set to 1 if the grade is above gc or 0 if the
grade is below gc; indicator variables are coded similarly for each desired cut-offs.
The resulting indicator estimates may be interpreted as either the probability that the block
will be above cut-off or the percentage of the block that is above cut-off.
Variograms are modelled for each indicator variable and an expected value for each indicator
is estimated using ordinary kriging and the appropriate indicator variogram.
Ordinary Kriging
It is a distance weighting technique where weights are selected via the variogram according
to the samples distance and direction from the point of estimation. The weights are not only
derived from the distance between samples and the block to be estimated, but also the
distance between the samples themselves. This tends to give much lower weights to
individual samples in an area where the samples are clustered.
Lognormal Kriging:
It is a method of non-linear kriging that was developed to improve estimation when the
underlying data are distributed according to a lognormal probability distribution.
The variogram is computed using the natural logs of the data; the kriging system is solved to
provide a weighted average of the natural logs of the data; the kriged log average is then
transformed back to normal values using lognormal transformation.
It has complex mathematics and presents many complications in its application, such as the
strict requirement for a lognormal distribution and a variogram which is stationary over the
field of estimation. Serious global and local biases may occur if either these conditions are not
met. In addition, there is a tendency for lognormal kriging to overestimate the high-grade end
of the population when the coefficient of variation is greater than 2.0.
Lognormal kriging is recommended only for special purposes where the results can be
monitored closely and adjusted to prevent biases.
Method
Nearest Neighbor
Advantages
Simplicity
Effective if training data is large
Inverse Distance
Simplicity
Fast calculation
Reasonable Results
Ordinary Kriging
Sensitive to clustering
Unbiased estimation
Error estimation and mapping
Suitable for distributions that are
positively skewed
Improves estimation when the
underlying data are distributed
according to a lognormal probability
distribution
Improves estimation when ore zones
are erratic and grade distributions
are highly variable and complex
Less smoothing of estimated grades
than ordinary kriging
Robustness in handling nonstandard
grade distributions.
Log-normal kriging
Indicator kriging
Disadvantages
Computation Complexity
Memory limitation
Computationally Slow
Easily fooled by irrelevant
attributes
Choice of weighting function may
introduce ambiguity
Not sensible to cluster regions
Does not have a measure of error
Difficulty to define the variogram used.
Not a suitable method for data sets
which present boundaries
Complex mathematics
Many complications in its application
Strict requirement for a lognormal
distribution and a variogram stationarity
Possible biases and overestimations
Time and effort to do full indicator
variography
Order relation problems needs to be
controlled
QUESTION 5
Input data:
Number of samples
Minimum value
Maximum value
Mean
Median
Variance
Standard Deviation
Thickness
194
1
8
3.197165
3
0.861087
0.927948
Output data:
Variable
Number of samples
Minimum value
Maximum value
Mean
Variance
Standard Deviation
gold_id
gold_nn
gold_ok
gold_thk
Gold Ok
756
0.355292
16.315715
2.701896
7.706975
2.776144
0.9754
0.7841
1
-0.1661
Gold Thk
756
1
8
3.146296
0.613591
0.78332
-0.1316
-0.0672
-0.1661
1
Q-Q Plot 4 GC OK
Discussion
The true value of gold and thickness (input data) and the estimate value (output data) should
be the same or near the same. And the statistic parameters such as mean, variance and
standard deviation, should decrease, since the data values were treated with the estimation
methods. It is clear that all estimated data have better statistics parameters after comparing
Table 7 to Table 8. And it is also noticeable the efficiency of the methods, just comparing
those parameters to each other. From that we can notice that the best estimation method in
this case was the ordinary kriging, and that the nearest neighbour algorithm did not present
satisfactory results. It is also noticeable that the thickness value were already well distributed,
so the estimation methods did not change its statistics parameters much.
Q-Q plots are used to compare two data sets where samples are not necessarily paired. They
use the values of the percentiles of the analysed samples, plotted against each other in order
to check how close the data sets distributions are. In the Q-Q plot charts above, it is possible
to compare the distributions of input and output thickness and gold grade. The estimated and
input values for thickness follow the 1:1 line very closely for most of the values. That indicates
they are very close, and there are not many deviations.
For the GC Q-Q plots it is noticeable that kriging did not achieve such a good estimation of the
data. Despite being clearly better than nearest neighbour and inverse distance estimations,
kriging also resulted in a few deviations. That might be explained by the skewness of the gold
grade data as proved by the histograms (Figures 17to 22). Kriging does not deal very well this
type of distribution.
QUESTION 6
Plan views
Cut-off grade
>0
>0.25
>0.5
>0.75
>1.0
>1.25
>1.5
>1.75
>2
>2.25
>2.5
>2.75
>3.0
Grand Total
Nearest Neighbour
Mineralised
Gold (g/t)
Tonnes
33288
93423
27892
28861
23788
9538
36917
23294
12084
25802
15314
14288
107445
451934
0.22
0.32
0.61
0.87
1.12
1.34
1.6
1.83
2.14
2.33
2.55
2.84
6.93
2.5
0
64866
46645
30552
38570
33193
31645
20045
19884
17157
11818
10564
126996
451934
Table 9: Grade and tonnages
0
0.42
0.61
0.88
1.13
1.35
1.63
1.87
2.12
2.37
2.65
2.9
6.06
2.6
Ordinary Kriging
Mineralised
Gold (g/t)
Tonnes
0
47766
56031
33953
44080
33649
27208
21518
16302
21480
10165
12825
126958
451934
0
0.44
0.6
0.87
1.13
1.39
1.61
1.86
2.12
2.36
2.62
2.87
5.95
2.59
Cumulative tonnes
8.00
7.00
6.00
5.00
4.00
3.00
Cumulative tonnes
2.00
Cumulative grade
1.00
0.00
Cut-off grade
Figure 25: Grade tonnage chart - NN
Cumulative tonnes
7.00
6.00
5.00
4.00
3.00
Cumulative tonnes
2.00
Cumulative grade
1.00
0.00
Cut-off grade
Figure 26: Grade tonnage chart - ID
Cumulative tonnes
7.00
6.00
5.00
4.00
3.00
Cumulative tonnes
2.00
Cumulative grade
1.00
0.00
Cut-off grade
Figure 27: Grade tonnage chart - OK
500000
450000
400000
350000
300000
250000
200000
150000
100000
50000
0
8.00
7.00
6.00
5.00
4.00
3.00
Axis Title
Cummulative tonnes
2.00
1.00
0.00
Cut-offs
Figure 27: Grade tonnage chart for all estimates
Discussion
Plan views:
Plan views are a visual form of analysis employed in model validation, in which we should
check for large differences between drillhole composites and the estimated grade. If those
exist, they must be analysed and explained.
References