Computers & Geosciences 79 (2015) 47–57

Contents lists available at ScienceDirect

Computers & Geosciences

journal homepage: www.elsevier.com/locate/cageo

Multiscale hierarchical domaining and compression of drill hole data

E.J. Hill n, J. Robertson, Y. Uvarova
CSIRO, PO Box 1130, Bentley, WA 6102, Australia

art ic l e i nf o a b s t r a c t

Article history: New drilling methods, currently under development for minerals exploration, combined with rapid data
Received 15 December 2014 collection by a range of sensors means that the end-user is confronted with increasingly large data sets.
Received in revised form In order to reduce the stream of data into objects which represent meaningful geological features we
9 March 2015
need to incorporate spatial information into the analysis. Boundary detection incorporating multiscale
Accepted 9 March 2015
Available online 11 March 2015
considerations has previously been carried out using a scale–space plot from a wavelet transform. We
present a method which applies a rectangular tessellation to the wavelet transform, this has the ad-
Keywords: vantage of being easier to interpret, as it resembles a geological log. In addition, the tessellation records
Wavelet transform hierarchical information for different scale objects. The tessellation can be filtered in order to remove
Tessellation
unwanted variation (including noise) from the results. When applied to geochemical data, the resulting
Lithochemistry
tessellation provides a basis for classification of lithochemical units that is more reliable than classifi-
Filter
Logging cation by considering individual samples without spatial context.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction Traditional methods of classification and discrimination of li-

thological units based on geochemical data generally do not take
Discoveries of mineral resources are declining in Australia and spatial information into account (see for example, Mullen, 1983)
future exploration is being targeted at deep basement rocks that and this may result in spurious lithological boundaries when the
lie beneath barren sedimentary cover. In order to facilitate the classification is plotted as a down hole geological log (see ex-
collection of geological information on deep basement rocks, low amples in next section).
cost, rapid drilling methods are being developed (Hillis et al., A different approach has been taken by geophysicists using
2014). Traditionally, geological information is extracted from the downhole geophysical tools such as gamma-ray and resistivity
drilling products by visual logging by trained geologists. However, logs, where signal processing techniques have been applied (re-
this method is slow and expensive, and more importantly, there is placing the time axis with depth) so that the spatial relationships
a lack of confidence in the consistency of logging between between data points are preserved. In particular, wavelet trans-
geologists. forms are popular due to their mathematical efficiency and mul-
As a result, automated methods of collecting data at the drill tiscale-edge detection ability (e.g. Panda et al., 2000; Arabjamaloei
et al., 2011; Perez-Munoz et al., 2013). In this paper we propose an
site are being developed; designed to be cheaper and faster than
alternate method for visualising the wavelet transform of geolo-
sending samples to an analytical laboratory. These include col-
gical data using the wavelet tessellation method of Witkin (1983).
lection of geochemical and mineralogical data from portable X-ray
The advantage of this method is that it provides a diagram that is
fluorescence spectrometer (XRF) and X-ray diffraction (XRD) de-
similar in appearance to a geological log, and therefore a familiar
vices (“Top-of-hole sensing”, Hillis et al., 2014). Converting the
format to geologists, it preserves the hierarchical multiscale in-
streams of geochemical and mineralogical data into geological logs formation provided by the wavelet transform and it is a very
provides a useful visual summary of the information. Geological compact summary of the data.
logs can be used to document important lithological boundaries,
classify lithological units and recognise corresponding lithological 1.1. Deriving lithochemical units from geochemical data
units in different drill holes in order to build up a 3D model of the
local geology. When a geologist manually logs a drill hole they usually in-
corporate spatial relationships in their interpretation, in particular,
n
Corresponding author.
the concept that samples that are in close proximity may be
E-mail addresses: [email protected] (E.J. Hill), grouped into a single class even if they have somewhat different
[email protected] (J. Robertson), [email protected] (Y. Uvarova). characteristics. When lithological units are classified using only

http://dx.doi.org/10.1016/j.cageo.2015.03.005
0098-3004/& 2015 Elsevier Ltd. All rights reserved.
48 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57

desirable to be able to summarise the data in a manner that not

only preserves information at different scales (so that it can be
extracted as required) but also summarises the hierarchical
structure of the higher level groupings. The tessellation method of
Witkin (1983) was chosen as it provides both multiscale and
hierarchical information in a succinct format. Furthermore, we
propose a user-controlled filtering procedure that allows un-
desirable low level variation to be removed from the tessellated
log as required.

1.2. Multiscale edge detection

Edge detection algorithms are used to identify points of sharp

variation in a measured signal. These edges (or boundaries) se-
B parate regions of low variation, which can be interpreted to re-
present objects in the signal. Edge detection algorithms have been
widely used in image analysis to help identify objects in images;
however, traditional edge detecting algorithms work at a pre-
defined scale (e.g. Marr and Hildreth, 1980; Canny, 1986). This
means that the scale of edge detection has to be determined be-
fore the calculation proceeds. If too small a scale is selected, then
we may have a very large number of edges, many of which may
represent noise or information of no interest at the scale of the
problem under consideration. If, on the other hand, we select a
scale that is too large, we may lose useful information about in-
teresting small scale features. Using a continuous wavelet trans-
form to detect edges has the advantage that it works simulta-
neously across a range of scales. Multiscale edge detection meth-
Fig. 1. Data from DETBrukunga02 classified into three groups according to location
on a compositional ternary plot (Fe–Al–S). The column plot shows the drill hole ods for images using the continuous wavelet transform were de-
intervals coloured according to the same classification scheme. Selected points on veloped by Mallat (1991), Mallat and Hwang (1992) and Mallat and
the drill hole are linked to their location on the ternary plot. In (A) the section of Zhong (1992). Their aim was to be able to compress image files by
unit Y in the column plot labelled “a” is compositionally distinct from the sur-
removing unwanted noise and preserving important edges. This is
rounding unit Z, but the section of unit Y labelled “b” is very similar to the sur-
rounding unit Z. In (B) a narrow section of unit Y indicated on the plot may be a
directly analogous to deriving geology logs from analytical data,
mixed sample of the units X and Z. where the geologist only wishes to detect important changes in
the measured property and ignore weak variations.
The wavelet transform has been used for detecting significant
non-spatial features spurious lithological boundaries may result. lithological boundaries in geophysical log data (typically from
Two examples of which are illustrated in Fig. 1, which shows a petroleum or groundwater wells). These include applications to
ternary plot of geochemical data from an experimental drill hole gamma ray, sonic and resistivity logs (e.g. Arabjamaloei et al.,
from the Deep Exploration Technologies Cooperative Research 2011; Cooper and Cowan, 2009; Davis and Christensen, 2013). In
Centre (DET CRC) project in Brukunga, South Australia. In the the petroleum industry, the technique is referred to as blocking or
figure hard boundaries have been assigned to the ternary plot in zoning, the aim of which is to smooth the log data in regions of
order to separate different lithological units (without reference to relatively homogeneous rock properties (i.e. regions of weak var-
any spatial information). These lithochemical classifications can be iation) while leaving the sharp boundaries (i.e. points of strong
plotted in a column to represent the lithological units in the drill variation). A comprehensive overview of recent applications in the
hole. Spurious lithological boundaries can result from a unit whose use of the wavelet transform for analysing well log data is given in
chemical variation crosses a classification boundary. Fig. 1A shows Perez-Munoz et al. (2013).
an example of this problem: if the spatial location had been taken A wavelet is a small oscillation that is finite in extent and has a
into account the unit labelled “b” would probably have been as- specified frequency. The wavelet transform is a convolution of the
signed to the same rock type as the surrounding units due to their wavelet with the signal. The wavelet can be translated in space
similarity in composition. Spurious lithological boundaries can (along the signal) and stretched (to operate at different scales).
also occur when a sample is a mixture of two rock types. For ex- There are a large number of different wavelet shapes which are
ample, in Fig. 1B it is possible that the unit classified as unit Y may suitable for different purposes. For example, Chandrasekhar and
actually be a mixture between adjacent X and Z units. Eswara Rao (2012) performed a comparison study using a number
To correct this spatial data problem we propose to first de- of wavelet transforms, which demonstrated that the Gaussian
termine the boundaries of the lithological units and then classify wavelet transforms were the most useful in delineating lithologi-
the units by using all the data points between a pair of boundaries cal boundaries from geophysical logs.
as a single lithological unit rather than classifying each data point We use the Gaussian wavelet transform (known as the DoG or
separately. We do not address the problem of classification in this Hermitian wavelet) for edge detection as it has certain properties
paper, only the problem of boundary identification. which make it favourable for this application, these are discussed
When logging a drill hole geologists do more than simply de- further in Section 2. The second order (i.e. second derivative) DoG
tect boundaries and classify units, they also make decisions re- is also known as the Mexican hat wavelet. Using the DoG wavelet
garding splitting and grouping of adjacent lithological units (see as an edge detector is equivalent to applying a Gaussian smoothing
for example, the combining versus separating and single versus function to the signal at various scales and then detecting inflec-
many dichotomies documented by Shipley et al., 2013). The degree tion points in the smoothed signal using the derivatives of the
to which units are grouped depends on the scale of the study. It is signal. Inflection points are interpreted to represent edges of
 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57 49

values. The wavelet coefficients are proportional to the values of

the derivatives for a Gaussian transform of the signal (Mallat and
Zhong, 1992). For drill hole signals, “space” is depth down the drill
hole. Note that scale on the x-axis is plotted as log(scale) to make
the smaller scale features clearer.
The boundary of the cone of influence (coi) is plotted in white
on the scale–space plots; it is a region in which the transform is
affected by the boundaries of the data. Because the drill hole signal
is of finite length the signal is padded at either end by sufficient
zeros to allow coefficients to be calculated near the ends (Torrence
and Compo, 1998). However, the zeros introduce discontinuities at
the ends of the signal. Caution is required when interpreting the
transform in the coi as the power of the transform is decreased in
this region (Torrence and Compo, 1998).

2. Methods

2.1. Continuous wavelet transform

We use a continuous wavelet transform rather than a discrete

wavelet transform (dwt). The continuous wavelet transform cal-
culates the transform for a continuous range of scales while the
Fig. 2. (A) synthetic signal containing a series of sharp variations with differing dwt only calculates the transform for a predefined discrete set of
scale and gradient; (B) smoothed signal; (C) and (D) signal convolved with first and
scales. This means that the dwt methods cannot accurately locate
second order DoG wavelet, respectively. Dotted lines illustrate the location of se-
lected signal inflection points which correspond to extrema and zero crossings in sharp boundaries due to the small number of scales used in the
the first and second order derivatives, respectively. transform (unless the scale of the boundary fortuitously matches
the scale of the transform, e.g. Chandrasekhar and Eswara Rao,
objects in the signal. The inflection points in the smoothed signal 2012). In a practical application, Cooper and Cowan (2009) de-
are represented by extrema (i.e. maxima or minima) in the first monstrated that the continuous wavelet transform performed
derivative or by zero crossings in the second derivative (Fig. 2). better than the dwt for analysing both noisy synthetic data and
The plots in Fig. 2 show the wavelet transform for a single scale
real geophysical data from iron ore deposits. The continuous wa-
only; a scale–space plot (Fig. 3) is used in order to illustrate the
velet transform is less computationally efficient than the dwt;
transform over a range of scales, they are also known as scalo-
however, this is not an issue in the case of drill hole data where the
grams. In these plots, the values of the wavelet coefficients are
data sets are 1 dimensional and not particularly large.
indicated by shading; stronger colours representing more extreme

Fig. 3. Scale–space plots of the wavelet transform of a synthetic signal (signal shown to the left of the transform): (A) first derivative, (B) second derivative and (C) zero
crossings of second derivative shown as grey lines, they separate regions in the scale–space plot that may represent objects in the signal. Note: zero crossings at very small
scales in (C) are spurious (
6–7), they are created by numerical errors in regions where the wavelet transform is close to zero.
50 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57

2.2. Wavelet transform and the scale–space plot Fig. 5A, the two separate triangular objects near “depth” 150 on
the synthetic signal exist as two separate features at small scales
The important properties of the DoG wavelet that make it (
101), but become merged into a single feature at larger scales,
useful for detecting boundaries in signals and images are: (1) it is i.e. between 101 and 102. In addition, the three lows in the signal at
continuously differentiable so it produces a smooth set of curves
50, 90 and 130, which form separate features at small scales
for the zero crossings and (2) the zero crossings will not disappear (
101), merge into a single feature at a scale of
102. At 102 the
as scale decreases (Section 6.2, Mallat, 2009; Mallat and Hwang, original 5 positive features on the signal have been reduced to the
1992; Witkin, 1983). Furthermore, the DoG wavelet conforms to 2 strongest features. If we add random noise to the artificial signal
the “localisation” assumption of Witkin (1983), which states that (see Fig. 5D), the hierarchical structure of the domains outlined by
the true location of an edge in the signal (at any scale) is given by the zero crossings becomes more complex at small scales, but the
the location of the zero contours as the scale approaches zero. This same features observed in the non-noisy signal still dominate at
is important because smoothing of the signal at large scales causes the larger scales.
broadening and flattening of the features and can result in dis-
location of the zero contours in space. This tendency is illustrated 2.3. Tessellation of a wavelet transform
in Fig. 3C where the zero contours are relatively parallel to the x-
axis at small scales but may diverge considerably at larger scales Witkin (1983) noted that applying a wavelet transform to a
(4 102). signal to produce a scale–space image (such as Fig. 5A) actually
Applying a wavelet transform to a signal increases the volume increases the volume of data and that this is an undesirable out-
of data to be interpreted and the scale–space plot of a signal can be come because a signal description should be as compact as pos-
difficult for people unfamiliar with the wavelet transform to un- sible and its units should correspond as closely as possible to
derstand. In the rest of this section we provide some guidance in meaningful objects. For this reason, Witkin (1983) proposed that
interpreting the scale–space plot using synthetic signals; we show the scale–space plot can be summarised by a rectangular tessel-
how varying the object wavelength and strength affects the lation which concisely describes the structure of the signal over all
transform and how weak features (including noise) are combined scales of observation. The advantage of the tessellation when
or absorbed at larger scales. Then in the next section we show the compared to traditional blocking or zoning methods (e.g. Hoyle,
information in the scale–space plot can be succinctly summarised 1986; Cooper and Cowan, 2009) is that it preserves the scale in-
as a tessellation. formation in a hierarchical structure, unlike the blocking method
The location and relative value of extrema (2D maxima and which compresses the scales and presents the multiscale edges as
minima) in the coefficients is important in interpreting the scale– an undifferentiated set of boundaries.
space plot. We use two simple synthetic signals to illustrate how The tessellation subdivides the scale–space plot into a set of
the location and size of the extrema are affected by the amplitude rectangles with a hierarchical structure. The tessellation is a rec-
and wavelength of objects in the signal. The scale–space plot in tangularised version of the scale–space domains bounded by the
Fig. 4A shows that increasing the amplitude without changing the zero crossings of the second order DoG transform. The boundaries
wavelength increases the value of the extrema; however, the scale of the scale–space domains appear as a set of arches terminating
at which the extrema occur is unchanged. Fig. 4B shows that in- on the y-axis, see Fig. 5A and D (the tops of larger arches maybe
creasing the wavelength without changing the amplitude changes truncated by the right edge of the plot). The first step in the tes-
both the value and scale of the extrema. sellation is to convert each arch to a rectangle (Fig. 5B). The top
The hierarchical structure illustrated by the wavelet transform and bottom of each rectangle is given by the locations at which the
is demonstrated by the scale–space plots in Fig. 5A and D. In zero contour intersects the y-axis; i.e. the locations at the smallest

Fig. 4. Scale–space diagrams illustrate the effects of changing the signal's (A) amplitude and (B) wavelength on the second order DoG transform.
 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57 51

Fig. 5. Scale–space plots of the second order DoG transform: (A) synthetic signal with zero contours highlighted in grey, zero contours enclose scale–space domains;
(B) conversion of irregularly shaped scale–space domains into rectangles; (C) tessellation of scale–space domain into a hierarchical parent–child structure; (D) synthetic
signal with added random noise and (E) tessellation of scale–space domain for noisy signal (note that at larger scales the tessellation is almost identical to that of the non-
noisy signal).

scale. The right hand edge of the rectangle is given by the max- each for the scale and location). The tessellation is also a good
imum scale of the domain. In the second step, for each domain representation of a geological log because it mimics the traditional
that contains a sub-arch, the corresponding rectangle is sub- style of geological logging by placing straight line boundaries at
divided into new rectangles as illustrated in Fig. 5C and E. The significant changes in composition, so the output is more intuitive
subdivision results in a hierarchical parent–child relationship (i.e. to a geologist than that of the wavelet transform scale–space plot.
a ternary tree structure), this structure is illustrated in a hy- It is also a very useful representation because it preserves the
pothetical example in Fig. 6. The ternary tree structure is useful for hierarchical structure according to the scale of the objects, so that
filtering the tessellation (Section 2.4). At the edges of the data (i.e. smaller, weaker objects can be grouped into larger, more dominant
top and bottom of the drill hole) it may be necessary to pad the objects.
data in order to preserve the ternary structure; this is hidden in
the figures of the tessellations. 2.4. Filtering the tessellation
The tessellation forms a very compact encoding of the wavelet
transform's scale–space geometry as the complex pattern of scale– The wavelet transform can be filtered to reduce undesirable
space domains can be reduced down to a simple array containing information in the signal, for example, the signal may contain
four values for each rectangle (i.e. minimum and maximum values noise from measurement error. Another source of undesirable
52 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57

Fig. 6. Schematic plot showing the relationship between scale–space domains and
the tessellation rectangles. Domains in the scale–space plot are defined by arches in
the zero contours. The arches terminate on the smallest scale (i.e. the y-axis). In this
example, the larger domain contains a sub-domain. As a result the rectangle re-
presenting the larger domain is subdivided into four new rectangles in a hier-
archical tree structure.

information occurs when there are small variations in the mea-

sured property that are considered by the geologist to represent
natural variation within a single lithological unit. We suggest that
a flexible approach to selecting a filter threshold is most appro-
priate for geological problems as the degree to which the data
should be filtered is dependent on the purpose of the geological
logging. When filtering is used in order to remove noise from a
signal there are a number of methods derived for calculating a
reasonable value for the threshold, based on certain assumptions
about the stochastic nature of the noise process (e.g. Donoho and
Johnstone, 1994). In the case of geological measurements, the
amplitude of the unwanted natural variation is typically much
Fig. 7. Small second order wavelet coefficients in the scale–space plot are overlain
larger than that of any measurement noise. Therefore the thresh- by a gray mask indicating which regions would be filtered out by a simple coeffi-
old required for removing unwanted natural variation in the signal cient threshold filter.
will usually exceed that of the noise. The concept of what con-
stitutes a lithological unit will vary depending on the purpose of
the interpretation of the data. A regional study may result in larger transform for blocking signals, their methods are applicable to a
units with more internal variation than would be acceptable for a true wavelet transform. In this paper we have adapted the concept
detailed study. For these reasons it is desirable that the filtering of Davis and Christensen (2013) of filtering by region but applied it
threshold be set by the user. to the tessellation. We do not follow the filtering method proposed
It is straightforward to filter the wavelet transform by removing by Witkin (1983) as this is designed to detect the edges in the
the smallest scale features; this would simply require truncation of signal or image which are considered most perceptually salient to
the left hand side of the scale–space plot (e.g. Cooper and Cowan, humans, a factor which may not be relevant to finding boundaries
2009). However, in minerals exploration, small objects may be in geological data.
important, as valuable mineral deposits often occur in narrow The tessellation is filtered by removing rectangles that fall be-
tabular bodies, and large objects that have a weak signal may be low the threshold and will be described by referring to Fig. 6. The
unimportant as they are not sufficiently different from their tessellation filter considers only the middle child rectangles in the
neighbours to be considered as a separate unit. Filtering by wa- ternary tree structure, if all the wavelet coefficients in the region
velet coefficient is more desirable because it removes weak edges represented by the middle child fall below the selected threshold,
irrespective of their scale. This means that boundaries between then the middle child is merged into its parent and siblings to
similar objects will be filtered out, keeping only those boundaries form a new rectangle which spans a larger scale range as well as a
that occur between contrasting objects. larger space. This is equivalent to simply removing the zero con-
If a simple coefficient threshold is applied to the scale–space tour which separates the below-threshold region inside the arch
plot (of the second order DoG transform) then the result is an ir- from the enclosing region. The filter iterates until there are no
regular filtered region, as illustrated in Fig. 7, and it is not clear more children below the threshold value.
how to interpret this result in terms of geological boundaries and By varying the filter size we can reduce the complexity of the
objects. Davis and Christensen (2013) segregated a transformed tessellation by removing unwanted information. In the examples
geophysical signal into regions bounded by multiscale edges (re- in this paper the threshold is expressed as a percentile of the
gions similar to the domains in Fig. 5A) and, instead of filtering out coefficient values as the actual coefficient values have little
individual coefficients, they filtered entire regions based on aver- meaning to a geologist. Fig. 8 shows that, for this example, we can
age coefficient values for each region. Their goal was to cleanly completely remove random noise added to the artificial signal
remove unimportant sedimentary layers from the signal. Although with a 40 percentile filter (Fig. 8C). And we can achieve a very
they used a piece-wise linear approximation of the wavelet simple structure with only the strongest parts of the signal
 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57 53

Fig. 8. Tessellation coloured by absolute maximum wavelet coefficient of the sub-domain. At each step filtering removes the rectangles with the weakest coefficients (i.e.
dark colours) while preserving the sub-domains with high values (paler colours). (A) unfiltered tessellation, (B) 30 percentile filter, (C) 40 percentile filter and (D) 95
percentile filter.

preserved by applying a 95 percentile filter; this extreme level of indicators of lithology. We demonstrate how the tessellation is useful
filtering may be useful for a regional scale study (Fig. 8D). for isolating the zones of mineralisation at various scales and how
selective filtering can remove unwanted variation from the results.

3. Experimental data 4.1. Lithochemical log as a tessellation

The practical application of the wavelet tessellation is illu- Ca/(Feþ Si) is a useful ratio for distinguishing the narrow do-
strated here using geochemical data from the DET CRC Brukunga lerite dykes in the data, the values of the Ca ratio and the wavelet
Drilling Research and Training Facility, located at the old Brukunga tessellation are shown in Fig. 9A. Once the small variation ele-
sulphide mine in the Adelaide Hills, South Australia. Brukunga has ments are filtered from the tessellation (here achieved using a 50
been mined for iron sulphides (pyrite and pyrrhotite). The DET- percentile filter) the dolerite dykes stand out clearly from the
Brukunga02 diamond drill hole penetrated a sequence of meta- background, Fig. 9B. The rectangles in the tessellation can be col-
sediments (psammitic and pelitic) which have sulphide-rich our coded in order to provide further information: in Fig. 9A the
bands. The main sulphide ore body was intersected between 130 rectangles are coloured by wavelet coefficient; in Fig. 9B and C
and 280 m. The sulphide-rich sedimentary beds are characterized they are coloured by the mean and variation of the signal (over the
by high Fe and S concentrations in the chemical signals. The me- depth interval of the rectangle), respectively. Because the dykes
tasediments are intruded by narrow dolerite dykes, which are are quite small features we look for them in the small scales of the
distinguished in the chemical signal by high Ca content associated tessellation, at larger scales the rectangles represent groups of
with actinolite. Fine rock powders are a by-product of the drilling dykes. If the tessellation is compared to the geologist's log (Fig. 10)
and these were analysed using portable X-ray fluorescence spec- it is clear that the geologist had trouble identifying dolerite in
trometer (pXRF) and X-ray diffraction analysis, in addition the core regions with high sulphide content (
120–280 m); rocks with
was examined by a geologist and manually logged. high sulphide content are very dark and may be difficult to dis-
The elemental geochemical data was collected using a Delta tinguish visually from the mafic dolerites. Otherwise there is a
hand-held pXRF spectrometer with a large area SDD detector, in- very good correspondence between the logged dolerite and the
tegrated vacuum technology and customised 4 W X-ray tube narrow small scale rectangles with high mean value in the tes-
providing increased light-element sensitivity. The data were col- sellation. Fig. 9C shows that the host rocks to the dolerite have
lected in Soil and Mining Plus modes, which gave a range of ele- very low variance for the Ca ratio.
ments analysed from Mg to U. Prior to measurement of element The ratio S/(S þFeþAl) distinguishes the sulphide-rich from the
concentrations by pXRF, the instrument was calibrated using a sulphide-poor metasediments. As well as the main zone of sul-
metal disk. An instrument check was done prior to measurement phide-rich psammitic rocks (including porphyroblastic and lami-
and then after every 15 samples using standards NIST 2702, NIST nated varieties, Fig. 10) between 125 and 300 m, a narrow sul-
2781, NIST 2710a, NIST 2711a and a SiO2 blank (pure quartz). De- phide-rich psammitic layer at
90 m is clearly identifiable
tection limits were estimated on the basis of reproducibility of the (Fig. 9E, filtered using a 40 percentile filter). By considering the
measurements for certified reference materials (NIST standards). hierarchical structure of the tessellation, we can see that the main
band of sulphide mineralisation lies between 130 and 240 m (from
4. Results and discussion the rectangles on the right hand side of the tessellation). Moving
left on the tessellation plot, this zone splits into two smaller high
In this section we show how the tessellation provides a visual sulphide zones separated by
25 m of lower sulphide content.
summary of the data for two geochemical ratios selected as reliable Local variations in the sulphide content can be identified by
54 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57

Fig. 9. Tessellations for 2 geochemical ratios are shown in (A) and (D) coloured by wavelet coefficient. The tessellations are filtered to remove small changes and coloured by
mean value in (B) and (E) or by signal variance in (C) and (F). Depth is in metres.

moving further left on the plot. When the plot is coloured by signal (Fe þSi) and S/(S þFe þAl) ratios, respectively. This level of filtering
variance, it is clear that, except for the psammitic layer at
90 m, removes the fine details leaving a clearer picture of the hier-
there is only low to moderate variance within the sulphide units archical structure of the pseudo-lithological units identified by the
defined by the tessellation. geochemistry ratios. Depending on the level at which the data is to
Fig. 10 allows direct comparison between the lithochemical log be analysed, it may be desirable to use a stronger filter. For ex-
which resulted from classification using a ternary plot and that ample, if the purpose is for a building a regional 3D geology model
achieved using a filtered tessellation. This figure illustrates how a geologist would typically choose a high level grouping on the
the filtered wavelet tessellation, by taking spatial information into assumption that the higher level units will be more spatially
account, does not suffer from the spurious boundary problems continuous and hence more likely to result in successful correla-
illustrated in Fig. 1. tion of units between different drill holes. Fig. 11A shows the Ca/
(Fe þSi) ratio with 80 percentile filtering. At this level of filtering
4.2. Filtering the lithochemical tessellation the dolerites marked by strong peaks in the signal are still pre-
served, dolerites which have a weak signal disappear into the
The effect of filtering of the original tessellation is shown in background and dolerites which are weak but occur in sets are
Fig. 9B and E; these are filtered by 50 and 40 percentiles for Ca/ grouped into single entities.
 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57 55

Fig. 10. (A) Column plot of S/(S þ Feþ Al), warm colours represent higher values; (B) column plot of lithological classification from ternary plot of Fe–Al–S, small triangles
show locations illustrated in Fig. 1; (C) filtered tessellation of S/(S þFeþAl); (D) Geological log of the drill hole and (E) filtered tessellation of Ca/(FeþSi). Depth is in metres.

Evaluating the usefulness of the results of the filtered tessel- variance is useful information when deciding how far to filter the
lation is highly subjective. However, spatial skill dichotomies tessellation. The variance of a unit is also a useful measure when
mentioned in Section 1 (Shipley et al., 2013) can be used as a guide planning for selective sampling campaigns. For example, if a lim-
when evaluating how the filtering of the tessellation imitates the ited number of samples are required for testing geometallurgical
cognitive processes of a geologist. The three observations on do- properties then it is economical to take only one sample from each
lerite detection in the tessellation described in the last sentence of lithochemical domain; however, it would be prudent to take more
the previous paragraph could be considered to be imitating a than one sample from a domain that has a high variance.
geologist's habits as follows. Deciding which dolerite units to re-
tain and which to ignore at larger scales is a combining versus 4.3. Comparison to blocking algorithm
separating problem (Shipley et al., 2013). In this case, at the larger
scale the prominent Ca peaks remain separated from the back- The results from the filtered tessellation can be compared to
ground and the weak Ca peaks are combined with the background. results from a typical blocking algorithm, Fig. 12. A blocking al-
Deciding when to group dolerite units into a set is a single versus gorithm picks many of the same boundaries as the tessellation
many problem (Shipley et al., 2013). In this case, at the larger scale (depending on what noise threshold is used) but the boundaries
the filtering process combines adjacent small Ca peaks into a are compressed into a 1D plot, which means that the hierarchical
single more spatially continuous entity; which may be more useful information is lost. The removal of the space dimension in the
for regional correlation than individual peaks. blocking algorithm also means that the choice of threshold is
Fig. 11B shows the S/(S þ Feþ Al) ratio with 70 percentile fil- critical to the information preserved.
tering. In this example, the filtering has reduced the complexity of
the plot significantly. The main region of high sulphide forms a
single entity approximately 100 m wide, while the narrow band of 5. Summary
strong sulphide
90 m is preserved as a separate entity. When
coloured by variance of the signal, Fig. 11C, it is clear which of the The rectangular tessellation provides a succinct and intuitive
large entities encompass more variation than others. For example, visualisation of the structure of lithochemical domains in the drill
the top of the hole above the main sulphide zones (o 90 m) is hole sample data. Unlike blocking methods (the current method
homogeneously low; while the lower section of the hole used for geophysical data), a rectangular tessellation plot pre-
(4 245 m) below the main sulphide zone is more variable. This serves the hierarchical structure of the objects at all scales and can
56 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57

Fig. 11. Effects of strong filtering of the tessellation for the Ca/(Feþ Si) and S/(S þFeþ Al) ratios. (A) and (B) are coloured by mean signal value, (C) is coloured by signal
variance; darker colours represent higher values. Depth is in metres.

units than classification by considering individual samples without

spatial context.

Acknowledgements

The work has been supported by the Deep Exploration Tech-

nologies Cooperative Research Centre (DET CRC) whose activities
are funded by the Australian Government’s Cooperative Research
Centre Programme. This paper is DET CRC document no. 2014/610.
The authors would like to thank Professor David Giles for pro-
viding the core logging data and Anya Reading, Michael Gazley and
Heather Sheldon for suggestions for improving the manuscript.
The authors gratefully acknowledge James Cleverley and Louise
Fisher for discussions and inspiration which helped drive new
ideas for this research.

References

Arabjamaloei, R., Edalatkha, S., Jamshidi, E., Nabaei, M., Beidokhti, M., Azad, M.,
2011. Exact lithologic boundary detection based on wavelet transform analysis
and real-time investigation of facies discontinuities using drilling data. Pet. Sci.
Technol. 29, 569–578.
Canny, J., 1986. A computational approach to edge detection. IEEE Trans. Pattern
Anal. Mach. Intell. 8, 679–698.
Davis, A., Christensen, N., 2013. Derivative analysis for layer selection of geophysical
Fig. 12. Comparison between the information preserved in (A) a tessellation and borehole logs. Comput. Geosci. 60, 34–40.
(B) a conventional blocking algorithm of Fe composition. Cooper, G., Cowan, D., 2009. Blocking geophysical borehole log data using the
continuous wavelet transform. Explor. Geophys. 40, 233–236.
Chandrasekhar, E., Eswara Rao, V., 2012. Wavelet analysis of geophysical well-log
data of Bombay Offshore Basin. India. Math. Geosci. 44, 901–928.
be filtered to remove objects which are represented by relatively Donoho, D.L., Johnstone, I.M., 1994. Ideal spatial adaptation by wavelet shrinkage.
weak signals. The result of the tessellation is a high level summary Biometrika 81, 425–455.
Hillis, R.R., Giles, D., van der Wielen, S.E., Baensch, A., Cleverley, J.C., Fabris, A.,
of the geochemical information (or other numerical data), which Halley, S.W., Harris, B.D., Hill, S.M., Kanck, P.A., Kepic, A., Soe, S.P., Stewart, G.,
provides a less problematic basis for classification of lithochemical Uvarova, Y., 2014. Coiled tube drilling and real-time sensing–enabling
 E.J. Hill et al. / Computers & Geosciences 79 (2015) 47–57 57

prospecting drilling in the 21st century. Soc. Econ. Geol. Spec. Publ. 18, Lett. 62, 53–62.
243–259. Panda, M.N., Mosher, C.C., Chopra, A.K., 2000. Application of wavelet transforms to
Hoyle, I.B., 1986. Computer techniques for the zoning and correlation of well-logs. reservoir-data analysis and scaling. Soc. Pet. Eng. J. 5, 92–101.
Geophys. Prospect. 34, 648–664. Perez-Munoz, T., Velasco-Hernandez, J., Hernandez-Martinez, E., 2013. Wavelet
Mallat, S., 1991. Zero-crossings of a wavelet transform. IEEE Trans. Inf. Theory 37, transform analysis for lithological characteristics identification in siliciclastic oil
1019–1033. fields. J. Appl. Geophys. 98, 298–308.
Mallat, S., 2009. A Wavelet Tour of Signal Processing, third ed. Elsevier, Burlington, Shipley, T.F., Tikoff, B., Ormand, C., Manduca, C., 2013. Structural geology practice
MA, pp. 205–261, Chapter 6 – Wavelet Zoom. and learning, from the perspective of cognitive science. J. Struct. Geol. 54,
Mallat, S., Hwang, W.L., 1992. Singularity detection and processing with wavelets. 72–84.
IEEE Trans. Inf. Theory 38, 617–643. Torrence, C., Compo, G.P., 1998. A practical guide to wavelet analysis. Bull. Am.
Mallat, S., Zhong, S., 1992. Characterisation of signals from multi-scale edges. IEEE Meteorol. Soc. 79, 61–78.
Trans. Pattern Anal. Mach. Intell. 14, 710–732. Witkin, A.P., 1983. Scale–space filtering. In: Proceedings of 8th International Joint
Marr, D., Hildreth, E., 1980. Theory of edge detection. Proc. R. Soc. Lond. 207, Conferences on Artificial Intelligence, Karlsruhe, Germany, pp. 1019–1022.
187–217.
Mullen, E.D., 1983. MnO/TiO2/P2O5: a minor element discriminant for basaltic rocks
of oceanic environments and its implications for petrogenesis. Earth Planet. Sci.