See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/283185832

EVALUATION OF WAVELET DENOISING METHODS FOR SMALL-SCALE JOINT

ROUGHNESS ESTIMATION USING TERRESTRIAL LASER SCANNING

Article · August 2015

DOI: 10.5194/isprsannals-II-3-W5-81-2015

CITATIONS READS

8 285

3 authors:

Maja Bitenc D. Scott Kieffer

Graz University of Technology Graz University of Technology
12 PUBLICATIONS   84 CITATIONS    47 PUBLICATIONS   373 CITATIONS   

SEE PROFILE SEE PROFILE

Kourosh Khoshelham
University of Melbourne
172 PUBLICATIONS   3,032 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Smartphone-based Indoor Localization and Navigation View project

Deep learning-based activity recognition for indoor localization View project

All content following this page was uploaded by Maja Bitenc on 23 February 2016.

The user has requested enhancement of the downloaded file.

 EVALUATION OF WAVELET DENOISING METHODS FOR SMALL-SCALE JOINT
ROUGHNESS ESTIMATION USING TERRESTRIAL LASER SCANNING

M. Bitenc a,* , D. S. Kieffer a, K. Khoshelhamb

a
Institute of Applied Geosciences, Graz University of Technology, Austria - (bitenc, kieffer)@tugraz.at
b
Department of Infrastructure Engineering, The University of Melbourne, Victoria 3010, Australia – [email protected]

Commission VI, WG VI/4

KEY WORDS: terrestrial laser scanning, joint roughness, range noise, discrete wavelet transform, stationary wavelet transform,
denoising performance

ABSTRACT:

The precision of Terrestrial Laser Scanning (TLS) data depends mainly on the inherent random range error, which hinders extraction
of small details from TLS measurements. New post processing algorithms have been developed that reduce or eliminate the noise
and therefore enable modelling details at a smaller scale than one would traditionally expect. The aim of this research is to find the
optimum denoising method such that the corrected TLS data provides a reliable estimation of small-scale rock joint roughness. Two
wavelet-based denoising methods are considered, namely Discrete Wavelet Transform (DWT) and Stationary Wavelet Transform
(SWT), in combination with different thresholding procedures. The question is, which technique provides a more accurate roughness
estimates considering (i) wavelet transform (SWT or DWT), (ii) thresholding method (fixed-form or penalised low) and (iii)
thresholding mode (soft or hard). The performance of denoising methods is tested by two analyses, namely method noise and method
sensitivity to noise. The reference data are precise Advanced TOpometric Sensor (ATOS) measurements obtained on 20×30 cm rock
joint sample, which are for the second analysis corrupted by different levels of noise. With such a controlled noise level experiments
it is possible to evaluate the methods’ performance for different amounts of noise, which might be present in TLS data. Qualitative
visual checks of denoised surfaces and quantitative parameters such as grid height and roughness are considered in a comparative
analysis of denoising methods. Results indicate that the preferred method for realistic roughness estimation is DWT with penalised
low hard thresholding.

1. INTRODUCTION Due to historical reasons, including limited amount of 1D or 2D

surface measurements and low computational power, roughness
In the last few decades Terrestrial Laser Scanning (TLS) is is traditionally parameterized by a single value as e.g. Joint
increasingly used in the field of engineering geology for a Roughness Coefficient (JRC) (Barton, 1973) or dilation angle
growing array of applications (Fowler, 2011; Fernández-Steeger (Patton, 1966). Taking profile measurements rather than
et al., 2009). Because of remote and fast acquisition that results considering 3D surface topography can lead to biased estimates
in relative high resolution and accuracy data, TLS is valuable of roughness as for example explained in (Rasouli and Harrison,
for in-situ rock mass characterization (Tonon and Kottenstette, 2004). Therefore newer studies took advantage of available 3D
2006); especially for estimating geometrical parameters (Slob, remote sensing data and developed new roughness parameters
2010). One of them is rock joint roughness, which significantly such as fractals (e.g. Fardin, 2004) or the angular threshold
influences rock mass stability, and for this reason needs to be method (Grasselli, 2001). Nevertheless, those parameters have
precisely estimated for each engineering project. limited ability to describe the three attributes of roughness,
namely amplitude, direction and scale dependency. A
Rock joint roughness describes the morphology of joint surfaces combination of precise and dense 3D measurements of in-situ
(roughness amplitude) in certain direction and at certain scales. large scale rock joints and truly 3D roughness parameters would
Roughness direction should be consistent with expected rock improve rock mass stability analysis and better support
mass movements (shear direction). When the shear direction is engineering geologist decisions. In our research the efficacy of
not known a-priori, roughness should be measured in all TLS as a means for in-situ characterization of joint surfaces is
possible directions. Roughness scale dependency was attributed investigated.
to sample size (Barton and Choubey, 1977), but the studies,
which are summarized in (Tatone and Grasselli, 2012), show Sturzenegger and Stead (2009) stated that TLS technology and
contradictory results. In case of large discontinuities, the joint careful fieldwork allow the extraction of first-order roughness
roughness consists of large-scale (waviness or primary profiles, but for the secondary roughness estimation the raw
roughness) and small-scale (unevenness or secondary TLS point clouds do not fully satisfy requirements for data
roughness) components (ISRM, 1978). Waviness can be resolution and accuracy. TLS data resolution refers to the ability
described as a large-amplitude and low-frequency signal. It has to resolve two objects on adjacent sight lines and depends on
been defined by (Priest, 1993) as “surface irregularities with a sampling interval (nominal point spacing) and laser beamwidth
wavelength greater than about 10 cm”. Unevenness can be (footprint size) (Lichti and Jamtsho, 2006). The smallest
described as a small-amplitude and high-frequency signal. It resolution is therefore defined by laser scanner capability to
covers finer scales of 5 to 10 cm and is superimposed on the steer the laser beam for small angular increments and to focus
waviness. the laser beam in a small footprint. Because the footprint size
depends on scanning geometry (the range and incidence angle),
it is generally advised to scan a surface as close as possible and (2001), appeared to be least sensitive to measurement random
optimally in perpendicular direction. The resulting (effective) errors.
TLS data resolution, which was studied in (Lichti and Jamtsho,
2006; Pesci et al., 2011), defines the smallest size of observable The aim of this research is to find a wavelet based denoising
feature, this is the smallest roughness scale. TLS data accuracy procedure that is most suitable for reliable estimation of rock
depends on several factors: (i) imprecision of laser scanner joint roughness. Roughness is characterized in 2D space (as
mechanism, (ii) geometric properties of scanned surface, this is surface) by Grasselli parameter. The paper is organized as
scanning geometry, (iii) physical properties of scanned surface follows: in Section 2 an overview of applied wavelet transform
material and (iv) environmental (atmospheric) conditions methods and thresholding procedures is given. In Section 3
(Soudarissanane et al., 2011). Resulting measurement errors are computational background of the Grasselli roughness parameter
composed of systematic and random errors. Systematic errors is presented. Section 4 describes our experiment of testing
are usually removed by a proper calibration procedure (Lichti, wavelet-denoising procedures and presents results of
2007). The remaining random errors, which are attributed comparison analysis. Discussion of results and concluding
mainly to range error and are therefore referred to as range remarks with an outlook are given in Section 5.
noise, are in order of a few millimetres. An appropriate
denoising method would improve capabilities of TLS for 2. WAVELET BASED DENOISING
modelling small-scale details.
Denoising by discrete wavelet transform (DWT) has its origin
Advances in computer technology have enabled development of in research of Donoho (1995). The data are processed at
fast and efficient post-processing algorithms that reduce or different scales (levels) or resolutions, which enable us to see
ideally eliminate noise in TLS data and at the same time general surface morphology trend as well as details (or a more
preserve surface details. Existing methods to estimate and illustrative example: to see both the forest and the trees
eliminate TLS range noise include empirical methods, where together). Since noise is characterized by high frequency
reference data of higher accuracy or a known reference model fluctuations, it is more likely (compared to other denoising
are used, interpolation methods using simple averaging or more techniques) that thresholding high frequency components of
elaborate smoothing functions, and theoretical methods based DWT reduces noise and preserves low frequency components
on error propagation models. Interpolation methods are that present general trend. For this reason DWT is considered as
commonly used since no additional data or assumptions are an interesting and useful tool for denoising.
needed with respect to reference data or a theoretical model.
Inputs for those denoising methods are either 3D point clouds or Figure 1 shows a general DWT denoising procedure with three
2.5D surfaces. A short list of 3D point cloud or mesh methods is main steps: (i) decomposition, (ii) thresholding detail
given in (Smigiel et al., 2008). Commonly the complexity of a coefficients and (iii) reconstruction. In the first step, surface is
3D randomly scattered point cloud is reduced by gridding. decomposed into several levels of approximation (general trend)
Having Cartesian coordinates (x,y,z), a grid of certain cell size and detail coefficients. Decisions have to be made about an
is constructed in the chosen two directions and the third appropriate DWT method, a mother wavelet and number of
coordinate is assigned to each grid point. Another approach for decomposition levels (N). Denoising process (called also
obtaining a 2.5D surface was used by Smigiel et al. (2008). wavelet shrinkage) rejects noise by thresholding in the wavelet
They transformed Cartesian coordinates to polar coordinates domain, which is the second step. For thresholding detail
(horizontal and vertical angle, and range), which are actually coefficients one must decide about thresholding method and
TLS original measurements, and computed the so-called range thresholding mode. The choice eventually involves a trade-off
image. An advantage of 2.5D surface is that a whole range of between keeping a bit of noise in the data and removing a bit of
existing image processing algorithms can be used. An overview actual surface details. Finally, the denoised surface is
of image denoising methods and further references can be found reconstructed using approximation coefficients of the last level
in (Buades et al., 2005; Smigiel et al., 2011; Zhang et al. 2014). (N) and thresholded detail coefficients of all levels (1 - N).

An objective of image denoising methods is to decompose

smooth and oscillatory part. Therefore, image filtering
techniques result in data smoothing (Heckbert, 1989). One of
frequency domain filtering method is discrete wavelet transform
(DWT), which represent signals with a high degree of sparsity
(scale-based components). Since denoising is performed at
different scales or resolutions, it potentially removes just noise
and preserves the signal characteristics, regardless of its Figure 1. General DWT denoising procedure.
frequency content. The DWT was successfully applied on TLS
data in (Smigiel et al., 2008) and (Khoshelham et al., 2011). In In the following section the basic principles of multi-level
first research the aim was to improve a model of a one-cubic- wavelet transform are briefly described. A more detailed and
meter big object (Corinthian capitals). In general the denoised comprehensive explanation of DWT can be found in numerous
model appeared smoother, but the sharp edges were preserved. publications, e.g. (Daubechies, 1992; Fugal, 2009). The
In the second research denoised data were used to compute the thresholding process with its variants, threshold methods and
roughness of a rock surface profile. 1D wavelet denoising threshold modes, is also presented.
resulted in more realistic estimates of the roughness; however
the fractal parameters, which were used to describe the 2.1 Multi-level discrete wavelet transform
roughness, were found to be too sensitive to noise. Therefore, in
our previous study (Bitenc, 2015) different methods were tested In general a large number of wavelet transforms (WT) exist,
considering noise sensitivity. The angular threshold method, each suitable for different applications. The two main types are
hereafter referred to as the Grasselli parameter after Grasselli Continuous Wavelet Transform (CWT) and Discrete Wavelet
Transform (DWT). The multi-level wavelet transform
decomposes the surface into different scales (levels) with developed, namely non-decimated or Undecimated Wavelet
different space and frequency resolution by translating (shifting) Transform (UDWT). The UDWT requires more storage space
and dilating a single function, the mother wavelet. Thus, the than DWT and is computationally intensive, but is shift
power of wavelets is its combined space-frequency invariant, has no aliasing and provides more precise information
representation of a surface. for the frequency localization. The principles of conventional
(decimated) DWT and undecimated DWT are compared, for the
At wavelet decomposition of a surface a resemblance or 1D case and for three levels, in Figure 2. By DWT the signal is
correlation index coefficients between the surface profile and downsampled by two, in contrast to UDWT, where the signal is
the wavelet are computed. The indices are called wavelet left unchanged and filters are upsampled by two on each
coefficients and are denoted by family C(a,b), where a is a scale consecutive level (Fugal, 2009). UDWT has been shown to
and b is a shift. CWT can operate at every scale and is perform better at image denoising (Gyaourova, 2002).
continuous in terms of shifting, but is computationally
demanding, extremely redundant and does not enable exact
reconstruction. Therefore, the DWT was developed, which
typically takes scales and translations based on powers of two
(in a dyadic grid). It is easier implemented in the discrete
computer environment and, with a careful choice of wavelets or
bandpass filters, enables a perfect reconstruction. This basic
concept of DWT is called subband coding (filtering). Two- Figure 2. Downsampling the signal in case of DWT (left) and
channel subband coding is an efficient algorithm that is upsampling the wavelet in case of UDWT (right) (Fugal, 2009).
performed in two steps: convolution (filtering) and
downsampling. In first step the original surface passes through Different UDWT algorithms have been developed that differ in
two complementary filters (high- and low-pass), and emerges as decomposition and/or reconstruction method, but are all shift
two signals of halved frequency resolution. Resulting surfaces invariant. To mention the most important and frequently used
hold redundant information, therefore the second step is algorithms: algorithm `a trous (Holschneider et al., 1990),
introduced – downsampling. Without loss of information, half Beylkin’s algorithm (Beylkin, 1992), undecimated WT (Mallat,
of the sample is discarded, which changes the surface size (i.e 1991), stationary WT (Nason and Silverman, 1995) and Cycle-
resolution) while doubling the scale. Spinning (Coifman and Donoho, 1995). The last algorithm is
implemented in Matlab, where it is called the ε-decimated DWT
For a multi-level DWT, the decision has to be made about and denoted by SWT. It averages some slightly different DWTs
mother wavelet and number of levels. To choose the wavelet, to define stationary wavelet transform.
two options (matched filtering) are to correlate finite wavelets
(i) with the transient surface or (ii) with the feature of interest, 2.2 Thresholding procedure
e.g. noise. In the first case, the surface is extracted from noise
and in the second case the feature (e.g. noise) is extracted and 2.2.1 Thresholding method. Success of DWT denoising,
then subtracted from the signal. The choice regarding suitable this is to reduce noise while preserving details of the original
number of levels is based on (i) the nature of the signal (e.g. surface, depends on the threshold limit. Different threshold
signal to noise ratio), (ii) a suitable criteria (e.g. entropy), (iii) a methods have been proposed sharing a common approach,
desired low-pass cut off frequency or (iv) precedent. In our which is, to estimate the standard deviation of noise and
research the wavelet was chosen based on visual inspection, as multiply it by a certain value. The threshold can be applied
the best match with the surface shape, and the decomposition globally or locally. In the global case one single value is applied
level was determined by the surface size. Considering those to all detail coefficients and in the local case a different
criteria, our experience indicates that the choice of wavelet (e.g. threshold value is chosen for each wavelet level. The
db3 or db4) and number of levels (e.g. 4 or 5) influence on the thresholding method is selected based on data characteristics
shape of components, but have negligible effects on denoising and overall noise distribution. In this research two thresholding
results. A possible reason is that the thresholds are computed methods are compared, namely the fixed form and the
each time from detailed coefficients, thus denoising results in penalized.
similar output. Therefore, the effects of wavelets and number of
levels are not further studied here. If the signal-to-noise ratio is small, the fixed-form (universal)
threshold proposed by Donoho and Johnstone (1994) is used.
Reconstruction or synthesis is an inverse discrete wavelet For 2D space, the fixed-form threshold TF is defined as:
transform procedure, where the components are assembled back
into the original surface with no loss of information. 𝑇𝐹 = 𝜎 ∗ √2 ∗ 𝑙𝑜𝑔(𝑟 ∗ 𝑐) (1)
Reconstruction consists of upsampling and filtering.
Upsampling is a process of lengthening a surface component by Where σ is the noise standard deviation and [r,c] is the surface
inserting zeros or interpolated values between samples. The grid size. If the noise level is not known a-priori, σ is computed
choice of filters is crucial to achieve perfect reconstruction of as the Median Absolute Deviation (MAD) of detail coefficients
the original surface, i.e. to cancel out the effects of aliasing at the first decomposition level (global threshold) or at each
made in the decomposition phase. Filters for the decomposition decomposition level (local threshold).
and reconstruction phases are closely related (but not identical).
They form a system of quadrature mirror filters. A variant of the fixed form strategy of the wavelet shrinkage is
the penalized thresholding method based on denoising results
A disadvantage of DWT downsampling is that the transform is by Birgé and Massart (1997). The penalized threshold TP is
not shift invariant. This means that the DWT of the translated defined as:
and original surfaces are not the same. Since shift-invariance is
important for many applications such as change detection, 𝑇𝑃 = |𝑐(𝑡 ∗ )| (2)
denoising and pattern recognition, a new type of DWT was
Where t* is the penalty function to be minimized and equals: Based on curve fitting and regression analysis, Grasselli
proposed an empirical equation to express potential contact area
𝑡 ∗ = 𝑎𝑟𝑔𝑚𝑖𝑛 [− ∑ 𝑐 2 (𝑘), 𝑘 < 1 + 2𝑡𝜎 2 (𝛼 + 𝑙𝑜𝑔(𝑛⁄𝑡))] (3) ratio Aθ∗ as a function of θ∗ :
𝑡 = 1, … , 𝑛
Where α is the sparsity parameter, coefficients c(k) are sorted in Aθ∗ = Ao ∗ ((θ∗max − θ∗ )⁄θ∗max )C (4)
decreasing order of their absolute value, and σ2 is the noise
variance computed as explained above (for fixed-form). Three Where Ao is the maximum possible contact area in the shear
intervals for the sparsity parameter are proposed, namely direction (when θ∗=0°), which is usually around 50% of the
penalized low 0 < α <1.5, penalized medium 1.5 < α <2.5 and total surface area for fresh mated joint, θ∗max is the maximum
penalized high 2.5 < α <10. apparent dip angle of the surface in the shear (analysis)
direction and θ∗ is the threshold inclination, this is the minimum
2.2.2 Thresholding mode. The selected threshold can be apparent dip angle for applied normal load σn.
applied to a surface in either a soft or hard thresholding mode.
For a case of surface profile, the principles as well as effects of C is an empirical fitting parameter calculated via a non-linear
thresholding modes are shown in Figure 3. In the hard mode, least-squares regression that characterizes the shape of the
coefficients (black solid line) that are in absolute value lower cumulative distribution. Surface parameters Aθ∗ , C and θ∗max
than a threshold value (black dashed lines) are simply set to depend on the specific shear direction and 3D surface
zero. In the soft mode, additional coefficients that are above the representation. A higher proportion of steeply inclined
threshold value are reduced for the threshold value. Therefore, triangulated areas is indicative of a rougher surface, and is
soft thresholding results in a smoother profile (blue dotted line) reflected by a larger area under the curve given by Eq. 4. This
and hard thresholding introduces profile discontinuities (red area, assuming Ao is 0.5, is taken as the roughness parameter R
dashed line). Hard thresholding provides an improved signal to (Tatone and Grasselli, 2012) and is computed by:
noise ratio (Shim et al., 2001). In terms of image denoising, the
soft thresholding is preferred, because it makes algorithms R = θ∗max ⁄(C + 1) (5)
mathematically more tractable and avoids artificial and false
structures (Donoho, 1995). 4. EXPERIMENTS AND RESULTS

Experiments were executed on ATOS measurements which

have much higher precision than any TLS data. Therefore
ATOS data are taken as a reference noise-free input data into
performance analyses of denoising methods. The objective of
this noise-controlled test is to make solid and firm guidelines
for denoising TLS data acquired by Riegl VZ400 laser scanner
(Riegl, 2015).

In the following section a detailed description of experimental

workflow is followed by presentation of results, including the
Figure 3. The principle of soft and hard thresholding.
threshold values, the method noise, and the method sensitivity
to noise.
3. ROUGHNESS PARAMETER
4.1 Experimental workflow
The roughness parameter introduced by Grasselli (2001) is
based on the angular threshold concept and was initially Input data for experiments were acquired with highly accurate
developed to identify potential contact areas during direct shear ATOS I measurement system (Capture3D, 2013). A rock joint
testing of artificial rock joint. Highly accurate and detailed formed in fossiliferous limestone having dimensions of
ATOS (Advanced TOpometric Sensor) measurements were 20×30 cm was imaged in the laboratory at approximately 0.5 m
used to reconstruct (triangulate) the surface of the rock joint. distance. On average, point density was 15 points per square
Based on joint surface damage patterns, it was found that only millimetre. Further processing was performed in Matlab. First,
those areas of the joint surface that face the shear direction and the acquired dense point cloud was linearly interpolated into
are steeper than a threshold inclination θ∗ provide shear 1 mm grid, hereafter referred to original ATOS data (see Figure
resistance. The sum of triangulated areas that are steeper than θ∗ 9, top). Z-direction corresponds to roughness amplitude.
(denoted as A θ∗ and referred to as the total potential contact area
ratio) is plotted against θ∗ (Figure 4). All together 12 wavelet denoising procedures were tested,
which are a combination of:
 Two 2D wavelet transform methods: DWT and SWT.
 Three thresholding methods: fixed-form global, fixed-
form local and penalised low.
 Two thresholding modes: soft and hard.

DWT and SWT transforms were executed on four levels using a

general purpose Daubechies wavelet db3 that has three
vanishing moments and the filter length of 6 points. Reasons for
the choice are given already in Section 2.
Figure 4. Cumulative distribution of potential contact area ratio
Aθ∗ having a minimum inclination θ∗ [ᵒ], which is computed Threshold values are computed with equations given in
with respect to average joint plane and analysis direction Paragraph 2.2.1. Sigma for fixed-form global threshold equals
(Grasselli, 2001). the known standard deviation of added noise (values are written
in the following paragraph). Standard deviation for fixed-form definition of our computation, the same for DWT and SWT.
local threshold is re-computed for each decomposition level as The other two SWT thresholds (fixed-form local and penalised
MAD of detail coefficients of the level. Penalised low threshold low) are compared to corresponding DWT’s thresholds higher
is a global value computed by sigma equal to MAD of detail and increase slower with the noise.
coefficients of the first level and alpha of 1.5. The three
threshold values are applied on detail coefficients in the two
modes: soft and hard. Six different threshold combinations are
denoted on the following figures as given in Table 1.

Denotation Threshold combination

FGS Fixed-form Global threshold, Soft mode
FGH Fixed-form Global threshold, Hard mode
FLS Fixed-form Local threshold, Soft mode
FLH Fixed-form Local threshold, Hard mode
PLGS Penalised Low Global threshold, Soft mode
PLGH Penalised Low Global threshold, Hard mode
Table 1. Denotation of threshold values and modes used in the
following figures.

Two analyses were performed in order to justify suitability of Figure 5. Fixed-form global and local, and penalised low global
the wavelet denoising methods for rock joint roughness threshold values for the DWT and SWT decomposition.
estimation. First, the noise of a denoising method itself is
analysed (method noise analysis) by applying denoising 4.3 Method noise
methods on original ATOS data. If assuming that ATOS data
have no or very little noise and that a denoising method Method noise analysis is performed on the original ATOS
removes just noise and not also details, the output (denoised) surface. Figure 6 shows mean and standard deviation of the two
surface should match the input surface. Second, original ATOS performance measures, ΔZ (left plot) and ΔR (right plot), versus
data is corrupted with different levels of Gaussian white noise denoising methods. Mean of ΔZ is as expected zero for all
and the method sensitivity to noise is studied as the noise denoising methods, however the standard deviation, which
increases. The aim of this controlled noise level experiment is to indicates noise produced by denoising method, ranges from
study dependence of thresholds and denoising method approximately 0.1 mm for DWT penalised low hard
performance on amount of noise. Noise levels were chosen thresholding (left plot, red mark at PLGH) to 0.6 mm for SWT
based on empirical noise estimation for the Riegl VZ400 laser fixed-form local soft thresholding (left plot, blue mark at FLS).
scanner in (Vezočnik, 2011). His experiment showed that noise Similar pattern can be observed in error plot of ΔR, where the
on concrete surfaces reaches maximum 2.2 mm for scanning DWT penalised low hard thresholding surface shows the
distances up to 65 m and incidence angles up to 60ᵒ. Therefore smallest mean roughness difference -0.3ᵒ (right plot, red mark at
in our experiment five noise levels were chosen, namely 0.5, 1, PLHG) compared to biggest mean difference of -9ᵒ by SWT
1.5, 2 and 2.5 mm. Noise was added to grid points of original fixed-form local soft thresholding (right plot, blue mark at
ATOS data and those noisy surfaces entered the denoising FLS). In general, considering the wavelet transform and
procedure. As an example, surface of 2 mm added-noise is thresholding mode, method noise is lower in case of DWT and
shown in Figure 9, middle. hard thresholding.

The Grasselli parameter was computed for the 12 denoised

surfaces and original surface, as explained above. The analysis
direction changes clockwise from 0ᵒ (+Y-axis direction) to 355ᵒ
in 5ᵒ steps.

A comparative study of denoised surfaces and the original

surface is based on three performance measures: (1) qualitative
visual check of denoised surfaces, (2) height differences ΔZ and
(3) differences of roughness parameters ΔR. Height differences
ΔZ are computed as Δ𝑍 = 𝑍𝑑𝑒𝑛𝑜𝑖𝑠𝑒𝑑 − 𝑍𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑙 for each grid Figure 6. Error plot of ΔZ (left) and ΔR (right) for 12 denoising
point and roughness differences as Δ𝑅 = 𝑅𝑑𝑒𝑛𝑜𝑖𝑠𝑒𝑑 − 𝑅𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑙 methods.
for each analysis (shear) direction.
4.4 Method sensitivity to noise
4.2 Threshold values
Adding the five levels of noise to ATOS grid, the height
Threshold values computed for decomposed original (0 mm differences ΔZ and roughness parameter differences ΔR were
noise) and decomposed noisy (0.5 – 2.5 mm noise) surfaces are computed. Mean ΔZ is very close to zero for all 12 denoised
shown in Figure 5. All values linearly increase with noise level. surfaces (as seen already in Figure 6). Standard deviation 𝜎Δ𝑍 ,
Fixed-form local thresholds for DWT and SWT are presented which shows how much noise is left in denoised surface,
by four lines – one for each decomposition level. Lines are increases with the noise level and is bigger in case of soft
following each other with the level number; from the upper line
thresholding for both, DWT and SWT transforms (see Figure
(higher threshold) of first level to the lower line (smaller
7). Figure 7 shows that the noise is greatly reduced, except in
threshold) of fourth level. Thus, threshold value decreases with case of 0.5 mm noise level, when fixed form global and local
decomposition level. Fixed-form global threshold values are, by soft threshold were applied for DWT (left plot, blue and red
line) and all three soft thresholds for SWT (right plot, red, blue
and black line).

Figure 7. Standard deviation 𝜎Δ𝑍 versus noise levels for DWT

(left) and SWT (right) denoised surfaces applying six
thresholds.

Figure 8 shows distribution of roughness parameter differences

ΔR versus noise level. Mean ΔR decreases (i.e. roughness
estimation error ΔR increases), whereas the standard deviation
of ΔR does not show clear trend. Surfaces contaminated with
higher level of noise become smoother after wavelet denoising.
Decrease of roughness with increase of noise can be explained
with higher threshold values (Figure 5), which filter more
surface details. An exception is SWT penalised low hard
method, which does not show a clear trend.

Figure 9. Visual comparison of 1 mm grid surfaces: original

ATOS (top), noisy surface with 2 mm added noise (middle) and
same surface denoised by DWT, penalised low hard
Figure 8. Error plot of ΔR versus noise levels for DWT (left) thresholding (bottom).
and SWT (right) denoised surfaces applying six thresholds.
5.1 Concluding remarks
Comparing DWT and SWT results for the common fixed-form
global threshold (in Figure 8 – blue and turquoise line for soft DWT versus SWT: For the investigated thresholding
and hard thresholding mode, respectively), SWT denoising procedures and for the five noise levels, DWT provides better
results in much underestimated roughness. results considering our application; this is for reliable rock joint
roughness estimation. SWT smoothens surfaces more than
Hard thresholding performs better than soft thresholding since DWT. The first reason is the underlying computation of SWT,
standard deviation of ΔZ is lower and absolute values of ΔR are which averages many different DWTs. Secondly, the same
smaller. However, the visual check of denoised surfaces shows thresholding methods (fixed-form local and penalised low)
that hard thresholding especially in combination with result in higher SWT thresholds. DWT is preferable also
conservative penalised-low thresholding results in a number of because MAD of first level DWT detail coefficients equals the
spikes (see Figure 9, bottom). Spikes are left overs of unfiltered standard deviation of noise.
noise. Spikes appear in DWT and SWT denoised surfaces, and
become higher with increasing noise level, as shown especially Fixed-form versus penalised threshold: It is known that the
for penalised low hard thresholding. penalised threshold is more conservative than fixed-form. This
means that the penalised threshold retains more coefficients and
5. DISCUSSION AND CONCLUSIONS the fixed form threshold removes noise more efficiently. In our
experiments the penalised low threshold results in noise spikes
Experiments on the simulated noisy ATOS data show positive in denoised surfaces (as reported also in Coifman and Donoho,
results: by applying a wavelet-based denoising procedure noise 1995), and the fixed-form filters out surface details and returns
is successfully reduced and roughness estimates are much closer lower roughness values.
to the reference values. All investigated denoising methods
significantly reduce standard deviation of height differences. Soft versus hard mode: Soft thresholding, compared to hard
Denoised surfaces contain less than 1 mm noise. Roughness is thresholding, smoothens the surface, since unfiltered
on average (for different noise levels and denoising methods) coefficients are reduced for the threshold. This phenomena
underestimated, which means that in addition to noise, some appears unfavourable for rock surface roughness estimation. On
surface details are also removed.
the other hand, soft thresholded surfaces look smoother and Volume 6. Lollino, G., Giordan, D., Thuro, K., Carranza-
more natural (without artificial spikes). Torres, C., Wu, F., Marinos, P., Delgado, C. (eds.), pp. 835–38.
Springer International Publishing.
Local versus global threshold: Local thresholds decrease with
level, which agrees with the assumption that higher levels of Buades, A., Coll, B., Morel, J., 2005. A Non-Local Algorithm
lower resolution contain less noise. On the contrary, global for Image Denoising. In IEEE Computer Society Conference on
thresholds computed from first level detail coefficients remains Computer Vision and Pattern Recognition, vol. 2, pp. 60–65.
constant and therefore removes more actual details on higher
levels. This is indicated by results of DWT denoising, where Capture3D, 2013. Atos I, Configurations.
fixed-form local thresholds provide more realistic roughness http://www.capture3d.com/products-ATOSI-configuration.html.
estimation than the fixed-form global threshold.
Coifman, R. R., Donoho, D. L., 1995. Translation-Invariant de-
In conclusion, the suggested denoising method is DWT Noising. Springer New York.
penalised low hard local thresholding. Having smaller mean and
standard deviation of roughness differences is preferred over Daubechies, I., 1992. Ten Lectures on Wavelets. CBMS-NSF
visually smoother surface. Regional Conference Series in Applied Mathematics. Society
for Industrial and Applied Mathematics.
5.2 Outlook
Donoho, D. L., Johnstone, J. M., 1994. Ideal Spatial Adaptation
Results of this research show that inherent TLS range noise can by Wavelet Shrinkage. Biometrika, vol. 81 (3), pp. 425–55.
potentially be removed by post- processing using wavelet based
denoising. However, further research is needed to support the Donoho, D. L., 1995. De-Noising by Soft-Thresholding.
proposed method for arbitrary scanning geometry and rock joint Information Theory, IEEE Transactions on Information Theory,
characteristics. vol. 41 (3), pp. 613–27.

In our experiment we assumed that the rock surface is scanned Fardin, N., Stephansson, O., Feng, Q., 2004. Application of a
in the perpendicular direction; thus z-direction (in which New in Situ 3D Laser Scanner to Study the Scale Effect on the
denoising is performed) corresponds to range measurement Rock Joint Surface Roughness. International Journal of Rock
direction. In the case of non-perpendicular acquisition, Mechanics and Mining Sciences, vol. 41 (2), pp. 329–35.
denoising in polar coordinates space (range image) would be
needed. Fernández-Steeger, T., Wiatr, T., Azzam, R., 2009. Terrestrial
Laser Scanning in Engineering Geology. In 17. Tagung Für
Beside wavelet transform, other image denoising methods exist, Ingenieurgeologie Und Forum „Junge Ingenieurgeologen.
for example Non-Local Mean (Buades et al., 2005). A short trial
showed promising results. However, an elaborated investigation Fowler, A., France, J. I., Truong, M., 2011. Applications of
is needed to find the optimum input parameters (patch and Advanced Laser Scanning Technology in Geology. Riegl USA.
search window size) for rock surface roughness estimation.
Fugal, D. L., 2009. Conceptual Wavelets in Digital Signal
ACKNOWLEDGEMENTS Processing. San Diego, Calif: Space & Signals Technical
Publishing.
The Slovenian National Building Institute and Civil
Engineering Institute enabled data acquisition with ATOS Grasselli, G., 2001. Shear Strength of Rock Joints Based on
measuring system. Klemen Kregar from University of Ljubljana Quantified Surface Description. PhD thesis. Lausanne, EPFL.
processed TLS reference targets’ centres with his image
matching algorithm. Gyaourova, A., Chandrika, K., Imola K. F., 2002. Undecimated
Wavelet Transforms for Image de-Noising. Report, Lawrence
REFERENCES Livermore National Lab., CA 18.
Barton, N., Choubey, V., 1977. The Shear Strength of Rock Heckbert, P. S., 1989. Fundamentals of Texture Mapping and
Joints in Theory and Practice. Rock Mechanics and Rock Image Warping.
Engineering, vol. 10, pp. 1-45.
Holschneider, M., Kronland-Martinet, R., Morlet, J.,
Barton, N., 1973. Review of a New Shear-Strength Criterion for Tchamitchian, P., 1990. A Real-Time Algorithm for Signal
Rock Joints. Engineering Geology, vol. 7, pp. 287-332. Analysis with the Help of the Wavelet Transform. In Wavelets.
Inverse Problems and Theoretical Imaging. Combes, J.-M.,
Beylkin, G., 1992. On the Representation of Operators in Bases Grossmann, A., Tchamitchian, P. (eds,), pp. 286–97. Springer
of Compactly Supported Wavelets. SIAM Journal on Numerical Berlin Heidelberg.
Analysis, 29 (6), pp. 1716–40.
ISRM, 1978. Suggested Methods for the Quantitative
Birgé, L., Pascal, M., 1997. From Model Selection to Adaptive Description of Discontinuities in Rock Masses. International
Estimation. In Festschrift for Lucien Le Cam. Pollard, D., Journal of Rock Mechanics, Mining Sciences and
Torgersen, E., Yang, G. L. (eds.), pp. 55–87. Springer New Geomechanics Abstracts, vol. 15, pp. 319–68.
York.
Khoshelham, K., Altundag, D., Ngan-Tillard, D., Menenti, M.,
Bitenc, M., Kieffer, D. S., Khoshelham, K., Vezočnik, R., 2015. 2011. Influence of Range Measurement Noise on Roughness
Quantification of Rock Joint Roughness Using Terrestrial Laser Characterization of Rock Surfaces Using Terrestrial Laser
Scanning. In Engineering Geology for Society and Territory -
 Scanning. International Journal of Rock Mechanics and Mining Discontinuity Characterization on Rock Cuts. Engineering
Sciences, vol. 48 (8), pp. 1215–23. Geology, vol. 106 (3–4), pp. 163–82.

Lichti, D. D., 2007. Error Modelling, Calibration and Analysis Tatone, B., Grasselli, G., 2012. An Investigation of
of an AM–CW Terrestrial Laser Scanner System. ISPRS Discontinuity Roughness Scale Dependency Using High-
Journal of Photogrammetry and Remote Sensing, vol. 61 (5), Resolution Surface Measurements. Rock Mechanics and Rock
pp. 307–24. Engineering, vol. 46 (4), pp. 657-681.

Lichti, D. D., Jamtsho, S., 2006. Angular Resolution of Tonon, F., Kottenstette, J. T., 2006. Laser and Photogrammetric
Terrestrial Laser Scanners. The Photogrammetric Record, vol. Methods for Rock Face Characteization. Report on a Workshop
21 (114), pp. 141–60. Held June 17-18, 2006 in Golden, Colorado. American Rock
Mechanics Association.
Mallat, S., 1991. Zero-Crossings of a Wavelet Transform.
Information Theory, IEEE Transactions on Information Theory, Vezočnik, R., 2011. Analysis of Terrestrial Laser Scanning
vol. 37 (4), pp. 1019–33. Technology for Structural Deformation Monitoring. PhD thesis.
University of Ljubljana, Ljubljana, Slovenia.
Nason, G. P., Silverman, B. W., 1995. The Stationary Wavelet
Transform and Some Statistical Applications. In Wavelets and Zhang, Y., Liu, J., Li, M., Guo, Z., 2014. Joint Image Denoising
Statistics, vol. 103, pp. 281–300. Springer-Verlag. Using Adaptive Principal Component Analysis and Self-
Similarity. Information Sciences, vol. 259 (0), pp. 128–41.
Patton, F. D., 1966. Multiple Modes of Shear Failure in Rock.
Presented at the 1st ISRM Congress. International Society for
Rock Mechanics.

Pesci, A., Teza, G., Bonali, E., 2011. Terrestrial Laser Scanner
Resolution: Numerical Simulations and Experiments on Spatial
Sampling Optimization. Remote Sensing, vol. 3 (1), pp. 167–84.

Priest, S. D., 1993. Discontinuity Analysis for Rock

Engineering. Springer.

Rasouli, V., Harrison, J. P., 2004. A Comparison of Linear

Profiling and an in-Plane Method for the Analysis of Rock
Surface Geometry. International Journal of Rock Mechanics
and Mining Sciences, vol. 41, Supplement 1 (0), pp. 133–38.

Riegl, 2015. Laser Scanner VZ-400, Datasheet. (April, 20015)

http://www.riegl.com/nc/products/terrestrial-
scanning/produktdetail/product/scanner/5/.

Shim, I., Soraghan, J. J., Siew, W. H., 2001. Detection of PD

Utilizing Digital Signal Processing Methods. Part 3: Open-Loop
Noise Reduction. Electrical Insulation Magazine, IEEE 17 (1),
pp. 6–13.

Siefko, S., 2010. Automated Rock Mass Characterisation Using

3D Terrestrial Laser Scanning. PhD thesis. International
Institute for Geo-information Science and Earth Observation,
Enchede, The Netherlands.

Smigiel, E., Alby, E., Grussenmeyer, P., 2008. Terrestrial Laser

Scanner Data Denoising by Range Image Processing for Small-
Sized Objects. The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information
Sciences, vol. 37.

Smigiel, E., Alby, E., Grussenmeyer, P., 2011. TLS Data

Denoising by Range Image Processing. The Photogrammetric
Record, vol. 26 (134), pp. 171–89.

Soudarissanane, S., Lindenbergh, R., Menenti, M., Teunissen,

P., 2011. Scanning Geometry: Influencing Factor on the Quality
of Terrestrial Laser Scanning Points. ISPRS Journal of
Photogrammetry and Remote Sensing, vol. 66 (4), pp. 389–99.

Sturzenegger, M., Stead, D., 2009. Close-Range Terrestrial

Digital Photogrammetry and Terrestrial Laser Scanning for

View publication stats