International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Contents lists available at ScienceDirect

International Journal of
Rock Mechanics and Mining Sciences
journal homepage: www.elsevier.com/locate/ijrmms

Semi-automatic extraction of rock discontinuities from point clouds using T

the ISODATA clustering algorithm and deviation from mean elevation
⁎ ⁎⁎
Peng Zhanga,d, , Junhui Lia, Xin Yangb,c,e, , Hehua Zhuf
a
College of Transportation Science & Engineering, Nanjing Tech University, Nanjing 210009, China
b
Key Laboratory of Virtual Geographic Environment, Ministry of Education, Nanjing Normal University, Nanjing 210023, China
c
Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China
d
School of Engineering, University of British Columbia, Kelowna, British Columbia, Canada V1V 1V7
e
Department of Earth, Environmental and Geographic Sciences, University of British Columbia, Kelowna, British Columbia, Canada V1V 1V7
f
College of Civil Engineering, Tongji University, Shanghai 200009, China

A R T I C LE I N FO A B S T R A C T

Keywords: This paper presents a methodology for semi-automatic extraction of rock discontinuities from point clouds on the
Rock discontinuities resulting 3D digital surface models of rock masses acquired by terrestrial laser scanning or photogrammetry.
Point clouds First, point clouds need to be pre-processed to generate a digital elevation model from a 3D digital surface model
Raster grid of a rock mass. Second, co-directional features of rock discontinuities are identified by classifying the standard
ISODATA
deviation of the elevation, slope aspect and slope gradient of raster grids in a digital elevation model using the
DEV
Iterative Self-organizing Data Analysis Techniques Algorithm (ISODATA). Third, planar features of rock dis-
continuities are inspected by judging whether the value range of the deviation from the mean elevation (DEV) in
the classified subset of raster grids is between + 1 and − 1 using the Topographic Position Index of terrain
analysis. Fourth, point clouds data corresponding to the target surfaces of rock discontinuities are acquired by
data post-processing using the ArcToolbox of ArcGIS. Finally, geological analysis is executed by calculating the
geometric features of the extracted point clouds. A case study of a rock slope along a highway is presented using
the proposed method, the results of which demonstrate that the proposed method is so rapid and accurate that it
could greatly reduce the amount of manual labour, improve the efficiency of surveying, and utilize the powerful
terrain analysis and visualization capabilities of ArcGIS to avoid the computational challenges associated with
directly analysing point cloud data.

1. Introduction However, there is an intractable technical bottleneck in that improve-

Collecting geological information on rock discontinuities is difficult,
time-consuming, and often dangerous when using field mapping and
handheld direct measuring devices with a scan-line method or a
window method that require direct access to the rock face, especially in
steep, inaccessible areas.1–4 Recently, several developed techniques,
such as digital photogrammetry5–8 and terrestrial and aerial LiDAR,9
have enabled the construction of 3D point clouds of rock surfaces to
rapidly obtain 3D geometric information about inaccessible terrain and
rock exposures. The virtual geometry of rock faces and terrains formed
by this dense point clouds allows engineers to perform the information
extraction on a computer with the aid of a mathematical algorithm.
ments in the automated extraction of rock discontinuity parameters are
needed to gain maximum value from these surface models.10
Several approaches have been used to extract rock discontinuities
from point clouds by creating 3D surface models of the rock masses.
First, the simplest approach for viewing point clouds is to manually
select at least three points on the target plane using commercial soft-
ware, such as Cyclone,11–13 Vulcan,14 Jointmetrix3D,15 Surpac,16 Sir-
ovision,17 3DM Analyst,18 or 3DGeomec.19 However, this approach is
time-consuming and laborious when analysing numerous rock dis-
continuities with complex distributions. Second, geological information
about the rock discontinuities can be acquired through statistical ana-
lyses of the geometric features (normal vectors) of triangular mesh units

⁎
Correspondence to: Urban Underground Space Research Centre, Nanjing Tech University, 200 North Zhongshan Road, Nanjing, Jiangsu Province 210009, China.
⁎⁎
Correspondence to: School of Geographic Science, Nanjing Normal University, No.1, Wenyuan Road, Xianlin University Distrct, Nanjing, Jiangsu Province
210023, China.
E-mail addresses: zhangpeng-mail@njtech.edu.cn (P. Zhang), JunhuiLi@njtech.edu.cn (J. Li), xxinyang@njnu.edu.cn (X. Yang),
zhuhehua@tongji.edu.cn (H. Zhu).

https://doi.org/10.1016/j.ijrmms.2018.07.009
Received 29 June 2017; Received in revised form 26 June 2018; Accepted 28 July 2018
Available online 07 August 2018
1365-1609/ © 2018 Elsevier Ltd. All rights reserved.
P. Zhang et al.
in a TIN generated from point clouds, such as the stereonet method11,20
and a data clustering algorithm.20,21 However, this method only pro-
vides statistical results and not the spatial distribution of rock dis-
continuities. Meanwhile, the statistical results are likely influenced by
the meaningless data of non-discontinuities. Third, rock discontinuities
can be automatically or semi-automatically extracted by identifying the
principal curvature changes of the vertices22 or by searching for the
best-fit planes9,23 by moving a sample window or cube on the point
clouds using geometric regional trend analysis software, such as Pla-
neDetect,20,24 DiAna 3D,9 Split-FX25,26 or Coltop 3D.27 The accuracy of
these approaches is severely susceptible to spurious points caused by
vegetation or other objects captured in the point clouds. The size of the
sample window or cube is difficult to determine and needs to be ad-
justed according to the characteristics of the rock discontinuities. If the
cube dimension is too large, small-scale discontinuity surfaces will be
omitted; if the cube dimension is too small, some pseudo-surfaces will
be counted in the extracted results.
This study aims to identify the advantages of statistical analysis on
the clustering and feature extraction of geometric contours. Previous
literature28 has discussed a similar idea as this study; however, it only
focused on the co-directional features of discontinuities with different
orientations, but not their planar features, which mistook the identities
of discontinuities and rock faces. By using the clustering algorithm on
Fig. 1. Flowchart of the proposed method.
77
International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87
the standard deviation of elevation, slope aspect and slope gradient
based on the Iterative Self-organizing Data Analysis Techniques Algo-
rithm (ISODATA), a semi-automatic method for extracting the exposed
surface of rock discontinuities from point clouds by identifying co-di-
rectional features is proposed. Planar features of rock discontinuities
are quantifiably judged by the deviation from the mean elevation
(DEV). Geological information of rock discontinuities can be obtained
by analysing the geometrical features and spatial relationships of those
extracted point clouds. The proposed method is applied to a Mingtang
Mountain Tunnel project and verified using a traditional method.
2. Methodology
A 3D digital surface model transformed from point clouds acquired
by terrestrial laser scanning or photogrammetry is composed of trian-
gular mesh units in a TIN. In a 3D digital surface model of a rock mass,
rock discontinuities are represented as different groups of gathering
triangular mesh units with near-identical normal vectors (co-directional
features) meeting a certain limitation of undulation (planar features).
The goal of this study is to identify, classify, and extract those triangular
mesh units with co-directional and planar features by developing a
series of algorithms using the powerful terrain analysis and visualiza-
tion capabilities of ArcGIS software.29
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Fig. 2. Flowchart of the ISODATA algorithm.

Fig. 3. Point cloud after pre-processing.

The proposed method is divided into seven steps, as illustrated by the
flowchart in Fig. 1. In Data Pre-processing, the first sub-step removes ve-
getation, resamples point clouds, reduces noise, and trims boundaries using
the open software of CloudCompare. The second sub-step constructs a di-
gital elevation model of a rock mass from point clouds using ArcGIS terrain
analysis. The aim of Identify Co-directional Features is to search the groups
of raster grids with a near-identical normal vector and limited undulation.
Planar features of rock discontinuities are identified by calculating the DEV
of raster grids. The value range of the DEV region is between + 1 and − 1,
which means the corresponding region is a plane. To ensure that these
parameters are meaningful, the coordinate system of the point cloud should
match the geodetic coordinate system in use at the site of interest. The dip
and dip direction of a discontinuity surface can be obtained by calculating
an arithmetic average of the normal vector from the triangular mesh units
or by determining a best-fit plane through points assigned to a discontinuity
using least squares.30,31

78
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Fig. 4. 3D surface model and digital elevation model of rock mass: (a) TIN data, (b) raster data, and (c) hill-shading map.

2.1. Identifying co-directional features of rock discontinuities Y axis is represented as the north/south direction. The range of slope
gradient α is set as [0°, 90°], and the range of slope aspect β is set as [0°,
Co-directional features of rock discontinuities in digital elevation 360°]. The third-order finite difference weighted by the reciprocal of
models are characterized by a series of groups gathering raster grids the squared distance is used to calculate the slope gradient and slope
with a near-identical normal vector and limited undulation. A near- aspect of raster grids in a digital elevation model of rock mass.
identical normal vector of groups gathering raster grids can be acquired Elevation standard deviation is a parameter that reflects the mac-
by classifying the slope aspect and slope gradient of the raster grids. roscopic parameters of the intensity of the overall ups and downs of the
Limited undulation of groups gathering raster grids can be achieved by terrain, as follows:
classifying the elevation of the standard deviation of raster grids. m n
Slope gradient and slope aspect are two correlated point location 1
z̅ =
mn
∑ ∑ z (i, j)
functions of surface morphology in digital elevation models. For surface i=1 j=1 (3)
morphology H = f(x, y) , slope gradient α and slope aspect β of point P
m n
can be represented as: 1
D(z) =
m (n−1)
∑ ∑ (z (i, j ) − z ̅ )2
i=1 j=1 (4)
α = arctan f x2 + f y2 (1)
σ= D (z ) (5)
fy fx
β= 180o−arctan + 90o where z(i, j) is the value of elevation, z̅ is the mean elevation, D(z) is
fx fx (2)
variance of elevation, and m and n are the grid numbers of the calcu-
where fx is the gradient of elevation in the X direction and f y is the lating window.
gradient of elevation in the Y direction. The initial direction of slope The ISODATA algorithm32 is an unsupervised classification method
aspect β is due north and slope aspect β is measured in the clockwise used to simultaneously classify raster grids of a rock mass in a digital
direction. The X axis is represented as the east/west direction, and the elevation model with three data layers, such as slope aspect, slope

79
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Fig. 5. Three kinds of distribution maps: (a) slope aspect, (b) slope gradient, and (c) elevation standard deviation.

Table 1 centres along each axis; θc is the minimum distance between two
Input parameters for the ISODATA clustering algorithm and DEV. cluster centres. Two clusters will merge if their distance is less than
Input parameters Nc K θs θc θn I R θc . L is the maximum number of cluster pairs that can be merged in
one iteration; I is the maximum number of iterations.
Value 1 8 20 20 20 10 10 × 10 grids Step 2: Assign the data to the clusters. If Di = min { x i
−yj ,j = 1, 2, …, Nc } , then x i ∈ Sj .
Step 3: If the number of points in Sj is less than θn , leave the subset
gradient, and elevation standard deviation. The ISODATA algorithm out of the sample; meanwhile, Nc is reduced by 1.
uses the following process: (1) Choose initial parameters. (2) Calculate Step 4: Recalculate the cluster centres:
the distance index function of each cluster. (3) Merge or split the
clusters according to the given requirements. (4) Repeat iteratively. 1
Zj =
Nj
∑x
Calculate new indexes and determine whether the results meet the x ϵSj (6)
clustering requirements. The detailed steps of the ISODATA algorithm
where Nj is the number of samples in Sj .
are listed as follows (see Fig. 2):
Step 5: Let Dj̅ be the average distance between Sj and the associated
cluster centre zj .
Step 1: Input the dataset with N points {x i,i = 1, 2, …, N } . In this
study, the input dataset {x i,i = 1, 2, …, N } is set as three data layers, 1
Dj̅ = ∑ x −zj
including the slope gradient of the raster grids, the slope aspect of Nj x ∈ Sj (7)
the raster grids and the elevation standard deviation of the raster
grids. N is the number of grids. Set Nc as the initial cluster centres
{z1, z2, …, z Nc } . There is no need for the value of Nc to equal the Step 6: Let D̅ be the overall average of these distances.
number of clusters. The initial values of the centres can be selected N
1
randomly. K is the expected number of clustering centres; θn is the D̅ = ∑ Nj D̅j
N j=1 (8)
minimum number of points that can form a cluster. A cluster will be
combined with another if the number of points in it is less than θn . θs
is the maximum standard deviation of the points from their cluster Step 7: Determine whether split or merge is needed. If the iterative

80
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Fig. 6. Clustering results of ISODATA clustering: (a) preliminary result and (b) results after removing meaningless data.

computation number has reached I , then set θc = 0 and advance to then return to step 1; otherwise, return to step 2. Update the itera-
step (11). If Nc ≤ K /2 , then advance to step (8) and split the cluster. tion by adding 1.
Step 8: For each cluster Sj , calculate vector σj ; σij is the standard
deviation of the i th coordinate of σj directed from zj to every point of 2.2. Judging planar features of rock discontinuities
Sj .
σj = (σ1j, σ2j, …, σnj )T (9) Topographic position index (TPI) measures the difference between
the elevation at centre point z 0 and the average elevation z̅ around it
within a predetermined radius of neighbourhood R 33:
Nj
1 TPI = z 0 − z ̅ (14)
σij =
Nj
∑ (xik − z ij)2
k=1 (10) 1
z̅ =
nR
∑i ∈R z i (15)
Here, i(i = 1, 2, …, n) is the dimensions of the sample feature vector,
j(j = 1, 2, …, Nc ) is the number of clusters, and Nj and Sj are the Positive TPI values indicate that the central point is higher than its
numbers of samples. average surroundings, while negative values indicate a lower-than-
Step 9: Let σjmax = max (σ1, σ2, σ3, …, σj ) . average position. In terrain analysis, TPI > 0 indicates that the in-
Step 10: For each cluster Sj , if σjmax > θs and either (Dj̅ > D̅ and vestigated point is located near the ridge, TPI < 0 indicates that the in-
Nj > 2(θN +1) ) or Nc ≤ K /2 , then calculate two new clusters according vestigated point is located near the valley, and TPI ≅ 0 indicates that
to σjmax . Then, replace zj with the two new clusters, increment k , and the investigated point is located in plain terrain or the middle part of
split Sj into two clusters. If a splitting operation is activated, advance hill slope. However, the range of TPI depends not only on elevation
to Step 2. differences but also on R .
Step 11: Compute the pairwise inter-cluster distances between all DEV measures the topographic position of central point z 0 using TPI
distinct pairs of cluster centres. and the standard deviation of elevation SD34:
Dij = z i − zj , i = 1, 2, …, Nc −1; j = i+1, …, Nc (11)
1
SD =
nR − 1
∑i =1 (z i − z ̅ )2
(16)
Step 12: Compare Dij and θc , and select all Dij less than θc as a new
subset. Then, sort the elements of the subset in ascending order: TPI z0 − z̅
DEV = =
SD 1
∑i = 1 (z i − z ̅ )2
{Di1j1, Di2j2, …, DiLjL} (12) nR − 1 (17)
where Di1j1 < Di2j2 <⋯<DiLjL . DEV measures the topographic position as a fraction of local relief
Step 13: Combine the two cluster centres z ik and zjk , and calculate normalized to local surface roughness. DEV > 0 indicates that the
the new cluster centres: central point is higher than its neighbourhood, DEV< 0 indicates that
1 the central point is lower than its neighbourhood, and DEV ≅ 0 in-
Zk* = [Nik z ik + Njk zjk ],k = 1, 2, …, L dicates that the central point is in a plain with its neighbourhood. The
Nik + Njk (13)
output values of DEVs mostly range between + 1 and − 1, and values
outside this range may indicate anomalies within the DEM. Hereby,
Step 14: If this is the last iteration (i.e., I times), then the algorithm planar features of rock discontinuities can be identified by calculating
ends. Otherwise, if it is necessary to change the input parameters, the DEVs of raster grids, statistically analysing the DEV Region in the

81
P. Zhang et al.
classified subset of raster grids, and judging whether the DEV Region is
between + 1 and − 1, which means the corresponding region is a
plane.
3. Case study
3.1. Data acquisition
The 3D point clouds data used in our study include data acquired by
a Reigl LMS-Z420i terrestrial laser scanner and the LiDAR data of a
steep rock slope along the Wu-Yue Highway in China. Scanning was
performed from three observation stations spaced approximately 10 m
apart. The observation stations were approximately 8 m from the rock
face. The resulting combined point clouds from the three observation
stations had 1,682,432 points with a point resolution of approximately
5 mm.
Fig. 7. Classifications of raster grids represented by rock discontinuities: (a) class 1 (rock stratum), (b) class 2 (rock joint 1), (c) class 3 (rock joint 2), (d) class 4 (rock
fault), (e) class 7 (rock faces), (f) class 8 (broken rock faces), and (g) final results. (For interpretation of the references to color in this figure, the reader is referred to
the web version of this article.)
82
International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87
3.2. Data pre-processing
Collecting the point cloud data inevitably involves a large amount of
noise and unnecessary points that could affect the detection of rock
discontinuities due to vegetation sheltering, terrain masking, light in-
terference and shooting or scanning angles. Vegetation removal is
performed using Riscan Pro software, which comes with the Reigl
scanner. The Moving Least Squares method is used to reduce noise.
Resampling points with a minimum distance of 5 mm and boundary
trimming are accomplished using CloudCompare software. Fig. 3 is the
point cloud of a rock mass after the above pre-processing.
In the data import process of ArcGIS, the coordinates of X, Y and Z
are changed to achieve coordinate transformation. Delaunay triangu-
lation and Rasterization from vector to grid with a 5-mm spatial re-
solution are performed on the point cloud to obtain a 3D surface model
and a digital elevation model of the rock mass slope using the spatial
analysis tools of ArcGIS, as shown as Fig. 4.
P. Zhang et al.
Fig. 7. (continued)
3.3. Identifying rock discontinuities
Fig. 5 shows the distribution maps of the slope aspect, slope gra-
dient, and elevation standard deviation on raster grids, in which the
slope aspect is visualized according to eight equal parts in 360°, the
slope gradient is visualized according to nine equal parts in 90°, and the
elevation standard deviation is visualized according to eight equal parts
of maximum deviation of the elevation standard deviation.
The input parameters of the ISODATA clustering algorithm and DEV
83
International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87
are listed in Table 1. The initial value Nc of the clustering centre can be
selected randomly and set as 1. K , which must be greater than the
number of clustering centres, is set as 8 because there are no more than
eight sets of rock discontinuities in this case. θs , θc , and I are set as the
default values of the ISODATA clustering algorithm in ArcGIS. θn is set
as 20 to identify the rock discontinuities with an area of 0.0005 m2. The
size of neighbourhood window R is 10 × 10 grids.
Fig. 6(a) demonstrates 8 clustering sets by ISODATA clustering the
slope gradient, slope aspect, and elevation standard deviation. To
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

exclude meaningless data interference for subsequent work, the

meaningless data can be removed by filtering those classifications of
smaller groups of raster grids, which are less than 50 grids in ap-
proximately 5 mm spatial resolution using ArcGIS, as shown in
Fig. 6(b).
Figs. 7(a)–(f) show the spatial distribution of the raster grids in
Classes 1, 2, 3, 4, 7, and 8 represented by the suspected rock dis-
continuities that have the co-directional and planar features. Through
visual identification, Class 1 of the raster grids is rock stratum, Classes 2
and 3 are a set of rock joints with a certain amount of undulation and
that need to be merged into one class, Class 4 is a set of rock faults, and
Classes 7 and 8 are broken and flat rock face surfaces. Thus, the final
extracted results of the rock discontinuities are raster grids of rock
strata labelled with red, raster grids of rock joints labelled with green,
and raster grids of rock faults labelled with blue, as showed in Fig. 7(g).

3.4. Data post-processing

The process outputs of data post-processing are shown in Fig. 8. The

data voids in Fig. 8(a) are filled, as shown in Fig. 8(b). The highlighted
line in Fig. 8(b) is the boundary of the surface discontinuities. The blue
scattered points are the raw point cloud data corresponding to the ex-
tracted surface discontinuities, as shown in Fig. 8(c).

3.5. Discontinuities characterization

Table 2 lists the geological information of the top 20 surface dis-

Fig. 8. Process outputs of data post-processing: (a) grid to vector, (b) void fill
and boundary trim, and (c) vector to point cloud. (For interpretation of the
references to color in this figure, the reader is referred to the web version of this
article.).
continuities in rock strata, rock joints, and rock faults. The position of
the top 20 surface discontinuities in rock strata, rock joints, and rock
faults are shown in Fig. 9.
In Fig. 9, those discontinuity surfaces that belong to a set of dis-
continuities can be easily identified visually or judged by calculating
the plane equation from a subset of the point cloud data. Geological
information of a set of discontinuities can be acquired through the
mean dip and mean dip direction of those discontinuity surfaces. For
example, F2, F1, and F12 belong to a set of rock faults, and the dip and
dip direction of this set have mean values of F2, F1, and F12, the dip of
which is 8.1 and the dip direction of which is 214.0. A comparison of
the results by mapping using a geological compass shows that using a
point cloud typically yields better data.
The characteristics of the nine sets of rock discontinuities in Fig. 9,
which belong to S1, S2, S4, J6, J9, J14, F1, F3 and F4, can be acquired
by using a geological compass, as listed in Table 3. A comparison of the
results shows that using a point cloud typically yields better data.

Table 2
Dip and dip direction of the top 20 area discontinuities on the surface of rock strata, rock joints and rock faults.
ID Point count Dip (°) Dip Direction (°) Area (m2) ID Point count Dip (°) Dip Direction (°) Area (m2)

S1 45,396 62.4 043.6 1.135 S11 9999 58.2 045.4 0.250

S2 43,648 69.0 044.5 1.091 S12 8956 62.2 046.3 0.224
S3 31,249 62.6 046.9 0.781 S13 8494 59.8 043.0 0.212
S4 27,872 64.4 044.5 0.697 S14 8121 65.8 043.1 0.203
S5 25,797 58.1 051.4 0.645 S15 8105 71.4 050.7 0.203
S6 21,503 68.1 053.0 0.538 S16 7589 67.4 050.1 0.190
S7 12,980 69.8 045.8 0.325 S17 7291 69.1 046.3 0.182
S8 12,402 59.9 043.8 0.310 S18 7160 66.7 042.4 0.179
S9 10,734 62.3 041.1 0.268 S19 6825 61.7 038.5 0.171
S10 10,536 66.8 048.1 0.263 S20 6713 52.5 042.8 0.168
J1 36,819 69.5 153.8 0.920 J11 6681 64.3 148.7 0.167
J2 23,882 63.3 177.5 0.597 J12 6491 54.8 168.3 0.162
J3 21,820 66.5 143.6 0.546 J13 6266 52.3 150.7 0.157
J4 17,955 63.7 150.0 0.449 J14 6028 64.4 150.3 0.151
J5 14,091 51.9 141.9 0.352 J15 6006 67.7 143.4 0.150
J6 13,041 70.9 143.1 0.326 J16 5811 56.9 178.3 0.145
J7 12,950 67.1 146.0 0.324 J17 5553 52.2 142.1 0.139
J8 8819 62.3 161.1 0.220 J18 5510 61.8 173.2 0.138
J9 8510 63.2 151.4 0.213 J19 5489 50.7 160.5 0.137
(continued on next page)

84
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Table 2 (continued)

ID Point count Dip (°) Dip Direction (°) Area (m2) ID Point count Dip (°) Dip Direction (°) Area (m2)

J10 6812 50.7 183.8 0.170 J20 5355 56.1 121.1 0.134
F1 16,253 7.4 213.4 0.406 F11 1183 5.3 236.4 0.030
F2 16,121 8.5 215.6 0.403 F12 1150 8.4 213.2 0.029
F3 9441 5.6 216.5 0.236 F13 1145 1.3 222.1 0.029
F4 6081 3.5 218.4 0.152 F14 1029 3.5 234.9 0.026
F5 2718 1.3 214.7 0.068 F15 883 0.4 225.7 0.022
F6 1999 11.1 235.4 0.050 F16 874 6.4 243.9 0.022
F7 1733 10.9 219.2 0.043 F17 815 4.5 236.8 0.020
F8 1336 0.1 231.8 0.033 F18 802 11.2 245.9 0.020
F9 1276 1.1 226.3 0.032 F19 795 9.1 222.4 0.020
F10 1264 4.5 231.3 0.032 F20 745 4.1 232.0 0.019

Fig. 9. Positions of the top 20 area discontinuities on the surfaces of rock strata, rock joints and rock faults.

Table 3 identified at an approximately 5 mm spatial resolution. If θn is set to 10,

Comparison of the use of a geological compass and the proposed method. rock discontinuities with an area of 0.00025 m2 can be identified at an
approximately 5 mm spatial resolution. Fig. 10 is the extracted result of
ID Point cloud Compass
Class 1 under θn = 20 (Red label) and θn = 10 (Red label + Green
Dip (°) Dip Direction (°) Dip (°) Direction (°) label).
The range of DEV mainly depends on the size of the neighbourhood
S1 62.4 043.6 63.1 044
window R . The appropriate size of the neighbourhood window is ex-
S2 69.0 044.5 67.9 044
S4 64.4 044.5 64.0 044 amined by the sensitivity analysis of DEV. Fig. 11 shows the DEV of
J6 70.9 143.1 71.4 144 raster grids for five neighbourhood window sizes, which demonstrates
J9 63.2 151.4 63.7 150 that planar features of rock discontinuities can be clearly sketched by
J14 64.4 150.3 63.9 150 positive or negative extremum of DEV when the size of neighbourhood
F1 7.4 213.4 8.0 213
window is greater than an 11 × 11 grid.
F3 5.6 216.5 6.1 217
F4 3.5 218.4 4.3 217 After statistically computing the DEV Region in Classes 1–8 of the
raster grids, the DEV Regions of Classes 5 and 6 are beyond the range of
+ 1 and − 1, as shown in Table 4. Table 4 is the Mean DEV Region of
4. Discussion the 8 classified subsets in Fig. 7 during sensitivity analysis of the
neighbourhood. According to the demonstration in Fig. 11, when the
In Table 1, the minimum number of points that can form cluster θn size of the neighbourhood window is greater than an 11 × 11 grid, the
and the size of neighbourhood window R are both related to the grid Mean DEV Region of Class 5 and Class 6 are outside the range from + 1
scale of the point cloud and the size of the target discontinuity surface. to − 1 for the appropriate size of the neighbourhood window, which
If θn is set to 20, rock discontinuities with an area of 0.0005 m2 can be means the raster grids represented by Class 5 and Class 6 do not meet

Fig. 10. Comparison of θn = 20 and θn = 10 in Class 1. (For interpretation of the references to color in this figure, the reader is referred to the web version of this
article.).

85
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Fig. 11. DEV of raster grids with different neighbourhood window sizes: (a) DEM hill-shading image, (b) 3 × 3 grids, (c) 7 × 7 grids, (d) 11 × 11 grids, (e) 15 × 15
grids, and (f) 19 × 19 grids.

the planar characteristic requirement. However, the raster grid classi- is inside the range from + 1 to − 1, the closer the Mean DEV Region
fications that meet the geometric characteristic of rock discontinuities gets to + 1 or − 1, the rougher the surface of the corresponding raster
are Classes 1, 2, 3, 4, 7, and 8. Meanwhile, even if the Mean DEV Region grids could be.

86
P. Zhang et al. International Journal of Rock Mechanics and Mining Sciences 110 (2018) 76–87

Table 4 2005. International Society for Optics and Photonics; 2005:17–27.

Mean DEV Region of 8 classified subsets with different neighbourhood window 6. Sturzenegger M, Stead D. Close-range terrestrial digital photogrammetry and ter-
sizes. restrial laser scanning for discontinuity characterization on rock cuts. Eng Geol.
2009;106:163–182.
Clustering 3×3 7×7 11 × 11 15 × 15 19 × 19 7. Tannant DD. Review of photogrammetry-based techniques for characterization and
number window window window window window hazard assessment of rock faces. Int J Geohazard Environ. 2015;1(2):76–87.
8. Li X, Chen J, Zhu H. A new method for automated discontinuity trace mapping on
Class 1 0.11 0.26 0.37 0.45 0.51 rock mass 3D surface model. Comput Geosci. 2016;89:118–131.
9. Gigli G, Casagli N. Semi-automatic extraction of rock mass structural data from high
Class 2 0.08 − 0.11 − 0.33 − 0.55 − 0.77
resolution LiDAR point clouds. Int J Rock Mech Min Sci. 2011;48:187–198.
Class 3 0.26 0.42 0.52 0.54 0.53
10. Vöge M, Lato MJ, Diederichs MS. Automated rockmass discontinuity mapping from
Class 4 − 0.01 0.16 0.46 0.69 0.87
3-dimensional surface data. Eng Geol. 2013;164:155–162.
Class 5 0.33 0.74 1.00 1.20 1.40 11. Mah J, Samson C, McKinnon SD. 3D laser imaging for joint orientation analysis. Int J
Class 6 0.36 0.78 1.14 1.43 1.66 Rock Mech Min Sci. 2011;48:932–941.
Class 7 0.02 − 0.24 − 0.51 − 0.74 − 0.93 12. Jason M, Claire S, Stephen DM. 3D laser imaging for joint orientation analysis. Int J
Class 8 0.22 0.41 0.50 0.64 0.74 Rock Mech Min Sci. 2011;48:932–941.
13. Deliormanli HA, Maerz HN, Otoo J. Using terrestrial 3D laser scanning and optical
method to determine orientations of discontinuities at a granite quarry. Int J Rock
Mech Min Sci. 2014;66:41–48.
5. Conclusions 14. Maptek. Vulcan home page. 〈http://www.maptek.com/products/vulcan/index.
html〉 Accessed January 2017.
This paper describes a new method for semi-automatic rock dis- 15. 3G Software and Measurement. Jointmetrix3D, home page. 〈http://www.3gsm.at/
eng/jminfo_eng.asp〉 Accessed January 2017.
continuity extraction from the 3D surface geometry of a rock face
16. Gemcom. Surpac, Geology and Mine Planning, home page. 〈http://www.
captured in a point cloud data set based on the ISODATA clustering gemcomsoftware.com/products/surpac/〉 Accessed January 2017.
algorithm and the DEV index. The method can eliminate outliers and 17. CSIRO. Sirovision, home page. 〈http://www.sirovision.com/〉 Accessed January
noise in the point clouds, such as those caused by points representing 2017.
18. Adam Technology. 3DM Analyst, Mining Analysis Software, home page. 〈http://
vegetation. The proposed method can extract not only large-scale dis- www.adamtech.com.au/3dm/Analyst.html〉 Accessed January 2017.
continuity surfaces but also small-scale discontinuity surfaces. 19. Gexcel. 3DGeomec, home page. 〈http://www.gexcel.it/en/prodotti/3dgeomec〉
The parameters influencing the accuracy of the proposed method Accessed January 2017.
20. Lato M, Diederichs MS, Hutchinson DJ, Harrap R. Optimization of LiDAR scanning
are the minimum number of points of cluster θn and the radius of and processing for automated structural evaluation of discontinuities in rockmasses.
neighbourhood window R . The two parameters are related to the inter- Int J Rock Mech Min Sci. 2009;46:194–199.
space of the point cloud and the size of the target discontinuity surface. 21. Olariu MI, Ferguson JF, Aiken CLV, Xu X. Outcrop fracture characterization using
terrestrial laser scanners: deep-water Jackfork sandstone at Big Rock Quarry,
The appropriate parameters can be acquired by a sensitivity analysis of Arkansas. Geosphere. 2008;4(1):247–259.
the ISODATA clustering algorithm and the DEV index according to 22. Umili G, Ferrero A, Einstein HH. A new method for automatic discontinuity traces
actual geological conditions. sampling on rock mass 3D model. Comput Geosci. 2013;51:182–192.
23. Otoo JN, Maerz NH, Duan Y, Li XL. LiDAR and optical imaging for 3D fracture or-
The proposed method as demonstrated by the case study can effi- ientations. In: Proceedings of the 2011 NSF Engineering Research a d Innovation
ciently and accurately extract joint surfaces from a point cloud dataset, Conference. Atlanta, Georgia; 2011.
and it can be used to replace traditional methods. Further research and 24. Lato MJ, Vöge M. Automated mapping of rock discontinuities in 3D lidar and pho-
togrammetry models. Int J Rock Mech Min Sci. 2012;54:150–158.
testing will focus on the feasibility of the proposed method in a wide
25. Slob S, Hack R, van Knapen B, Turner K, Kemeny JA. Method for Automated
range of rock mass conditions and rock face topologies. Discontinuity Analysis of Rock Slopes with 3D Laser Scanning. Washington, DC:
Transportation Research Board; 2005.
Acknowledgements 26. Slob S, Hack HRGK, Feng Q, Roshoff K, Turner AK. Fracture mapping using 3D laser
scanning techniques. :In: Proceedings of the 11th Congress of International Society for
Rock Mechanics. Lisbon, Portugal. Vol. 1, 2007:299–302.
The authors gratefully acknowledge the support of the National 27. Jaboyedoff M, Metzger R, Oppikofer T, et al. New insight techniques to anlayze rock-
Nature Science Foundation of China (NSFC No. 41372295, No. slope relief using DEM and 3D imaging cloud points. In: Rock Mechanics: Meeting
Society’s Challenges and Demands. Proceedings of the First Canada-US Rock Mechanics
U1703244, No. 41771415 and No. 41102178); the Youth Teacher Symposium. Vancouver. London: Taylor & Francis; 27–31 May 2007:61–68.
Overseas Research Project, Jiangsu Province, China; and the 28. Chen J, Zhu H, Li X. Automatic extraction of discontinuity orientation from rockmass
Postgraduate Research & Practice Innovation Programme of Jiangsu surface 3D pointcloud. Comput Geosci. 2016;95:18–31.
29. Tang GA, Yang X. Experimental Tutorial of Spatial Analysis Geographic Information
Province (No. KYCX17-0928). System by ArcGIS. 2nd ed. Peking: Science Press; 2012.
30. Shi XB. Research for Identifying Structural Plane Based on Three-Dimensional Laser
References Scanning Data. [Master Thesis]. Beijing, China: China University of Geosciences; 2010.
31. Zhu YF. The Theory and System Development of Discontinuity Identification Based on 3D
Laser Scanning Data. [Master Thesis]. Beijing, China: China University of Geosciences;
1. Barton N, Lien R, Lunde J. Engineering classification of rock masses for the design of 2012.
tunnel support. Rock Mech. 1974;6(4):189–236. 32. Li MC, Han S, Shi J. An enhanced ISODATA algorithm for recognizing multiple
2. ISRM. Suggested methods for the quantitative description of discontinuities in rock electric appliances from the aggregated power consumption dataset. Energy Build.
masses. Int J Rock Mech Min Sci Geomech Abstr. 1978;15:319–368. 2017;140:305–316.
3. Franklin JA, Maerz NH, Bennett CP. Rock mass characterization using photoanalysis. 33. Gallant JC, Wilson JP. Primary topographic attributes. Terrain Analysis: Principles and
Geotech Geol Eng. 1988;6(2):97–112. Applications. New York: Wiley; 2000:51–85.
4. Priest SD. Discontinuity Analysis for Rock Engineering. London: Chapman & Hall; 1993. 34. Reu DJ, Bourgeois J, Bats M, et al. Application of the topographic position index to
5. Roncella R, Forlani G, Remondino F. Photogrammetry for geological applications: heterogeneous landscapes. Geomorphology. 2013;186:39–49.
automatic retrieval of discontinuity orientation in rock slope. In: Electronic Imaging

87