Content uploaded by Candan Gokceoglu
Author content
All content in this area was uploaded by Candan Gokceoglu on Nov 10, 2020
Content may be subject to copyright.
© International Society for Rock Mechanics and Rock Engineering
Norwegian Group for Rock Mechanics
ISBN: 978-82-8208-072-9
C.C. Li, H. Ødegaard, A.H. Høien, J. Macias (Eds.)
ISRM International Symposium
Eurock 2020 –
Hard Rock Engineering
Trondheim, Norway, 14-19 June
Geological strength index prediction by vision and
machine learning methods
A.S. Bozkir
Department of Computer Engineering, Hacettepe University, Ankara, Turkey
H.A. Nefeslioglu
Department of Geological Engineering, Hacettepe University, Ankara, Turkey
O. Kartal
General Directorate of Turkish State Railways, Ankara, Turkey
E. Sezer
Department of Computer Engineering, Hacettepe University, Ankara, Turkey
C. Gokceoglu
Department of Geological Engineering, Hacettepe University, Ankara, Turkey
candan.gokceoglu@gmail.com (email of the corresponding author)
Abstract
In this study, we propose a novel methodology in order to predict Geological Strength Index (GSI)
values of rock outcrops by using computer vision and machine learning methods. For this purpose, we
separately employed two different global image descriptors namely GIST and HOG (Histogram of
Oriented Gradients) to extract a discriminative and “holistic” visual signature of the input images.
Subsequently, after visual feature extraction, we utilized three different machine learning methods (i.e.
SVM, Random Forests and XGBoost) for the estimation of geological strength index values of rock
outcrops by using digital images. We collected a corpus involving 1488 high-quality images of eleven
rock types acquired from twenty outcrops. According to the results of the experiments carried out by
using test data, GIST descriptor presents a superior regression performance (i.e. accuracy of 76.36%)
compared to the HOG (i.e. accuracy of 61.41%). The results show that GIST descriptor has a
promising estimation performance along with a fast and easy to compute scheme for both desktop and
mobile environments.
Keywords
Geological Strength Index (GSI), GIST, HOG, SVM, Random Forest, XGBoost
Eurock 2020 – Hard Rock Engineering
2
1 Introduction
One of the most important stages in rock engineering applications is the assessment of rock mass
environment. To evaluate a rock mass environment, strength and deformation characteristics of the
whole mass including intact rock and rock mass discontinuities should be defined. The critical
problem is the sampling size necessary for laboratory testing. Since, there is not yet a test equipment
to test a specimen having rock mass size in the laboratory, the assessment of the strength and
deformation characteristics of rock masses in the laboratory is almost impossible. By considering this
limitation, Hoek and Brown (1980) suggested an empirical failure criterion to evaluate the strength of
rock mass. The quality of rock mass, rock mass discontinuity properties and the properties of intact
rock material e.g. uniaxial compressive strength of intact rock (σci) and the material constant (mi)
constitute the basis of the failure criterion.
To evaluate the quality of rock mass, Hoek and Brown (1997) proposed a scheme namely Geological
Strength Index (GSI). The state of blockiness and the conditions of rock mass discontinuities are the
essentials of the GSI scheme. The experience of an engineering geologist affects the determination of
the GSI value of a rock mass. According to the GSI scheme proposed by Hoek and Brown (1997), the
GSI value of a rock mass is determined by visual interpretations of blockiness and rock mass
discontinuity conditions. The minimum and maximum values of GSI were 10 and 80 on the first
scheme. Foliate/laminated/sheared and massive or intact rock environments were also added in the
subsequent versions of the scheme (Hoek et al. 1998; Hoek 1999). Taking into account the natural
uncertainty of the geological environment, it is more appropriate to define the GSI values of a rock
mass by quoting a range, instead of precise determinations (Hoek and Brown, 1997; Hoek et al., 1998;
Hoek 1999). However, there are also some GSI schemes from which GSI values can be calculated
precisely (Sonmez and Ulusay, 1999; Cai et al., 2004; Russo, 2007; Russo, 2009; Hoek et al., 2013).
Marinos et al. (2005) explain this peculiarity as follows; the main reason for the attempt to obtain
more accurate GSI values is due to the over-deterministic expectations of the responsible engineers.
Considering the natural uncertainty of the geological environment, it is evident that precise GSI
determinations may not be reliable. The blockiness and rock mass discontinuity assessments may not
be so precise as the geological engineer expected.
In the present study, to reduce uncertainties and expert subjectivities, a novel GSI determination
method is represented. To predict Geological Strength Index (GSI) values of rock outcrops, computer
vision and machine learning (ML) methods were implemented. For the purpose, the digital images of
different rock masses of which GSI values were assigned by experienced engineering geologists were
used. To extract a discriminative and “holistic” visual signature of the input images, two different
global image descriptors namely GIST and HOG (Histogram of Oriented Gradients) were separately
employed. Subsequently, after visual feature extraction, 3 different ML methods (i.e. SVM, Random
Forests and XGBoost) for the estimation of geological strength index values of rock outcrops were
utilized by using the digital images.
2 Data
In the present study, the GSI values which were assigned for 11 rock types at 20 outcrops were used
(Fig. 1). The GSI values given for lava and peridotite were acquired from Marinos et al., (2006), while
sandstone-siltstone and siltstone-sandstone data were obtained from Marinos (2019). The remainder
was provided by the General Directorate of the Turkish State Railways. The GSI determinations were
carried out according to the GSI schemes suggested by Hoek et al., (1998) and Marinos and Hoek
(2001). The expressive GSI values that can be obtained from the schemes are only multiples of five
(Hoek et al. 1998; Marinos and Hoek 2001). For this reason, as can be seen from Fig. 2a, the GSI
values, which were implemented in training and test stages were assigned as multiples of five.
The number of digital images in our corpus is 62. Though our collection covers almost the whole
range of GSI scope acquired from various kinds of rocks, the total number of images was not adequate
to create a generalizable ML model. Thus, for increasing the count of image samples we utilized a
methodology called “data augmentation”. To achieve this goal by using a single frame, we first
cropped four equal-sized sub-regions of the image that each occupies top-left, top-right, bottom-left
and bottom-right parts of the input image. In addition, we also included an image-center crop having
the same width and height yielding 5 = 4+1 images. As a result, we obtained 5 additional sub-samples
from a single photo which makes a data expansion at a rate of 1 to 6.
Eurock 2020 – Hard Rock Engineering
3
Fig. 1 Example digital images from different rock types. The images of “Lava” and “Peridotite” were obtained from Marinos
et al., (2006); “Sandstone-Siltstone” and “Siltstone-Sandstone” were acquired from Marinos (2019); the remainder was
provided by the General Directorate of the Turkish State Railways.
Next, we have continued to increase the number of image samples by also rotating each image
obtained at the previous step by degrees of 90, 180 and 270. In this regard, we have produced 24
images from just 1 image sample implying an increment by factor of 24 (i.e., 6×4). Thus, the total
number of data samples reached up to 1488 (i.e., 62×24). With this expansion we aimed to achieve
more generalizable models. Moreover, we also investigated the performance of descriptors against
rotational variances. As stated in (Bozkir et al. 2020), data augmentation is a widely employed
technique in computer vision based tasks due to its ability to reduce bias and overfitting by also
creating more robust models against rotations and transformations.
Following the data augmentation stage, for each GSI category, we have randomly split the whole
dataset into training and testing sets having ratio of 80% and 20% respectively. Consequently, our
dataset has been restructured by covering 1193 training and 295 testing images. The distribution of the
samples in training/testing sets can be viewed in Fig 2a.
Fig. 2 Distribution of training and test sets per GSI category (a); the number of images per rock type (b).
Eurock 2020 – Hard Rock Engineering
4
3 Methods and Application
In this section, we first briefly explain the employed computer vision and ML methods. Next, we have
presented how we extract the visual descriptors based on GIST and HOG (Histogram of Oriented
Gradients) features, which then be fed to ML algorithms to predict GSI scores. Moreover, we have
demonstrated our experimental results on both training and testing data.
3.1 Vision based feature extraction
Recent years have witnessed widespread use of computer vision based studies in numerous diverse
fields such as human action recognition (Zhang et al. 2019), phishing web page detection (Dalgic et al.
2018) and medical image analysis (Diament et al. 2018). Moreover, the literature also covers many
successful examples of vision based studies applied to the domain of geology. For instance,
Farahbakhsh et al. (2018) have employed vision techniques in order to reduce the dimension of data
and remove noise to enhance the expression of geological lineaments. In this study, we have employed
GIST (Oliva and Torralba 2001) and HOG image descriptors (Dalal and Triggs 2005) in order to
reveal and represent the discriminative visual cues located in rock mass photos. To this end, the
mentioned descriptors – combined with conventional ML methods – have yielded the basis of our GSI
score estimator scheme.
3.1.1 GIST descriptors
Invented by Oliva and Torralba (2001), the aim of GIST descriptor is to produce a low dimensional
representation of a visual scene without the need for a segmentation process (Diament et al. 2018).
Furthermore, the GIST descriptors are often employed for scene understanding via focusing on the
visual outlines to reveal identifiable visual characteristics – the spatial envelope – of an image.
Similarly, Diament et al. (2018) state that the GIST descriptor deals with the shape of the scene by
considering the relationship between the outlines of the surfaces as well as their properties while
discarding the local objects and their relationships in the scene/image. To compute a GIST based
feature vector (960 dimensional in our case), first, the input image is divided into n × n blocks where
each block is processed with Gabor filters at different scales and orientations (Eroglu et al. 2019).
Finally, the histograms collected from different orientations and cells are concatenated to obtain a
unified and image resolution invariant feature vector. The GIST based feature vector generation
process along with ML methods has been depicted in Fig. 3 below.
3.1.2 Histogram of oriented gradients (HOG) descriptors
On the other hand, invented by Dalal and Triggs (2005), HOG features aim to represent the edge and
corner structures of an image by computing the intensity-based gradients and their magnitudes in
fixed-sized blocks and cells as well. Technically speaking, HOG descriptor first divides an image into
small neighbor areas called blocks having equal-sized cells and then it computes the histogram of
gradient and edge directions of the pixels within each cell followed by a block-level normalization
Fig. 3 The process of GIST based feature extraction and utilization of machine learning methods to estimate the GSI value.
Eurock 2020 – Hard Rock Engineering
5
stage. Next, normalized gradient magnitudes for each angle interval are collected to form a large
histogram which also takes the spatial ordering into account.
Though they serve fast and effective solutions, the HOG features are known to be invariant to
deformations/transformations (e.g. shifting and rotation) to a certain extent. Further, it should be
noted that, during the computation of HOG, the size of the input image affects the dimension of the
resultant feature vector while GIST does create a fixed-sized feature vector regardless of the
resolution. Besides, the main goal of HOG is to capture the edge and corner like structures for
characterizing the scene without requiring color information whereas GIST is also able to take the
color as a source of information.
3.1.3 Feature extraction in details
In order to compute the aforementioned descriptors regarding the training and test images, we created
a script coded in Python 3.6 programming language and employed “pyleargist” package in order to
extract color GIST descriptors. GIST descriptors can be obtained from grayscale and colored images.
In this study, we have preferred color GIST features since we did not want to lose color information
and allow the learning methods to work in mode information-rich data. During the GIST computation,
we have kept the parameters at their default values yielding two fine 8 orientations and one coarse 4
orientations that are computed for 16 equal-sized regions in a 3-channel (RGB) colored image. In this
way, our GIST vectors were generated by having 960 dimensions = 3 × (4 × 4) × (8 + 8 + 4).
As noted before, the vector dimension of HOG descriptors is highly dependent on various factors such
as (a) image size, (b) block size and (c) cell size. In order to create an ML model, we need fixed-sized
vectors with their ground truth GSI scores. Thus, we followed the following procedure. We first
converted the images into the grayscale color space and get the center crops. Second, we resized them
in order to have 256 × 256 wide by utilizing Python Image Library (PIL). In this way, we obtained a
scalable yet efficient workflow for collecting fixed-length feature vectors. Next, we have applied
various cell sizes (i.e 16, 32, 64 pixels) in order to identify which model suits best. For the
hyperparameter of cells_per_block, we have set 2×2 whereas 9 for the orientations. For the block_
norm technique, we have preferred “L2-Sys” scheme.
3.2 Learning to automatically estimate the GSI score
Having the discriminative visual features, our next goal was to find out whether the aforementioned
descriptors provide us a ML based predictable environment for unseen photos of rock masses.
Therefore, following the visual feature extraction stage, we have applied 3 well-known supervised ML
methods namely Random Forests (RF), XgBoost (XGB) and Support Vector Machine (SVM) for GSI
score prediction based on previously seen examples. These methods are used for both classification
(i.e. label assignment) and regression (i.e. numeric estimation) tasks. Nonetheless, in this study, we
have utilized these methods as a regressor.
Briefly, invented by Breiman (2001), Random Forests is an ensemble learning method designed for
classification and regression tasks. By building a diverse collection of decision trees by a random
selection of input features and samples, Breiman (2001) has targeted to avoid the problem of
overfitting which exists in traditional decision trees. In other words, it behaves as a meta estimator that
fits several decision trees on various sub-instances for enhancing the prediction capability along with
reducing overfitting (Bozkir et al. 2020). On the other hand, suggested by Chen and Guestrin (2016),
XGBoost (XGBoost, 2019) has been designed to be a highly efficient, portable and hardware
optimized open-source distributed gradient tree boosting framework. XGBoost scheme runs a parallel
tree boosting algorithm which is applicable to numerous data science problems such as classification,
regression, calculation of feature importances and ranking (XGBoost, 2019). As another method,
Support Vector Machine (SVM), first introduced by Vapnik (1995) attempts to find an optimal
hyperplane between positive and negative classes. For regression oriented tasks, on the other hand,
Support Vector Regression (SVR) has been derived from the conventional SVM classification scheme
for predicting continuous values instead of labeling categories. As a nonparametric method, SVR is an
appropriate regressor which can be employed for linear and nonlinear seperable datasets (Bozkir et al.
2020). It should be noted that SVM can be trained for linear and nonlinear datasets via linear and
radial basis kernels respectively.
As depicted in Fig. 3, following the stage of visual feature extraction from separated training and
testing images, we have created their corresponding comma-separated values (CSV) files in order to
Eurock 2020 – Hard Rock Engineering
6
be further trained via “sklearn” library in Python language. Next, we have coded a Python script in
order to build different ML models belonging to XGBoost, Random Forests, and SVR schemes. Since
XGB and RF discover the nonlinear patterns in data, we have trained SVR algorithm with linear
kernel in order to explore whether the underlying visual features have linear relationships with the
target value (i.e. GSI score). During the training, we have kept the hyperparameters at their default
values except the C (i.e. cost value) parameter in SVR as being set to 100. According to our
observations, on a platform having Intel 8750 CPU and 24 GB 2666 Mhz DDR4 memory, XGB
requires the lowest training time with 10 seconds on average.
3.3 Results of the experiments and findings
In order to assess the effectiveness of the created models, we have measured the obtained results in
terms of different performance metrics such as (a) mean absolute error (MAE), (b) mean squared error
(MSE), (c) root mean squared error (RMSE), (d) coefficient of determination (R2), and accuracy.
Evaluations were carried out on both training (1193 samples) and testing data (295 samples). The
results were listed in Tables 1, 2, 3, and 4 according to the employed descriptor and the dataset
predicted by the related ML model.
Table 1 Performance evaluation of GIST descriptors along with various learning methods on training data
Descriptor
Algorithm
MAE
MSE
RMSE
R2
Accuracy
GIST
XGBoost
0.11
0.022
0.150
0.999
99.69%
GIST
Random Forest
3.13
17.34
4.16
0.944
89.52%
GIST
Linear SVR
10.52
219.57
14.81
0.295
58.72%
Table 2 Performance evaluation of HOG descriptors along with various cells sizes and learning methods on training data
Descriptor – [Mode]
Algorithm
MAE
MSE
RMSE
R2
Accuracy
HOG – 64 px cells
XGBoost
0.11
0.022
0.149
0.999
99.66%
HOG – 32 px cells
XGBoost
0.10
0.016
0.128
0.999
99.68%
HOG – 16 px cells
XGBoost
0.09
0.013
0.118
0.999
99.71%
HOG – 64 px cells
Random Forest
4.29
29.21
5.404
0.906
85%
HOG – 32 px cells
Random Forest
4.47
31.33
5.59
0.899
84.34%
HOG – 16 px cells
Random Forest
4.7
35.14
5.92
0.887
83.36%
HOG – 64 px cells
Linear SVR
11.18
270.6
16.45
0.132
51.39%
HOG – 32 px cells
Linear SVR
10.85
243.4
15.60
0.219
54.88%
HOG – 16 px cells
Linear SVR
8.83
188.2
13.72
0.396
61.48%
Table 3 Performance evaluation of GIST descriptors along with various learning methods on test data
Descriptor
Algorithm
MAE
MSE
RMSE
R2
Accuracy
GIST
XGBoost
7.24
103.86
10.19
0.663
76.36%
GIST
Random Forest
8.2
126.94
11.26
0.588
71.87%
GIST
Linear SVR
11.23
235.04
15.33
0.2387
56.83%
Table 4 Performance evaluation of HOG descriptors along with various cells sizes and learning methods on test data
Descriptor – [Mode]
Algorithm
MAE
MSE
RMSE
R2
Accuracy
HOG – 64 px cells
XGBoost
11.94
215.59
14.68
0.301
61.41%
HOG – 32 px cells
XGBoost
12.33
230.73
15.18
0.252
59.83%
HOG – 16 px cells
XGBoost
13.04
257.61
16.05
0.165
57.00%
HOG – 64 px cells
Random Forest
11.69
209.45
14.47
0.321
60.49%
HOG – 32 px cells
Random Forest
12.34
230.94
15.19
0.251
58.40%
HOG – 16 px cells
Random Forest
12.85
257.13
16.03
0.167
55.61%
HOG – 64 px cells
Linear SVR
13.11
295.83
17.19
0.042
49.54%
HOG – 32 px cells
Linear SVR
13.07
295.38
17.18
0.041
48.88%
HOG – 16 px cells
Linear SVR
12.96
284.97
16.88
0.077
51.66%
For HOG based analysis, we have tested different cell sizes as an attempt to explore the degree of
details and the patch area processed by the descriptor generation algorithm. According to the results
we have obtained the following findings:
Eurock 2020 – Hard Rock Engineering
7
• GIST descriptor along with XGBoost scheme performs the best estimation (i.e. 76.36%
accuracy and 7.24 MAE) on testing data
• Compared the other regressors, XGBoost achieves the best scores. Moreover, it enables us to
fit the training data well (i.e. >99.9% accuracy on training data)
• In all evaluations, GIST outperforms HOG descriptors with a large margin which implies that
the GIST provides much more suitable and discriminative features in the problem domain.
• Reducing the cell sizes in HOG based estimation results better accuracy and lower error which
shows the positive effect of working on smaller patches.
• Linear SVR performs the worst results in terms of all metrics by also clearly indicating that
the dependent variable (GSI) has nonlinear relationships with the revealed visual features.
4 Conclusions
In this study, we attempted to utilize 2 visual features (GIST and HOG) along with ML methods in
order to compute the approximate GSI value of the rock mass when we are only given the photos of
the rock outcrops. In particular, GIST based scheme shows promising results. The accuracy scores
were found as higher than 99% and 76% for training and test images respectively, by using the GIST
descriptors and XGBoost algorithm.
It should be noted that prior works were also conducted for indirect estimation of GSI values (Sonmez
et al., 2003; Poulsen et al., 2015). Nonetheless, in these methods, the parameters were required to be
either computed or defined in advance. Likewise, extra requirements for calculations and definitions
prevent the method from being fully automated and hamper human and computer interaction (Bozkir
et al. 2020).
In this regard, the approach suggested in this study serves promising results towards a fully automated
scheme. Our observations and measurements show that the inference speed of the system is capable of
being run on a mobile cell phone platform too. Furthermore, 11 rock types at 20 outcrops were
implemented. We believe that, as the number of rock types evaluated in the corpus and the number of
GSI definitions evaluated at the training stage increase, the accuracy and generalization capability of
the offered approach will observably increase. As future work, we plan to employ different visual
descriptors in order to gain more robust estimators.
References
Chen T. and Guestrin C. (2016) XGBoost: A Scalable Tree Boosting System. KDD’16, San Francisco, USA,
2016.
Bozkir A. S., Nefeslioglu H. A., Akcapinar Sezer E., Kartal O., Gokceoglu C. (2020) Geological Strength Index
(GSI) determination by local image descriptors and machine learning methods, Human Computer
Interaction, Nova Science Publishers, USA (in press)
Breiman L. (2001) Random Forests, Machine Learning. 45(1):5– 32. 10.1023/A:1010933404324.
Cai M., Kaiser P. K., Uno H., Tasaka Y., Minami M., (2004) Estimation of rock mass strength and deformation
modulus of jointed hard rock masses using the GSI system. International Journal of Rock Mechanics
and Mining Sciences 41(1): 3-19 doi: 10.1016/S1365-1609(03)00025-X
Dalal N., Triggs B. (2005) Histograms of oriented gradients for human detection. IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, San Diego USA
Dalgic F.C., Bozkir A.S., Aydos M. (2018) Phish-IRIS: A New Approach for Vision Based Brand Prediction of
Phishing Web Pages via Compact Visual Descriptors. ISMSIT2018, Ankara, Turkey 2018
Diament I., Bar Y., Geva O., Wolf L., Zimmermann G., Lieberman S., Konen E. (2017) Greenspan H., Chapter
13 - Chest Radiography Pathology Categorization via Transfer Learning. Deep Learning for Medical
Image Analysis. Academic Press, USA
Eroglu E., Bozkir A.S., Aydos M. (2019) Brand recognition of phishing web pages via global image descriptors.
European Journal of Science & Technology Special Issue 2019: 436-443. 10.31590/ejosat.638397
Farahbakhsh E., Chandra R., Olierook H.K.H., Scalzo R, Clark C, Reddy S.M., Muller R.D. (2018) Computer
vision-based framework for extracting geological lineaments from optical remote sensing data.
https://arxiv.org/pdf/1810.02320.pdf. Accessed at 29.10.2019
Hoek E., Brown E.T. (1980) Underground Excavations in Rock” The Institution of Mining Metallurgy, London.
Eurock 2020 – Hard Rock Engineering
8
Hoek E., Brown E.T. (1997) Practical estimates of rock mass strength. International Journal of Rock Mechanics
and Mining Sciences & Geomechanics Abstracts 34:1165–1186.
Hoek E., Marinos P., Benissi M. (1998) Applicability of the geological strength index (GSI) classification for
very weak and sheared rock masses: the case of the Athens schist formation. Bulletin of Engineering
Geology and the Environment 57(2): 151-160 doi: 10.1007/s100640050031
Hoek E. (1999) Putting numbers to geology—an engineer’s view point. Quarterly Journal of Engineering
Geology 32.
Hoek E., Carter T. G., Diederichs M. S. (2013) Quantification of the Geological Strength Index Chart, 47th US
Rock Mechanics/Geomechanics Symposium, San Francisco, CA, USA, 2013.
Marinos P., Hoek E. (2001) Estimating the geotechnical properties of heterogeneous rock masses such as flysch.
Bulletin of Engineering Geology and the Environment 60(2): 85-92 doi:10.1007/s100640000090
Marinos P., Hoek E., Marinos V., (2006) Variability of the engineering properties of rock masses quantified by
the geological strength index: the case of ophiolites with special emphasis on tunnelling. Bulletin of
Engineering Geology and the Environment 65:129-142 doi: 10.1007/s10064-005-0018-x
Marinos V., Marinos P., Hoek E. (2005) The geological strength index: applications and limitations. Bulletin of
Engineering Geology and the Environment 64(1): 55-65 doi: 10.1007/s10064-004-0270-5
Marinos V. (2019) A revised, geotechnical classification GSI system for tectonically disturbed heterogeneous
rock masses, such as flysch. Bulletin of Engineering Geology and the Environment 78:899–912.
Oliva A, Torralba A (2001) Modeling the shape of the scene: A holistic representation of the spatial envelope.
International journal of computer vision, 42(3): 145-175 doi: 10.1023/A:1011139631724
Poulsen B. A., Adhikary D. P., Elmouttie M., Wilkins K. A. (2015) Convergence of synthetic rock mass
modeling and the Hoek-Brown strength criterion. International Journal of Rock Mechanics and Mining
Sciences 80: 171-180 doi: 10.1016/j.ijrmms.2015.09.022
Russo, G., (2007) Improving the reliability of GSI estimation: the integrated GSI-RMi system. ISRM Workshop
Underground Works under Special Conditions, Madrid, 2007
Russo, G. (2009) A new rational method for calculating the GSI. Tunneling and Underground Space Technology
24(1): 103-111 doi: 10.1016/j.tust.2008.03.002
Sonmez H., Ulusay R., (1999) Modifications to the geological strength index (GSI) and their applicability to the
stability of slopes. International Journal of Rock Mechanics and Mining Sciences 36(6): 7432-760 doi:
10.1016/S0148-9062(99)00043-1
Sonmez H., Gokceoglu C., Ulusay R. (2003) An application of fuzzy sets to the Geological Strength Index (GSI)
system used in rock engineering. Engineering Applications of Artificial Intelligence 16(3): 251-269 doi:
10.1016/S0952-1976(03)00002-2
Vapnik V. (1995) The nature of Statistical Learning Theory. Springer, New York, 1995.
XGBoost (2019) XGBoost Documentation. https://xgboost.readthedocs.io/en/latest/. Accessed at 26.10. 2019
Zhang H-B, Zhang Y-X, Zhong B, Lei Q, Yang L, Du J-H, Chen D-S (2019) A comprehensive survey of vision-
based human action recognition methods. Sensors 19 (1005). doi:10.3390/s19051005
