www.geologicacarpathica.sk
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
, OCTOBER 2011, 62, 5, 477—485 doi: 10.2478/v10096-011-0034-7
Introduction
Context and literature review
In geotechnical engineering, it is not uncommon to encounter
problems that are very complex and not well understood. In
most cases, the mathematical models used to solve such prob-
lems attempt to make up for our lack of physical understand-
ing by either simplifying the problem or incorporating several
assumptions. Such models may also be designed to produce
solutions with a specific structure, chosen in advance, which
may be suboptimal for the case at hand. Consequently, many
mathematical models fail to simulate the complex behaviour
of geotechnical engineering problems (Rizzo & Dougherty
1994). Analyses based on the finite element method provide a
better assessment of geotechnical engineering problems, but
these are often costly and time-consuming.
These limitations of mathematical and finite element mod-
els make the problem of landslide prediction exceedingly
difficult. Constitutive programs to determine the stability of
slopes require many free parameters and a large amount of
input data to calibrate. Therefore, practicing engineers prefer
to employ simplified methods to assess subsurface layering
and landslide risk. Such procedures are very useful in the
preliminary design stages of a new project. If the landslide
risk is high, then finite element analysis can be carried out to
determine the stability of the slope with accuracy suitable to
the subsequent design of structures. Artificial Neural Net-
works (ANNs) provide a very different approach to geotech-
nical prediction, with many advantages.
ANNs are model-independent, nonlinear systems capable
of learning physical patterns from a set of input—output data
pairs (Gardner & Dorling 1998). Since their results are based
entirely on the data provided, ANNs neither simplify the
physical nature of the problem nor incorporate any assump-
Predicting subsurface soil layering and landslide risk
with Artificial Neural Networks: a case study from Iran
FARZAD FARROKHZAD
1
, AMIN BARARI
2*
, LARS B. IBSEN
2
and ASSKAR J. CHOOBBASTI
1
1
Department of Civil Engineering, Babol University of Technology, Babol, Mazandaran, Iran
2
Department of Civil Engineering, Aalborg University, Sohnga°rdsholmsvej 57, 9000 Aalborg, Denmark; *ab@civil.aau.dk,
*amin78404@yahoo.com
(Manuscript received November 9, 2010; accepted in revised form March 17, 2011)
Abstract: This paper is concerned principally with the application of Artificial Neural Networks (ANN) in geotechnical
engineering. In particular the application of ANN is discussed in more detail for subsurface soil layering and landslide
analysis. Two ANN models are trained to predict subsurface soil layering and landslide risk using data collected from a
study area in northern Iran. Given the three-dimensional coordinates of soil layers present in thirty boreholes as training
data, our first ANN successfully predicted the depth and type of subsurface soil layers at new locations in the region. The
agreement between the ANN outputs and actual data is over 90 % for all test cases. The second ANN was designed to
recognize the probability of landslide occurrence at 200 sampling points which were not used in training. The neural
network outputs are very close (over 92 %) to risk values calculated by the finite element method or by Bishop’s method.
Key words: Modelling, subsurface soil layering, landslide, Artificial Neural Network.
tions. At worst, the data provided to train the network are bi-
ased. However, ANNs can always be updated to obtain bet-
ter results by presenting additional training examples as new
data become available (Ghaboussi & Sidarta 1998).
Many authors have studied applications of ANNs to geo-
technical engineering. A frequent theme is the prediction of
load capacity based on pile driving data. Ellis et al. (1992) de-
veloped an ANN model for sands based on grain size distribu-
tion and stress history. Chan et al. (1995) developed a neural
network as an alternative to pile driving formulae. Lee & Lee
(1996) utilized neural networks to predict the ultimate bearing
capacity of piles. Teh et al. (1997) used a neural network to es-
timate the static capacity of precast, reinforced concrete piles
with a square section based on dynamic stress-wave data.
Penumadu & Jean-Lou (1997) used neural networks to repre-
sent the behaviour of sand and clay soils.
It should be mentioned that over the last ten years, the au-
thors have undertaken a comprehensive research program us-
ing ANN (Choobbasti et al. 2009; Farrokhzad et al. 2011a,b).
It should also be noted that every ANN is highly specialized,
and entirely dependent on the input data of the particular
project. All aspects of training must be carefully controlled,
and we perform many robustness tests to show that the predic-
tions of the networks are independent of the dataset used.
Research goals
We will now discuss the central aspects of our research in
more detail, as a means of showcasing the broad applicability
of ANNs to geotechnical engineering problems. To this end, a
short introduction to subsurface soil layering is in order.
It is known that the engineering properties of soil vary
from point to point due to the complex and imprecise physi-
cal processes associated with its formation. In contrast, civil
engineering materials such as steel and concrete exhibit far
478
FARROKHZAD, BARARI, IBSEN and CHOOBBASTI
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
greater homogeneity and isotropy. In order to cope with the
complexity of geotechnical behaviour and the spatial vari-
ability of soil, traditional engineering design models have
been justifiably simplified (Jaksa 1995). Soil geology plays
an important role in core material selection (for example, the
construction of rock fill dams) and in geotechnical evalua-
tions preceding the construction of major structures.
Layering is caused by a variety of pedological and geological
processes, and has important consequences for the interpreta-
tion of soils and landscapes. Collecting adequate data on soil
layering (soil classification, thickness of layers) can be difficult
and expensive, sometimes requiring a series of experiments.
Therefore, a method of predicting the structure of subsurface
soil layers from existing data would lead to significant cost re-
ductions in soil geology and improved operational planning.
Accordingly, our first goal in this paper is to develop an
ANN capable of predicting the type and depth of subsurface
soil layers from easily collected geological data. Specifically,
we will use the three-dimensional positions of layers observed
in previously collected boreholes to train an ANN, then test
the network’s capacity to predict soil layers under an arbitrary
location on the surface. The training and testing data were col-
lected from the Mazandaran province of northern Iran.
Our second goal is to train an ANN to correctly predict
landslide risk in the region based on more elaborate geo-
physical data. We will compare the outputs of the network to
landslide risk assessments made using the finite element
method and Bishop’s method.
The remainder of this paper is organized as follows. In
Section 2, we point out some relevant theoretical aspects of
landslides and ANNs. Section 3 describes in detail our meth-
od of designing and training an ANN for soil layer predic-
tion and the performance of the network. Section 4 describes
our ANN for landslide risk assessment, while it is followed
by Section 5 as conclusions.
Theory
Landslides
“Landslides” are mass motions of soil resulting from par-
tially drained shear failures of the surface and subsurface
layers. Relatively small disturbances may trigger landslides,
especially in areas where previous slides have reduced the
residual shear strength of soil masses along their correspond-
ing slip planes (Choobbasti et al. 2008).
The deformation of soil masses generally begins at or
close to the toe of the slope, because this is where the shear
strength of the soil is first exceeded. The deformation then
progresses retrogressively along the shear zones. Movements
of the unstable soil masses take place at points where a large
local strain leads to structural failure. The characteristics of
landslides depend on the geometry of the unstable soil mass,
the pore pressure distribution (which in turn usually depends
on the location of the phreatic surface), the pattern of defor-
mation, and the speed of the movements.
Two common underlying factors are the mechanical prop-
erties of the soil skeleton and the fraction of saturation by
groundwater (Terzaghi & Peck 1967). Applying Terzaghi’s
concept of the effective stress in saturated soil, the stress-
strain behaviours of drained soil and dry soil can be assumed
to be identical. Terzaghi’s approach is detailed below.
The total stress
ij
can be separated into an effective stress
tensor
’
ij
and the pore pressure u as follows:
ij
=
’
ij
+u.
ij
(i, j=1, 2, 3) (1)
where
ij
is the Kronecker delta. The deformation of the soil
skeleton only depends on changes in the effective stresses
’
ij
.
Shear failures occur along the shear planes forming the
boundaries between moving masses of water-saturated dense
soils or rocks. These failures may be caused by the deforma-
tion of partially undrained masses (Bekele et al. 2010). The
word “landslide” is employed by the authors to describe
such failures. The term implies more than deformation of the
masses; one or more blocks of practically solid soil or rock
must actually move (Yesilnacar & Topal 2005). A shear fail-
ure is generally assumed to occur when the average shear
stress along the sliding (or slip) surface is equal to the shear
strength of the soil or rock. The shear strengths can be evalu-
ated by field or laboratory tests (Terzaghi & Peck 1967).
Given that landslides are a consequence of uneven draining
in saturated soils and rocks, they lead to progressive failure of
the structure and erode the residual strength of the layers.
Thus, existing landslides may continue to slip at an average
shear stress considerably less than the drained peak strength of
the soil or rock. The drained peak strength can be measured by
conventional tests (such as triaxial or direct shear tests of the
completely drained material – Yilmaz 2009a).
The first sign of an imminent landslide is the appearance
of surface cracks in the upper part of the slope. These cracks
are perpendicular to the direction of movement. They may
gradually fill with water, reducing the effective normal stress
in the soil along the shear planes. This process may further
degrade the soil’s resistance. The speed of a landslide during
the primary failure is controlled mainly by the nature of the
materials involved and the overall shape (including steep-
ness) of the failure surface.
With respect to the materials involved, it is known that soil
with more loosely packed particles has a larger difference
between the drained and undrained strengths. This is true for
both sands and clays. When saturated, these materials are
also known as “quicksand” and “quick clay”; even a small
disturbance can cause their structures to collapse. Both types
of soils are characterized by liquefaction failure (Lee et al.
2004).
In this paper, both subsurface layering and landslide anal-
ysis are assessed using artificial neural networks. Before
continuing, we will therefore briefly introduce this form of
modelling.
Artificial Neural Networks
Neural networks are intuitively appealing structures based
on a crude, low-level model of biological neural systems.
Even a simple network is capable of modelling extremely
complex nonlinear functions. They also excel at bypassing
the “dimensionality problem” that bedevils other forms of
479
PREDICTING SUBSURFACE SOIL LAYERING AND LANDSLIDE WITH NEURAL NETWORKS
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
non-linear modelling, which often require a very large num-
ber of variables.
Neural networks learn by example. The researcher gathers
representative data, then invokes a training algorithm to find
a network structure that reproduces the data. A typical feed-
forward network is shown in Fig. 1. The nodes are arranged
in layers, with any given node transmitting its information to
all nodes in the layer below. The input layer serves only to
introduce the values of the input variables. The output layer
is used to classify the data.
When the network is presented with data, the input vari-
able values are placed in the input nodes. These values are
used to generate signals in the first hidden layer, and then the
hidden values are used to generate the output layer. The net-
work calculates the signal for a hidden node or output node
by performing a weighted sum of the outputs from units in
the preceding layer and subtracting a threshold. This value is
then passed through an activation function to produce the
outputs of the node. The output layer signals are therefore a
result of the entire network (Farrokhzad et al. 2011b).
Predicting subsurface soil layering with ANNs
Data collection
Babol (Fig. 2), a city in the Mazandaran province of north-
ern Iran, is our study area for subsurface soil layering. The
city is located approximately 20 kilometers south of the Cas-
pian Sea, on the west bank of the Babolrood River. The re-
gion receives abundant annual rainfall. Babolrood has two
types of river terraces, denoted H
1
and H
2
. H
1
terraces have a
low level surface level, with heights of 1 to 2.5 m and widths
of 0 to 150 m. In some sections of the river, these terraces
mark the boundary of the active (yearly) flood plain. In other
sections, they represent an alternative flood plain (diacritic
of active-alternative flood plain from 20-year flood plain).
Fig. 1. Structure of Artificial Neural Network.
Fig. 2. Map of areas under study.
H
1
terraces consist of fine-grained and unconsolidated allu-
vial materials. H
2
terraces are referred to as river terraces.
They have higher surface levels, 4 to 6 m, and support com-
pact vegetation. They consist of Aeolian deposits (loess).
This research makes use of data from 40 borehole logs col-
lected in the study area by a different institute of geotechni-
cal engineering and seismology, for different research
purposes. The boreholes cover a region of about 7.8 square
kilometers, separated into five zones by imaginary lines
(Fig. 3). Thirty of the logs, with depths of 10 m to 30 m,
were used to train and test our ANN (Fig. 4).
The properties of the soil layers were determined by field ex-
amination and by laboratory index testing of benchmark soils
following standard procedures. During the survey, we bored
many shallow holes, examined and classified the samples, and
created a soil map of the region. The samples were tested in the
laboratory to determine their grain size distribution, plasticity,
and compaction characteristics (Allen & Tadesse 2003). Ta-
ble 1 shows the estimated properties of the soil types, including
grain size distributions and Atterberg limits, engineering classi-
fications, and the physical properties of the major layers.
The soil layer ANN
To train the Artificial Neural Network, we need a set of
known input and output pairs (Basheer 1998). The available
input—output pairs are usually divided into two sets. The
learning or training set is used to determine the connection
weights (w
k
ij
)
1
(Agrawal et al. 1997) in the network. The test-
ing set is used to measure the performance of the neural net-
work after training. In this study, 70 % of the data were used
for training and 30 % were used for testing. We tried other
ratios, but this one yielded the best results.
1
'
1
1
(
)
(
)
k
n
k
k
k
i
i
ij
i
j
y
f y k
f
w y
∑
1
480
FARROKHZAD, BARARI, IBSEN and CHOOBBASTI
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
A case study was conducted to evaluate the usefulness of
the ANN approach and determine the optimal structure and
algorithm. The inputs to the network were a set of soil for-
mative environmental factors; the output was a set of simi-
larity values mapped to soil classes. The classes are those
defined by the Unified Soil Classification System (USCS)
based on the particle size distribution and other properties
that affect their usefulness as construction materials.
We tested this network’s performance under five supervised
learning (training) algorithms: back-propagation, conjugate
gradient descent, Levenberg-Marquardt, quick propagation,
and delta-bar-delta. Each training method yielded different re-
sults. In this case, performance is defined as the root-mean-
squared (RMS) error of the output patterns produced by the
network on the testing dataset.
The RMS values shown in Table 2 are based on a back-
propagation neural network with one hidden layer. Back-
propagation is the best known training algorithm for neural
networks, and is still one of the most useful. It has a lower
Fig. 3. Plan of some borehole places in 7.80 km
2
in western part of Babol.
Fig. 4. Representative of soil profiles in study area.
481
PREDICTING SUBSURFACE SOIL LAYERING AND LANDSLIDE WITH NEURAL NETWORKS
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
Table 1: Samples of data collected from western part of Babol.
Table 2: Results of research in order to Learning/training algorithm
selection.
Table 3: Results of research in order to select the number of neu-
rons in the hidden layer.
Fig. 5. Network architecture for predicton of subsur-
face soil layering in Babol.
memory requirement than most algorithms, and
usually reaches an acceptable level of error very
quickly. However, it can be very slow to con-
verge to the error minimum.
We tested networks with 3, 5, 7, 8, 9, 10, 11,
12 neurons in the hidden layer. Table 3 shows
the results of this comparison. The network with
8 hidden neurons produced the smallest RMS er-
ror in this case study. This network’s architec-
ture is shown in Fig. 5.
The training phase is divided into epochs. In
each epoch, every input—output pair in the train-
ing set is fed through the network once and used
to adjust the network weights and thresholds.
After each epoch the state of the network and the
performance of the network on the testing set are
recorded, so that if over-learning occurs the best
network discovered during training can be re-
trieved. The number of epochs required to opti-
mize the network in this research varied between
100 and 500. It is good practice to “shuffle” the
training set in each epoch, presenting the input—
output pairs to the network in a random se-
quence. (If the sequence never changed, the
network might perform better for pairs at the end
of the epoch.)
In addition to the primary choices described
above, back-propagation models can be tuned us-
ing two parameters. The learning rate controls the
size of the weight changes made by the training
algorithm. A larger learning rate may lead to fast-
er convergence, provided that the error surface is
not too noisy (Penumadu & Zhao 1999). In the
present work, we select a learning rate of 0.5. The
momentum parameter causes the back-propaga-
tion algorithm to “pick up speed” if consecutive
training pairs push the weights in the same direc-
tion. This property helps the network reach a glo-
bal optimum by preventing it from getting
trapped in local minima of the parameter space.
Supervised
learning/
training
algorithms
Back
propagation
Conjugate
gradient
descent
Levenberg–
Marquardt
Quick
propagation
Delta-
bar-
Delta
RMS (%)
8 11
13.3
10.7
11.5
Number of neurons in the
hidden layer
3 5 7 8 9 10 12
RMS (%)
14 11.7 10.9 7.5 9.8 12.1 13.3
482
FARROKHZAD, BARARI, IBSEN and CHOOBBASTI
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
Results of the soil layer ANN
Figure 6 presents the ANN’s predictions for the thickness
of subsurface soil layers in three boreholes, and the actual
layers obtained from digging the boreholes. As mentioned in
the previous section, soils from the Babol study area were
classified into four types: gravel, sand, silt and clay. In the
case study, the input variables were simply the subsurface
coordinates (x, y, z). For a given value of x and y, the net-
work yields different soil types for different values of z
(depth). The neural network results are very similar to the
real data. In fact, the ANN model properly predicts the types
of soil present in all three boreholes. The predicted thick-
Fig. 6. a, b, c – Comparison between predicted results of ANN and soil profiles obtained from test
boreholes.
Fig. 7. Errors in the ANN for prediction of subsurface soil layers in test samples obtained from test boreholes.
nesses of the three layers are 1280 cm, 330 cm and 150 cm.
The corresponding actual thicknesses are 1300 cm, 300 cm
and 160 cm.
Figure 7 plots the ratio of the predicted values to the actual
values based on soil investigation for all samples in the test-
ing set. If the predicted value and the true value are similar in
all cases, then the points should all lie near the line y = 1. The
average correlation of the ANN predictions with the true
data in all cases is over 90 %. Thus, it can be concluded that
the ANN method can accurately predict the depths of subsoil
layers. This ANN is a viable prediction tool that can assist
geotechnical engineers in making accurate and realistic pre-
dictions of the subsoil structure.
An assessment of
landslide risk in Shirgah
Shirgah (Fig. 2) is located in
the north of Iran, in Mazanda-
ran province. The annual mean
temperature of the terrain is
12.5 °C, and the annual mean
precipitation is about 800 mm.
The climate (from the Dommar-
tan method) is humid. From a
geological point of view, most
of the units in this region are re-
lated to the Cenozoic Era. Marl,
shale and silty stone are there-
fore prevalent, and the region is
susceptible to landslides.
Before assessing risk on this
slope, we dug several boreholes
to obtain essential information
such as details of the strata,
moisture content, and the stand-
ing water level. Piezometer
tubes were installed in the
ground to measure changes in
water level over time. These
field investigations also includ-
ed in situ and laboratory tests,
photographs, a study of geolog-
ical maps and memoirs to indi-
cate probable soil conditions,
483
PREDICTING SUBSURFACE SOIL LAYERING AND LANDSLIDE WITH NEURAL NETWORKS
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
visiting and observing the slope, examining records of previ-
ous instabilities, and plotting the topography (Fig. 8).
As with the case study described in Section 3, we tested sev-
eral network structures, training parameters, and other criteria to
find the ANN model that was most suitable for predicting land-
slide hazard in the region. After a number of trials, the best-per-
forming ANN had four layers: an input layer of 8 neurons, an
output layer of 2 neurons, and two hidden layers. All other as-
pects of the ANN model are identical to that used in Section 3.
In selecting the input variables (Fig. 9a,b) we hope to recog-
nize the conditions which caused the slope to become unstable
Fig. 8. Topography of study area and unstable zone.
Fig. 9. a – Network architecture, b – Sample’s coordinate and Stable and Unstable sample description.
and the processes which triggered that movement. If the net-
work can achieve an accurate diagnosis, it will be possible to
understand the landslide mechanisms and propose effective
remedial measures. In the proposed model, the input parame-
ters are the coefficient of cohesion (C), the angle of the slope
( ), the angle of internal friction ( , distance from the slope
edge (X), the unit weight ( ), the slope elevation (H), the ef-
fective stress (
’), and the length of the slope (L). The output
variable is the slope stability.
In this research, we assume that the horizontal acceleration
of an earthquake is constant. During an earthquake, the hori-
zontal acceleration on the
blocks may cause instability.
However, this acceleration
also reduces the normal
stresses on the contact plane
and hence the friction to
shear along the plane. The
contribution of the cohesion
coefficient to shear strength
may be real (due to cementa-
tion) or apparent (due to as-
perities on the discontinuity
plane). During an earthquake
cementation may be broken,
and asperities may be broken
or overridden leading to non-
fitting roughness patterns. In
all cases cohesion and fric-
tion are permanently reduced.
Hence, an earthquake not
only adds unfavourable forc-
es to a slope but may also
permanently reduce the shear
strength along discontinuity
planes (Yilmaz 2009b).
The ANN models devel-
oped in this research were
used to predict the slope sta-
bility for 7 sections (labeled
) in our study area
(Fig. 8). Approximately 1000
input—output pairs, including
inputs and outputs, were col-
lected for the prediction of
landslide risk. Among these
data, 80 % were used for
training and 20 % were used
to validate the ANN.
The data cover a wide
spectrum of soil Fish pa-
rameters. In order to test the
performance of the ANN,
21 random cases were se-
lected from the testing set
for detailed predictions us-
ing the finite element meth-
od or Bishop’s method (see
Table 4).
484
FARROKHZAD, BARARI, IBSEN and CHOOBBASTI
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
Figures 10 and 11 show the accuracy of the testing set
(200 sampling points). In all cases the accuracy of the ANN
model is over 92 %.
Conclusion
In this investigation, we developed and trained two ANN
models to demonstrate the suitability of the method for prob-
lems in geotechnical engineering. The first ANN predicts
subsurface layering in the Babol region of Iran, and the sec-
ond assesses landslide risk based on the physical characteris-
Table 4: Selected case studies showing the actual modes of failure and prediction of the ANN.
Fig. 11. Prediction of FOS versus true value.
Fig. 10. Accuracy of ANN in prediction of slope stability.
Sample
C
(kPa)
L
(m)
(°)
X
(m)
kN
m
3
(
)
kN
m
2
(
)
H
(m)
Stability
(Actual
condition)
FOS
(Actual
condition)
Stability
(Prediction
of ANN)
FOS
(Prediction
of ANN)
-1
13
45
35
19
2
21.2
15.96
12
instable
0.68
instable
0.72
-2
13
45
35
19
5
21.2
39.91
12
instable
0.73
instable
0.79
-3
13
45
35
19
10
21.2
79.82
12
instable
0.80
instable
0.82
-1
13
40
35
19
2
21.2
15.96
10
instable
0.77
instable
0.83
-2
13
40
35
19
5
21.2
39.91
10
instable
0.68
instable
0.69
-3
13
40
35
19
10
21.2
79.82
10
instable
0.71
instable
0.77
-1
10
63
18
22
2
20
6.62
15
instable
0.91
instable
0.86
-2
10
63
18
22
5
20
16.57
15
stable
1.03
instable
1.8
-3
10
63
18
22
10
20
33.14
15
stable
1.10
instable
1.09
-1
13
72
22
19
2
21.2
9.21
9
instable
0.91
instable
0.97
-2
13
72
22
19
5
21.2
23.02
9
stable
1.10
instable
1.02
-3
13
72
22
19
10
21.2
46.05
9
stable
1.16
instable
1.05
-1
10
58
22
22
2
20
8.24
11
instable
0.87
instable
0.80
-2
10
58
22
22
5
20
20.60
11
instable
0.89
instable
0.92
-3
10
58
22
22
10
20
41.21
11
instable
1.04
instable
1.03
-1
10
40
36
22
2
20
14.82
5
instable
0.76
instable
0.85
-3
10
40
36
22
5
20
37.05
5
instable
0.81
instable
0.84
-3
10
40
36
22
10
20
74.10
5
instable
0.82
instable
0.84
-1
13
35
38
19
2
21.2
17.81
12
instable
0.72
instable
0.68
-2
13
35
38
19
5
21.2
44.53
12
instable
0.76
instable
0.71
-3
13
35
38
19
10
21.2
89.06
12
instable
0.75
instable
0.76
485
PREDICTING SUBSURFACE SOIL LAYERING AND LANDSLIDE WITH NEURAL NETWORKS
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA
GEOLOGICA CARPATHICA, 2011, 62, 5, 477—485
tics of the slope and soil. Both neural networks provide very
accurate results. The first network was tested on hand-col-
lected borehole data, and the second network was compared
to two classical mathematical models of landslide risk. The
trained neural networks developed in this study are capable
of predicting subsurface soil layering, the stability of slopes,
and landslide risk in a specific study area with an acceptable
level of confidence.
References
Allen A. & Tadesse G. 2003: Geological setting and tectonic subdi-
vision of the Neoproterozoic orogenic belt of Tuludimtu, west-
ern Ethiopia. J. African Earth Sci. 36, 1—2, 329—343.
Agrawal G., Chameau J.A. & Bourdeau P.L. 1997: Assessing the
liquefaction susceptibility at a site based on information from
penetration testing. In: Kartam N., Flood I. & Garrett J.H.
(Eds.): Artificial neural networks for civil engineers: funda-
mentals and applications. New York, 185—214.
Basheer I.A. 1998: Neuromechanistic-based modeling and simula-
tion of constitutive behavior of fine-grained soils. Ph.D. Dis-
sertation, Kansas State University, Manhattan, KS.
Bekele A., Dramis F., Fubelli G., Umer M. & Asrat A. 2010: Land-
slides in the Ethiopian highlands and the Rift margins. J. Afri-
can Earth Sci.
56, 1, 131—138.
Chan W.T., Chow Y.K. & Liu L.F. 1995: Neural network: An alter-
native to pile driving formulas. J. Computers and Geotechnics
17, 135—156.
Choobbasti A.J., Barari A., Safaei M. & Farrokhzad F. 2008: Miti-
gation of Flourd landslide using non-woven Geosynthetic. Bul-
garian Geol. Soc. Rev. 69, 1—3, 49—56.
Choobbasti A.J., Farrokhzad F. & Barari A. 2009: Prediction of
slope stability using artificial neural network – case study:
Noabad, Mazandaran. Arabian J. Geosci. 4, 311—319.
Ellis G.W., Yao C. & Zhao R. 1992: Neural network modeling of
the mechanical behavior of sand. Proc. Engineering Mechanics,
ASCE, 421—424.
Farrokhzad F., Choobbasti A.J., Barari A. & Ibsen L.B. 2011a: As-
sessing landslide hazard using artificial neural network: case
study of Mazandaran, Iran. Carpathian J. Earth Environ. Sci.
6, 251—261.
Farrokhzad F., Choobbasti A.J. & Barari A. 2011b: Liquefaction mi-
crozonation of Babol city using artificial neural network. J. King
Saud Univ. (Sci.). Doi:10.1016/j.jksus.2010.09.003, in print.
Gardner M.W. & Dorling S.R. 1998: Artificial neural networks
(The multilayer perceptron) – A review of applications in the
atmospheric sciences. Atmospheric Environment 32 (14/15),
2627—2636.
Ghaboussi J. & Sidarta D.E. 1998: New nested adaptive neural net-
works (NANN) for constitutive modeling. J. Computers and
Geotechnics 22, 1, 29—52.
Gomez H. & Kavzoglu T. 2005: Assessment of shallow landslide
susceptibility using artificial neural networks in Jabonosa River
Basin, Venezuela. Engineering Geol. 78 (1—2), 11—27.
Jaksa M.B. 1995: The influence of spatial variability on the geo-
technical design properties of a stiff, overconsolidated clay.
Ph.D. Thesis, Univ. Adelaide, Adelaide.
Lee I.M. & Lee J.H. 1996: Prediction of pile bearing capacity using
artificial neural networks. Computers and Geotechnics 18, 3,
189—200.
Lee S., Ryu J.H., Won J.S. & Park H.J. 2004: Determination and ap-
plication of the weights for landslide susceptibility mapping using
an artificial neural network. Engineering Geology 71, 289—302.
Penumadu D. & Jean-Lou C. 1997: Geomaterial modeling using ar-
tificial neural networks. Artificial Neural Networks for Civil
Engineers: Fundamentals and Applications, ASCE, 160—184.
Penumadu D. & Zhao R. 1999: Triaxial compression behavior of
sand and gravel using artificial neural networks (ANN). J. Com-
puters and Geotechnics 24, 207—230.
Rizzo D.M. & Dougherty D.E. 1994: Application of artificial neural
networks for site characterization using hard and soft informa-
tion. Peters A. et al. (Eds.): Proceedings of the 10th Interna-
tional Conference on Computational Methods in Water
Resources. Kluwer Academic, Dordrecht 12, 793—799.
Teh C.I., Wong K.S., Goh A.T.C. & Jaritngam S. 1997: Prediction
of pile capacity using neural networks. J. Computing in Civil
Engineering, ASCE, 11, 2, 129—138.
Terzaghi K. & Peck R.B. 1967: Soil mechanics in engineering prac-
tice. John Wiley & Son, Inc., New.
Yesilnacar E. & Topal T. 2005: Landslide susceptibility mapping: a
comparison of logistic regression and neural networks method
in a medium scale study, Hendek region (Turkey). Engineering
Geology 79, 251—266.
Yilmaz I. 2009: A case study from Koyulhisar (Sivas-Turkey) for
landslide susceptibility mapping by artificial neural networks.
Bull. Engineering Geol. Environment 68, 3, 297—306.
Yilmaz I. 2009: Landslide susceptibility using frequency ratio, lo-
gistic regression, artificial neural networks and their compari-
son: a case study from Kat landslides (Tokat-Turkey).
Computers and Geosciences 35, 6, 1125—1138.