You are on page 1of 18

Environmental Earth Sciences (2022) 81:487

https://doi.org/10.1007/s12665-022-10590-8

THEMATIC ISSUE

Deep learning models for large‑scale slope instability examination


in Western Uttarakhand, India
Vishnu Himanshu Ratnam Pandey1 · Ashutosh Kainthola1   · Vikram Sharma2 · Abhishek Srivastav1 · T. Jayal3 ·
T. N. Singh4

Received: 15 November 2021 / Accepted: 18 September 2022 / Published online: 3 October 2022
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2022

Abstract
Slope failures are avoidable accidents in most of the scenarios. The eventuality of a failure leads to loss of lives and destruc-
tion, especially in hilly areas. Investigation, analysis and prediction of slope failure is a reliable approach to avert such
mishaps. Hence, the present research work delves into the prediction of landslides and slope failures through numerical
simulation and a deep learning approach. Field attributes and laboratory-tested strength data from Lower Tons Valley, North-
ern India has been taken as a case study. Initially, a total of 185 slope models were simulated in a finite difference code by
varying four slope parameters, namely, slope angle, slope height, cohesion and angle of internal friction. These simulated
results were further divided into two parts, one part with 148 datasets for the training of models and other part consisting
of 37 datasets for testing of models. Two artificial neural network prediction models, along with a conventional multi-linear
regression model was developed and their accuracy was accessed. The developed neural network models superseded the
conventional model, in terms of performance and accuracy, as shown by statistical approaches R2 and mean squared error
values. Moreover, the neural network model with Adam optimizer achieved higher statistical accuracy than the one with
stochastic gradient descent optimizer. However, all these deep learning models demonstrate significant performance, and
can be used by geo-engineers for swift prediction of safety factors for excavated slopes in the study area.

Keywords  Slope stability · Artificial neural network · Multi-linear regression · Finite difference method · Factor of safety

Introduction

Early human societies settled along fertile river plains, which


provided them with water and agricultural produce (Harari
2014). However, the rapid rise in human population led to a
This article is part of a Topical Collection in Environmental Earth
crisis of habitable space and eventually it led to the move-
Sciences on “Geoenvironmental resources and risks: management
strategies and techniques for eco-sustainable development and ment of people towards rugged mountainous terrain. Steep
planning of fragile territories”, guest edited by Gioacchino hill slopes are rapt with instability issues, ranging from ero-
Francesco Andriani, Domenico Calcaterra, Juan Carlos Mora sion to slope failures (Shariati and Fereidooni 2021). These
Chaparro, Ratan Das, Lidia Loiotine, Michael Tarullo.
instances are further aggravated due to improper settlement
* Ashutosh Kainthola and unplanned excavations. Each year slope failures claim
ashu.py@zohomail.in several lives and cause huge economic damages (Cui et al.
2014). Slope failures and landslide incidents could be miti-
1
Department of Geology, Banaras Hindu University, gated or avoided with investigation, stability assessment,
Varanasi, Uttar Pradesh, India
stabilization and regular monitoring. Estimation of slope
2
Department of Geography, Banaras Hindu University, instability depends on rock/soil type, discontinuity distri-
Varanasi, Uttar Pradesh, India
bution, topography, hydrological environment and stress
3
State Environment Conservation and Climate Change regime (Hoang and Pham 2016; Qi and Tang 2018). The
Directorate, Dehradun, Uttarakhand, India
likelihood of failure is usually measured on the basis of
4
Department of Earth Sciences, IIT Bombay, Mumbai, the strength parameters of the slope material (Choobbasti
Maharashtra, India

13
Vol.:(0123456789)
487 
Page 2 of 18 Environmental Earth Sciences (2022) 81:487

et al. 2009; Qiu et al. 2017). These strength parameters are Therefore, in the present research work an attempt has
assessed through laboratory testing of the representative been made to predict the stability of slope through artifi-
samples and field investigation. The resulting data is either cial neural networks (ANNs), together with a conventional
analysed mathematically or fed into sophisticated numerical multiple linear regression (MLR) model. The input data
simulation tools for deciphering the factor of safety. In most viz., slope height, slope angle, angle of internal friction,
instances in the Himalayan terrain, the slope failure zones and cohesion were evaluated from the field visit and subse-
are not localized but extend over several kilometres and car- quent laboratory testing of the soil mass. These geometric
rying out simulation over the long stretch can be cumber- and strength parameters were used to numerically evalu-
some. Forecast of existing and future instability, owing to ate the factor of safety (FoS). Some of the test results have
the change in ground conditions adds more variability to the already been published (Kainthola et al. 2021). For the pre-
analysis. Therefore, while doing the stability analysis spe- sent work, a total of new 185 slope models were simulated
cial consideration is needed to include the changes in slope using a finite difference code. Afterwards, the variables:
geometry and variable geotechnical dimensions. slope height, slope angle, cohesion, angle of internal fric-
Conventional stability assessment methods employ the tion and factor of safety were analyzed  for correlation and
approximation of driving and resisting forces at a prede- prediction through MLR and neural network algorithms.
fined surface of failure to ascertain the overall stability of All the developed machine learning models were intuitively
the slope (Zheng et al. 2006; Raghuvanshi 2019). On the combined and correlated to support the study. The multiple
contrary, the numerical simulation uses the discretization linear regression model is an advanced form of simple linear
and solution of the partial derivatives of the problem to yield regression, which relates several independent parameters to
these results (Singh et al. 2008; Li et al. 2009). Numerical a dependent parameter (Eberly 2007). Moreover, the ANNs
tools possess an advantage over conventional methods as are a kind of deep learning prediction models capable of
they can calculate the deformation and no prior assumption identifying the intrinsic pattern in the training datasets to
of failure surface is entailed (Zou et al. 1995; Tiwari et al. predict the outputs.
2020; Hazari et al. 2021). However, numerical simulation
tools are expensive and require an admirable amount of
sophisticated computing to function efficiently (Zheng et al.
2014). Furthermore, the input parameters require exhaustive Previous works
testing of the geological material, which is also expensive.
In the recent past, sophisticated artificial intelligence (AI) Computers are arguably the most ingenious creation of
tools have become quire prevalent for carrying out classi- human mind and have played a dominant role in the techni-
fication and prediction for large datasets (Taha et al. 2018; cal advancement of the present world. The successful execu-
Benbouras et al. 2019; Lawal and Kwon 2020). Machine tion of the first computer program, evoked people’s curiosity
learning is one such facet of AI, where models after being for a time when machines will become self-aware and auto-
trained adequately, can predict the desired output to meet driven (Turing 2009; Aiello 2016). However, after that, it
high accuracy (Suman et al. 2016; Singaravel et al. 2018; took approximately a century to develop the idea of artifi-
Feng et al. 2018). Use of neural network, decision trees cial intelligence (AI) and machine learning (ML) methods.
and support vector machines are among the most popular The recent growth of these techniques can be attributed to
tools for prediction (Ma et al. 2017; Pham et al. 2018; Dou inexpensive and swifter computing accessories with higher
et al. 2019). Anitescu et al. (2019) explored the capabilities storage and processing capabilities. Furthermore, copious
of ANN, in line with the collocation method to frame an amounts of data generated from daily human activities are
entirely mesh-free technique for evaluating partial deriva- pumped as fuel in this industry. The major encouragement
tives and examining boundary conditions. The developed for the advancement of these techniques came from indus-
neural network turned out to be robust and swift PDEs tries, which were keenly interested towards insightful data-
approximation, and showed good results for enumerating driven decision models for their better investment.
inverse acoustic problems. Later, Samaniego et al. (2020) Machine learning tools such as ANN, support vector and
devised a deep neural network model to find the solutions to random forest are getting wide acceptance throughout the
partial differential equation, satisfying boundary value con- scientific community in different areas of interest. It is due to
ditions and further effectively applied them to problems of its ability to solve complex problems in intelligible, swifter
computational mechanics. The biggest advantage in favour and reasonable manner with higher accuracy in its own
of ML tools is they are free of cost and several libraries learning experiences. Geotechnical engineers and research-
exists for their usage (Erickson et al. 2017; Stančin and Jovic ers are already using machine learning techniques to deepen
2019; Khurma et al. 2020). their understanding of slope stability, underground excava-
tion and other civil infrastructure works. Factor of safety

13
Environmental Earth Sciences (2022) 81:487 Page 3 of 18  487

(FoS) is a term used to define the stability condition of a on preliminary field investigation, samples collection,
slope. FoS is the ratio of the shear strength to shear stress sophisticated laboratory testing and finite difference stabil-
on a slope. In turn, the shear strength of the material relies ity assessment. Several slope failure events were recorded
on the cohesion and angle of internal friction of the slope in the initial survey of the study area and numerous critical
material. Furthermore, the slope angle and height bear a slopes were studied (Fig. 2). A total of 185 slope geometry
direct relationship to the ensuing shear strength parameters. parameters were ascribed in field and GIS investigation; and
To examine the stability condition of slopes through tradi- shear strength parameters were obtained from the laboratory
tional methods; workers and practitioners have to scrupu- testing. These input parameters were fed into the finite dif-
lously determine slope geometry, cohesion, angle of internal ference tool to establish the factor of safety of every slope.
friction, young’s modulus, unit weight, and complex geology Afterwards, the multi-linear regression (MLR) analysis and
of the groundmass. ANN can use the existing knowledge to artificial neural network (ANN) techniques were imple-
make predictions; and not rely on fresh tests and simulation. mented on these inputs (Slope geometry and shear strength
Previously, workers have discerned the FoS of slopes properties) and output (Factor of Safety) dataset generated
using ANN and exemplified the comparison of predicted via finite difference technique. Both of these illustrative
outcomes with the results from other available techniques machine learning tools were supported by several open-
(Lu and Rosenbaum 2003; Pradhan and Lee 2009; Ray source libraries, available on R and Python platforms.
et al. 2020). Chakraborty and Goswami (2017) worked on
the comparative slope failure assessment between ANN and Data distribution
multi-linear regression approaches, and both methods have
been discussed in great detail. Sakellariou and Ferentinou The choice of parameters for the slope instability analysis
(2005) presented an intelligent ANN model for the deter- is contingent upon the employed technique. Slope’s shear-
mination of FoS and also dwells into the relative contribu- parameters and geometry have been widely used to under-
tion of input parameters for the enumeration of the safety stand landslide dynamics (Zhang and Xiao 2019; Bui et al.
factor. A dump slope stability prediction of coal open pit 2020; Tiwari et al. 2020; Kainthola et al. 2021; Chen et al.
mines using artificial neural networks have been pictured 2021a, b). The effect of the distribution of drained shear-
into space by Rahul et al. (2015). A sophisticated and robust strength on failure has also been detailed by Stark et al.
machine learning technique called neuro-fuzzy system eased (2005). The limit-equilibrium stability analysis has revealed
the prediction of Young’s modulus of rocks by feeding the a strong impact of cohesion, angle of internal friction, and
basic physico-mechanical inputs such as point load, density, slope’s geometry on the size of failed zone and dynamics
and water absorption (Cüneyt Aydin et al. 2006; Singh et al. (Frattini and Crosta 2013). Soil cohesion, weighted as the
2012; Armaghani et al. 2015). These workers have presented most prominent contributor in factor of safety, among the
excellent and incisive machine learning techniques to deal other applied parameters like height, slope, angle of inter-
with a particular problem in their research area, however, nal friction, and density, etc. (Gelisli et al. 2015). A hybrid
they have not undertaken all the possible predictor variables machine learning safety factor prediction model was devel-
which may alter the results significantly. Moreover, there is oped by Verma et al. (2016), which considered cohesion,
a need for a universal deep learning model to comply with angle of internal friction, slope angle, and pore water pres-
all kinds of geotechnical problems in every environmental sure. Guo et al. (2021), through extended—discontinuous
condition. deformation analysis demonstrated the reduction in cohe-
sion being the prime cause of mass-movement initiation at
Fukuchi-machi, Japan. Moreover, several published works
Methodology on susceptibility models observed the dominance of slope
angle, over several other landslide conditioning factors in
The manuscript details its findings on geotechnical investi- failure probability (Pradhan and Lee 2010; Pradhan 2011;
gation and deep learning approaches on the road cut slopes Abraham et al. 2021; Sardooi et al. 2021). Rukhaiyar et al.
along the Tiuni to Kalsi via Koti motorway. The study area (2018), emphasized the role of cohesion, internal friction
belongs to Lower Tons valley, and is a part of Garhwal angle, slope angle, slope height in FOS prediction model.
Himalayan division (Fig. 1). Damta and Jaunsar Groups Also, the size of the mass-movement is mostly controlled
cover the major section of the study area, and some parts by the initial slope angle and the shear values (Chen et al.
are composed of Tejam, Sirmur and Siwalik Groups (Auden 2016). As the algorithms calculates the safety factors using
1934; Valdiya 1980). The Himalaya is known for its rugged Mohr–Coulomb failure criteria, a major emphasis of the pre-
topographic terrain and complex geological setup, further diction models in the present analysis is on geometry and
aggravated by previous and ongoing tectonic plate motion shear strength of slope mass (Lollino and Andriani 2017).
(Srivastav et al. 2021). The research builds its foundation

13
487 
Page 4 of 18 Environmental Earth Sciences (2022) 81:487

Fig. 1  Geological and tectonic setup of the study area (after Thakur and Rawat 1992)

The ingredients of FoS like the angle of internal friction and


( )
ResistingForce (W ∗ cos𝜃 ∗ 𝜇) + cA
cohesion are important to resist forces, while the height and FOS = = (1)
DrivingForce W ∗ sin𝜃
slope angle will influence the skidding forces.
Therefore, the four most obvious and critical parameters where W is the weight of the sliding block, θ is the slope
viz., slope height, slope angle, cohesion and angle of internal angle, μ is the coefficient of friction, c is the cohesion and A
friction have been taken for the statistical and deep learning is the area occupied by the sliding block.
assessment of slope failure. Cohesion and angle of internal The selection of the four parameters stems from their
friction were discerned in the laboratory through direct shear pertinence and ease of applicability. Any change in either
testing, while the geometrical factors were adopted based on slope geometry or shear strength can swiftly be incorpo-
the topography of the study area. The cohesion and angle rated into the model. A preliminary geo-statistical analy-
of internal friction are the intrinsic properties of soil mass, sis of each geotechnical parameters involved in the pre-
whereas the driving force (W*sinθ) are the significant attrib- sent work shed light on their distribution and correlation
utes of slope height and slope angle (Eq. 1) (Fig. 3). (Fig. 4). The correlation coefficient (r) between each of
these properties helps to develop a mathematical associa-
tion of one parameter to other parameters (Asuero et al.
2006). Slope height has a correlation coefficient of − 0.14,

13
Environmental Earth Sciences (2022) 81:487 Page 5 of 18  487

Fig. 2  Evidences of slope failure in the study area

Fig. 3  Critical factors affecting slope failure

0.18 and 0.04 with slope angle, cohesion, and angle of


internal friction, respectively. Slope angle shows a weak
correlation coefficient of 0.04 and 0.03 with cohesion and
angle of internal friction, respectively. Moreover, cohe-
sion and angle of internal friction have a weak negative
r value of 0.28. These weak correlations among differ-
ent parameters show that the data is unbiased for present Fig. 4  Distribution of different slope parameters and their correlation

13
487 
Page 6 of 18 Environmental Earth Sciences (2022) 81:487

research work; and the developed ANN and MLR models of liquid and plastic limits as 23.3 and 19.5. Kainthola et al.
will have a wider applicability in the future (Alavi and (2021), further contribute other physico-mechanical attrib-
Gandomi 2011). Furthermore, the statistical illustration utes to the present work, such as unit weight (2200 ± 150 kg/
of these properties hastens the workers’ conception and m3), Poisson’s ratio (0.29 ± 0.02) and Young’s modulus
understanding of their distribution pattern. Beside all (1000 ± 300 MPa). However, in the present work, only mean
these, data sanity can be quickly reinforced with the help values of these physico-mechanical parameters were consid-
of a statistical examination of these geotechnical proper- ered in numerical simulation for all 185 slopes. The open-
ties (Rezania et al. 2008). source program FLAC/Slope was employed for the finite
difference method (FDM) calculations (Fig. 5). FDM uses
Finite difference analysis equations of motion to derive new velocities and displace-
ments from stresses and forces. Then, strain rates are derived
Existing literature is rich in numerical techniques that from velocities, and new stresses are derived from strain
could be utilized to examine slope instability (Igwe et al. rates (Beyabanaki 2020).
2014; Zhang et al. 2021). Several advanced simulations To simulate the slope stability in the software, the “Sim-
approaches viz. Finite Element Method (FEM), Cracking ple form” has been used to get the desired geometry param-
Particle Method (CPM), Discrete Element Method (DEM), eters (rise, slope, left, right and depth), which were assigned
FDM, dual-horizon peridynamics (DH-PD), Bound- as per field observations (Fig. 5). As the data for studied
ary Element Method, Discrete Fracture Network, and slopes was collected in summer season (drained condition),
extended-FEM have been devised to investigate the slope the present simulation didn’t consider wet density. Only the
failures (Andriani and Parise 2015; Cao et al. 2021). In slope face and upper slope portions have been ascribed free
CPM, failure (crack) is simulated as the intensification of boundaries, the rest of the boundaries have been bound in
discontinuities at each node/particle, configured randomly both vertical and horizontal direction. Coarse-mesh den-
in the body (Rabczuk et al. 2007). Rabczuk et al. (2010) sity has been applied uniformly with an automatic meshing
had demonstrated that a particle is fragmented once the scheme, conforming to material boundaries. Eventually, the
desired norms are satisfied and supplant crack segment in model has been solved for FoS under Mohr–Coulomb failure
the body enabling its failure mechanism simulation under (Handin 1969) criteria (including cohesion and internal fric-
static or dynamic conditions. Moreover, Rabczuk and tion angle) following the shear strength reduction technique.
Belytschko (2004) introduced the Element free Galerkin Finally, one will have the safety factor for the modelled slope
method–Particle for modelling crack development using along with distribution of maximum shear-strain rate in the
mesh-free particle technique. Furthermore, with promis- slope body.
ing outcomes in 2D and great potential to work in 3D, An intensive finite difference investigation for each 185
the method shows connecting fractures will replicate the slope models in FLAC/Slope ascribed their safety factor
continuous rupture in the model (Rabczuk and Belytschko (FoS). These enumerated safety factors of slopes along
2007). Additionally, the dual-horizon principle empowers with their respective slope attributes (input parameters)
the earlier Peridynamics technique to treat the problems of engendered a handsome amount of data for generation
altering horizon sizes and ghost forces in solid mechanics of predictive models (MLR and ANN). As expected, the
(Ren et al. 2016). Further, the method was applied effec- outcome of FDM demonstrates that FoS decreases as
tively to simulate the crack path pattern and comprehend the slope angle and slope height increases, whereas FoS
multiple material problems (Ren et al. 2017). However, in increases with increment in cohesion and angle of internal
the present work, the power of the finite difference method friction of soil mass (Fig. 6). Moreover, the simulated FoS
(FDM) was preferred over more sophisticated-powerful values of these studied slopes were matched and calibrated
techniques due to its availability as an open-source tool with the ground conditions to validate the FLAC/Slope’s
(FLAC/Slope) and no additional requirement of sophisti- outcomes. In 49 cases the FDM depicted critical condi-
cated hardware accessories. tions, however, in field 127 slopes had shown no visible
Initial stability assessment was conducted in a finite signatures of failure. In addition, on cross validation of
difference code to ascertain the factor of safety. A total of these FoS values and ground reality, only 175 outcomes
185 slope models with varying shear strength parameters were true in both the situations and remaining 10 cases
and geometrical attributes were developed. Based on the differ. Therefore, the FDM analysis was true in 94.59%
sieve analysis, approximately 70% of grain-sizes are greater cases out of 185 total slopes analysed. The further scru-
than 2 mm for the presently studied slopes (Kainthola et al. pulous analysis of this ground truth data and FoS values
2021). Thus, the coarse-grained particles are the major con- indicates that only in scenarios, where 1 < FoS < 1.04 had
stituents of the slope-forming materials with poor sorting. the anomalous ground condition. Additionally, the field
Further, the laboratory works determined the mean values investigation and numerical examination of these finite

13
Environmental Earth Sciences (2022) 81:487 Page 7 of 18  487

Fig. 5  A representation of finite difference calculation through FLAC/Slope software (cohesion: 69626.51 Pa, angle of internal friction: 33.9°
and mass-density: 2200 kg/m3)

slopes indicate that the majority of the instabilities were region. The Mohr–Coulomb failure criteria ascertain the
of shallow failure types, and in some cases deep-seated maximum shear stress ( 𝜏max ) on the slip surface using
failures were involved. Eq.  (2) (Labuz and Zang 2012). Moreover, the shear
Finite difference tool uses Mohr–Coulomb failure cri- strength reduction (SSR) techniques were employed in the
terion to assess the stability of slope masses in the study FLAC/Slope software program to enumerate the factor of

Fig. 6  Influence on FoS of a cohesion and angle of internal friction, b slope angle and slope height

13
487 
Page 8 of 18 Environmental Earth Sciences (2022) 81:487

Table 1  Input geotechnical Parameters Slope angle (°) Slope height (m) Cohesion (Pa) Angle of inter- FoS
parameters for slope stability in nal friction (°)
FLAC/Slope (FDM analysis)
software and its output of 185 Minimum 25.00 20.00 10,787.00 0.26 0.12
locations
25th percentile 40.00 50.00 30,877.99 32.50 0.98
50th percentile 45.00 66.00 50,000.00 37.70 1.21
75th percentile 50.00 90.00 66,684.50 42.90 1.48
Maximum 80.00 200.00 124,543.20 52.70 2.45
Mean 45.70 76.39 50,609.07 36.28 1.25
Standard deviation 9.55 35.26 27,304.58 9.04 0.42

Table 2  Variance inflation factor for each independent parameter where ao, a1, a2, a3, a4 are coefficients of independent (input)
ascribed through MLR variables and ε is the residual error of the MLR prediction
Independent vari- Height Slope angle Cohesion Friction angle model.
able The significance of accommodating any predictor (regres-
sor) in multiple linear regression models should be exam-
VIF value 1.049 1.008 1.140 1.109
ined before actually employing them. One such aspect of
assessing the predictors is multicollinearity, which examines
the effect of one variable onto another which may affect
safety (Eqs. 3, 4) (Labuz and Zang 2012; Bharati et al. the stability. A higher multicollinearity in the predictors
2021). A statistical summary of finite difference calcula- of any model poses a threat to soundness and reliability of
tion can be visualized in Table 1. that model. Variance inflation factor (VIF) is used to assess
multicollinearity in the multiple linear regression analysis
𝜏max = c + 𝜎n tan𝜑 (2) (Eq. 6). Outcomes of VIF for each independent variable
adopted in the present study for the MLR model is close to
c
ctrial = (3) 1; and evades probability of high multicollinearity among
FoStrial the predictors, moreover favours the orthogonality among
[ ] the predictors (Table 2) (Daoud 2017).
𝜑
𝜑trial = tan−1 (4) 1
FoStrial VIF = (6)
1 − R2i
where c is the cohesion, φ is the internal friction angle, ctrial
is the FoStrial (SSR) times reduced cohesion, and φtrial is the where R2i is the regression of ith regressor on the rest other
FoStrial (SSR) times reduced angle of internal friction. predictors.

Multiple linear regression model Artificial neural network (ANN) model

The study adopts the traditional multiple linear regres- Artificial neural network (ANN) is a computational tech-
sion (MLR) algorithm to generate a model or function to nique that works similar to neurons present in a human brain
predict FoS of hill slopes in the research area. MLR is a (Wang 2003; Lau et al. 2019). It is a subset of machine learn-
general extension of linear regression analysis by incorpo- ing, which itself is a further subset of Artificial Intelligence
ration of more than one independent variable. The model (Bini 2018). Analogous to the human brain, it predicts the
limns FoS based on input parameters like cohesion (c), outcomes based on its previous training datasets or past
angle of internal friction (f), slope height (h), and slope experiences. Here, the experience of ANN can be correlated
angle (θ) (Eq. 5). The MLR model tries to formulate a to the number of training datasets, the greater the amount
model (mathematical hyperplane) as a function of inde- of training datasets engender higher accuracy (Jain and
pendent variables (input), to predict the dependent vari- Kumar 2007). The learning techniques can be supervised
able (output). or unsupervised (Japkowicz 2001). If the training dataset
is accompanied by a given outcome, it will be supervised
FOSMLR = a0 + a1 ∗ c + a2 ∗ f + a3 ∗ 𝜃 + a4 ∗ h + 𝜀 (5) learning otherwise unsupervised learning (Fig. 7a) (Nam-
dev et al. 2015). The supervised technique is useful for the
prediction and classification, whereas unsupervised learning

13
Environmental Earth Sciences (2022) 81:487 Page 9 of 18  487

Fig. 8  The structure of ANN model developed for the present study


(n represents node with subscripts showing its number for given hid-
den layers) (after Khandelwal and Singh 2009)

Fig. 7  Workflow of a machine learning model and b traditional pro- (x − u)


gramming (modified after Sarkar et al. 2018) z= (7)
s

approaches aids in clustering. In contrast to machine learn- where z is the normalized value of a random input variable
ing paradigms, the traditional programming algorithms were x; u and s are the mean and standard deviation of the input
based on a predefined set of rules/ codes (Fig. 7b). training data, respectively.
The algorithms of ANN empower a model to assign Both the neural network models were structured, entail-
appropriate weights for each input variable and iterate ing four input parameters to predict an output. The inherent
epoch-times to minimize the cost function (loss or error) pattern in the variables such as slope height, slope angle,
(Siddique et al. 2018). The ANN has primarily 3-layers cohesion, and angle of internal friction and their contribu-
namely, input layer, hidden layer, and output layer (Khandel- tion in FoS, were recognised by both designed ANN models.
wal and Singh 2009). The hidden layer is the place where all The present ANN models incorporate two hidden layers of
the computational jobs are accomplished (Boob et al. 2022). 32 nodes each. The Gaussian Error Linear Unit (GELU) is
Hidden layers may contain single or multiple layers and each implemented in both the hidden layers of each model (Eq. 8)
layer consists of one or more nodes. These layers are decided (Hendrycks and Gimpel 2016).
on hit and trial basis or sometimes help of a few optimiza- � � ��
tion algorithms are incorporated for the best outcomes and x
GELU(x) = 0.5x 1 + erf √ (8)
there is no specific set of rules defined for it (Ramakrishnan 2
et al. 2013).
In the present study, two ANN models have been pro- However, one neural network model was compiled using
grammed on the Python platform with the help of listed Adam optimizer, while the other uses SGD optimizer. In both
open-source libraries such as pandas, NumPy, Keras, scikit the models, the loss is enumerated through mean_squared_
learn and matplotlib, etc. The dataset is imported into the error (MSE) function (Eq. 9) (Arafat et al. 2020). Finally,
model through pandas, after which the input and output vari- with a batch size of 5 and 100 epochs, two ANN models
ables are demarcated for further processing via NumPy. The have been structured (Eq. 10) (Fig. 8) (Dar and Shah 2021;
dataset is again divided into training and testing data using Chi et al. 2021). Moreover, both the structured networks are
train_test_split operation of sklearn.model_selection; 80% tested based on their outcomes and finite difference results.
data are used as training and remaining 20% as testing out Eventually, the comparison is made on the basis coefficient
of total of 185 datasets. The input variables were normalized of determination achieved by both the designed models.
using StandardScaler operation of scikit-learn preprocessing N
1 ∑( )2
algorithm based on Eq. (7) (Ruiz et al. 2020). MSE = t − tdi (9)
n i=1 i

13
487 
Page 10 of 18 Environmental Earth Sciences (2022) 81:487

where ti is the real value (FDM outcomes) and tdi is the the small steps will aid the convergence. Hence, the con-
predicted value of ANN model. vergence can be achieved without over-steeping (in case of
(m ) big learning rates), furthermore without compromising the
∑ speed (as in case of small learning rates). In Adam, one can
FoSANN = h xi wij + bi (10)
i=1
comprehend the mathematics of updating each parameter/
weight (wj) through a combination of Eqs. (12–15) (Bushaev
where h is a nonlinear function (GELU in the present mod- 2018; Kathuria 2018). To dodge the complexity in writing
els), xi is the input nodes or previous layers nodes, wij is the the equation, w is used instead of wj in the below equations.
assigned weight for the ith input variables, and bj is bias in
(12)
( )
each hidden layer. vt = 𝛽1 ∗ vt−1 − 1 − 𝛽1 ∗ gt
The optimization algorithms serve in minimizing any
mathematical function by identifying the optimal param- st = 𝛽2 ∗ st−1 − 1 − 𝛽2 ∗ g2t (13)
( )
eters (Zhang 2018). Earlier workers have illustrated several
optimizers that are capable of minimizing the cost/loss func- v
tion (mathematical function) of machine learning models. Δwt = −𝜂 √ t ∗ gt (14)
As discussed, the two ANN models were devised with same st + 𝜖
structure, however, with different optimizers, i.e., Adam and
Stochastic Gradient Descent (SGD). Both are commonly wt+1 = wt + Δwt (15)
used optimizer techniques in different machine learning
models (Saha et al. 2021). The SGD is more viable in mini- where η signifies “initial learning rate”, vt stands for “exponen-
mizing the loss functions than its earlier versions of gradient tial average of gradients along wj”, ­gt denotes the “gradient at
descent. The SGD does not consider whole parameters at time t along wj”, and st implies the “exponential average of the
each iteration step in updating the parameters, instead takes square of gradients along wj”. Besides, β1, and β2 are the hyper-
a single or small subset of the parameters as compared to parameters with common numerical values of 0.9 and 0.99,
batch gradient descent (BGD). So, SGD out-performs BGD respectively. Meanwhile, the ε generally takes a value ­10−10.
in terms of speed, besides converges at global minima earn-
ing a better accuracy (Hansen 2019; Gupta et al. 2021).
Eq. (11), mathematically elucidate the rule of parameter (θ) Results and discussion
update in case of stochastic gradient descent optimization
(Hansen 2019). The tested geotechnical parameters along with different
slope geometry for the study area were incorporated into
𝜃 = 𝜃 − 𝜂 ∗ Δ𝜃 J(𝜃;x, y) (11) the finite difference model. The FDM facilitates a factor
of safety for each 185 slope models, through mathematical
where the symbols η, Δ, J, x, and y represent learning
equations inbuilt in the FLAC/ Slope software program. The
rate, gradient, cost function, training example and label,
slope models with FoS greater than 1 were considered sta-
respectively.
ble, while those less than 1 were deemed unstable. The effect
Adam embodies properties of momentum and adaptive
of each parameter incorporated in the study through FDM
learning rates (ALR) to update the parameters involved in
demonstrates that slope angle and slope height are negatively
machine learning prediction models, consequently, mini-
correlated, whereas cohesion and angle of internal friction
mizes the cost function easily to converge swiftly (Kingma
are positively correlated (Fig. 9). The effect of slope angle
and Ba 2014; Jais et al. 2019). The momentum enhances
and angle of internal friction are more pronounced in com-
Adam to update the parameters of the model in the process
parison to that of cohesion and slope height. The outcomes
of optimization by partially enumerating the one-step earlier
of finite difference were taken in multi-linear regression and
update in the present update (Meng et al. 2015; Balogun
artificial neural network as its input and output datasets.
and Adebisi 2021; Zhao et al. 2021). Essentially, momen-
Multicollinearity in the MLR will reduce the precision of
tum aids the Adam to avoid converging the models at local
the estimated coefficients, which will weaken the statistical
minima and hence enables them to attain the global minima.
power of the regression model. Eventually, to check that
Whereas, the adaptive learning rate governs the learning
there was no multicollinearity, VIF values were discerned.
rate (size of step in updating the parameters). A big learn-
Moreover, the estimated VIF values indicate that the regres-
ing may overshoot the minima and a small learning rate
sors chosen for MLR are unaffected by other independent
will delay the convergence. The Adam will overcome the
variables. Estimates of MLR model can swing wildly based
issue with the ALR technique, that enable the optimization
on other independent variables in the model, moreover, the
with bigger steps initially and on approaching the minima
coefficients become very sensitive to small changes in the

13
Environmental Earth Sciences (2022) 81:487 Page 11 of 18  487

quite high degree of accuracy when compared with the real


datasets enumerated through FDM (Fig. 10a–d). The model
minimizes the loss with each epoch, and eventually mean
squared error is reduced close to zero (Fig. 10e, f). Besides,
the appropriate weights are assigned to each parameter with
a bias in each layer, and consequently robust, swift and intel-
ligible ANN models were evolved.
The regression analysis of MLR prediction model yields
a value of R2 equal to 0.8945 on testing datasets and 0.9204
on the training datasets (Fig. 11a, b). The coefficient of
determination of ANN (with Adam optimizer) is 0.9985 and
0.9996 on testing and training data, respectively (Fig. 11c,
d). Besides, the ANN (with SGD optimizer) earns a R2 value
of 0.9926 (on testing data) and 0.9945 (on training datasets)
(Fig. 11e, f). The coefficient of determination can be derived
using Eq. (17); and the higher R2 values indicate the greater
significance of the prediction model. Hence, in the present
Fig. 9  Correlation among geotechnical properties and FoS (obtained study the ANN model (with Adam optimizer) is more reli-
through FDM) able and robust than the other ANN model (with SGD opti-
mizer) and the MLR model. The comparison of these models
model. Standard error is a statistical approach to test the fit has been further made based on mean squared error (MSE).
of any regression model, and smaller values of it approves Eventually, MSE declares that ANN (with Adam optimizer)
the viability of the devised prediction model through MLR performs better than their counterpart ANN (with SGD opti-
(Table 3). The significance of the MLR model further vali- mizer) and MLR (Table 4).
dated by other statistical methods viz., t value and p value
(Table 3). The smaller numeric values of p and *** signi- SSresidual
R2 = 1 − (17)
fies that the MLR model is free from randomness and is an SStotal
extremely confidents model (Di Leo and Sardanelli 2020).
where, ­SSresidual means squared sum of residuals, and ­SStotal
The MLR model derived in the present research can be deci-
indicates squared sum of total.
phered and analysed using Eq. (16) and Table 3.
Besides, predicting the FoS, the developed models have
FOSMLR = 1.4282724 + 0.00619 ∗ c + 0.03223 ∗ f a fair scope for upgradation into the risk estimation mod-
(16) els with incorporation of vulnerability factors in future. As,
− 0.00375 ∗ h − 0.03009 ∗ 𝜃
Sardooi et al. (2021) proposed a hybrid ML model for esti-
The artificial neural network models involved in the pre- mation of landslide risk zone by coupling the susceptibil-
sent research consist of four input parameters and one output ity and vulnerability attributes, contingent upon landslide
variable. Similar to MLR model, a total of 185 datasets, out inventories. Moreover, Abraham et al. (2021) developed five
of which 148 datasets were taken for training the neural net- AI-prediction models for landslide susceptibility mapping
work model and the rest 37 datasets were utilised as testing (LSM) based on 12 critical parameters to comprehend the
ANN model. The structure of the ANNs has two hidden lay- consequences of spatial resolution and train-test ratios of
ers of 32 nodes with GELU acting as an activation function, input datasets over the performance of these algorithms. To
bridging the input and output layers. The prediction capa- have the best prediction algorithm for provided computa-
bility of ANN models devised for the present study shows a tional facilities, it was opined to check the functioning of
different LSM prediction models by varying splitting ratios
of input datasets and trying all the available spatial-data res-
Table 3  Outcomes of MLR and statistical synopsis of the model olutions for any given area. However, in the present research
the significance of spatial resolution cannot affect the per-
Coefficients Estimate Std. error t value Pr ( >|t|) formance of the developed models, as the model training
Intercept 1.4282724 0.0687572 20.77  < 2e–16*** datasets rely on laboratory and field datasets only. Moreover,
Height − 0.0037531 0.0003018 − 12.43  < 2e–16*** the choice of data splitting (80–20) was made considering
Slope angle − 0.0300983 0.0010171 − 29.59  < 2e–16*** the accepted recommendation in the literature.
Cohesion 0.0061953 0.0003911 15.84  < 2e–16*** Additionally, a lucid expression determining the angle of
Friction angle 0.0322318 0.0011272 28.59  < 2e–16*** repose (accommodating the angle of internal friction and

13
487 
Page 12 of 18 Environmental Earth Sciences (2022) 81:487

Fig. 10  Pictorial performance illustration of the a ANN (with Adam (with SGD optimizer) on training datasets; e loss (MSE) vs epoch in
optimizer) on testing data; b ANN (with SGD optimizer) on testing the training of the ANN model with Adam optimizer); f loss (MSE)
datasets; c ANN (with Adam optimizer) on training datasets; d ANN vs epoch in the training of the ANN model with SGD optimizer

cohesion) for granular heaps in terrestrial and extra-terres- Earth’s gravity is applied in numerical simulations of the
trial environments has been derived from the work of Elekes present work. Also, the future works in the line of Elekes
and Parteli (2021). The equation is a function of particle size and Parteli (2021), will be an asset in comprehending extra-
and Van-der Waals forces, along with appropriate simulation terrestrial slope instabilities.
of interacting gravitational force on different planets. Thus, In addition, there are numerous factors controlling the
the attained equation is open to extra-terrestrial bodies apart slope instability and its investigation, as discussed in the
from Earth and may be a good alternative to cohesion and above sections of the present manuscript, in the context
angle of internal friction in future terrestrial/extra-terres- of past publications (Zhang and Liu 2010; Andriani et al.
trial landslide studies. On contrary, the developed predictive 2015; Cao et al. 2016; Fazio et al. 2019; Andriani and Loio-
models are only credible to terrestrial slope failures, as only tine 2020). Often, the availability of resources and ground
conditions dictate the choice of parameters as well as the

13
Environmental Earth Sciences (2022) 81:487 Page 13 of 18  487

Fig. 11  Picturing the performance of models a MLR (testing); b MLR (training); c ANN (with Adam optimizer) (testing); d ANN (with Adam
optimizer) (training) e ANN (with SGD optimizer) (testing); f ANN (with SGD optimizer) (training)

13
487 
Page 14 of 18 Environmental Earth Sciences (2022) 81:487

Table 4  Statistical comparison of developed models the government/private bodies to assess the safety concerns
Models R 2
MSE
of infrastructure projects, quite well. A vivid comparison
with recently published work indicated how the developed
MLR (testing) 0.8945 0.01823 models may account vulnerability factor to estimate associ-
MLR (training) 0.9204 0.01386 ated risks along with FoS. Further, the work encapsulates
ANN—Adam optimizer (testing) 0.9985 0.00037 the discussion about landslide influencing variables and their
ANN—Adam optimizer (training) 0.9996 0.00006 complementarity with other ones.
ANN—SGD optimizer (testing) 0.9926 0.00171
ANN—SGD optimizer (training) 0.9945 0.00088
Conclusion

landslide analysis viz. empirical, numerical, remote sensing The aim of the present study is to devise a reliable model for
and GIS (Fall et al. 2006; Hong et al. 2016; Kahyaoğlu et al. the prediction of the safety factor of slopes, through machine
2017; Chen et al. 2021a, b; Francioni et al. 2021; Khamkar learning techniques. However, one needs copious amount
et al. 2022). Previously, the remote sensing-based landslide of data to develop such machine learning models with
conditioning factors, along with inventories, have supplanted higher accuracy. Moreover, to suffice the objective of the
the susceptibility mapping (Abraham et al. 2021; Sardooi research and generate required data, preliminary fieldwork
et al. 2021). Published research articles lay emphasis on was supplanted by laboratory assessment and finite differ-
the complementarity of parameters viz. grain-size distribu- ence examination. A total of 185 datasets were rendered in
tion, water content, lithology, particle-shape, and degree of this precursor study, which aided the further development
consolidation over the shear-strength variables of the slope of the present prediction models. Numerous machine learn-
(Parteli et al. 2014). Similarly, slope angle will determine ing models prevail in the literature; however, multiple linear
certain features like erosion, infiltration, land-use, and sur- regression and two artificial neural network (ANN) mod-
face-runoff (Hill and Peart 1998; Ribolzi et al. 2011; Peng els were assimilated in the present work due to their wide-
and Wang 2012), while parameters like type of vegetation, acceptability and robustness. The stability of a slope can be
geology, erosion, rainfall, or anthropogenic activities will attributed to several parameters, however, the present MLR
have an effect on slope geometry (Van Westen et al. 1997; and ANN prediction models consider the four most critical
Kouli et al. 2010). Thus, a parameter aggravating the slope parameters for the stability assessment. The four parameters
mass condition may have some relationship with the other chosen as predictors in the present research are cohesion
triggering factors. So, the attributes governing the instability (c), angle of internal friction (f), slope height (h), and slope
in the present work (Eq. 5), may complement the parameter angle (θ). Density was not used as a parameter in deep learn-
accounted for in earlier literatures. Besides, the interde- ing model, since it had low variance in the study area. The
pendency among the combination of parameters affecting datasets generated in the preliminary study and were cat-
slope failure should be examined scrupulously, before their egorised into training datasets (80%) and testing datasets
incorporation in landslide investigation or susceptibility (20%) to foster the further advancement of the research. Key
modelling. findings of the present work can be summarized as:
Evidently, the results are in harmony with the field con-
ditions and can be extended to other regions with similar a. Statistical examination of the predictor (independent)
geological and geo-environmental conditions. Further- variables resulted in weak correlation among themselves
more, with the addition of more data from different regions as well as multicollinearity lacks in the datasets. Hence,
in training the developed algorithms, the viability of the the developed ML models will have greater viability.
landslide estimation models can be extended to other areas. b. The FDM analysis shows that lower values of slope
The outcomes of the prediction models show the efficacy angle and slope height favours the stability, whereas
of open-source tools which are used in poor and develop- higher values of cohesion and angle of internal friction
ing nations to swiftly assess slope safety, before the com- render higher FoS.
mencement of the excavation. Further, the laboratory works c. All these prediction models have four independent vari-
conducted in the present study contribute well to advancing ables (predictor) namely, cohesion, angle of internal
the knowledge of shear-attributes of slope masses. Addi- friction, slope angle and slope height with one depend-
tionally, the work is the first of its type in the studied area, ent variable (output) i.e., factor of safety (FoS)
considering the significance of geotechnical and geometrical d. The multiple linear regression (MLR) prediction model
attributes of slopes with the power of finite difference code. earns a R2 value of 0.8945 (on testing datasets) with
Therefore, the developed machine learning models will aid other statistical terms declaring the coefficients’ such
as standard error, t value and p value for every predictor

13
Environmental Earth Sciences (2022) 81:487 Page 15 of 18  487

as highly significant. Moreover, the practicality of the F (eds) Engineering geology for society and territory, vol 5.
devised model can also owe to its lack of multi-colline- Springer, Cham, pp 387–392. https://​doi.​org/​10.​1007/​978-3-​
319-​09048-1_​74 
arity among its regressors. However, MSE calculated for Anitescu C, Atroshchenko E, Alajlan N, Rabczuk T (2019) Artifi-
the present model shows that model could lead to greater cial neural network methods for the solution of second order
error and compromise the performance. boundary value problems. Computer Mat Contin 59(1):345–359.
e. Both the ANN models developed in the present research https://​doi.​org/​10.​32604/​cmc.​2019.​06641
Arafat M, Sjafrizal T, Anugraha RA (2020) An artificial neural network
beats the MLR model in terms of coefficient of determi- approach to predict energy consumption and surface roughness
nation and mean squared error. Establishing the suprem- of a natural material. SN Appl Sci 2:1–1. https://d​ oi.o​ rg/1​ 0.1​ 007/​
acy of artificial neural networks over multi-linear regres- s42452-​020-​2987-6
sion in case prediction models. Armaghani DJ, Mohamad ET, Momeni E, Narayanasamy MS (2015)
An adaptive neuro-fuzzy inference system for predicting uncon-
f. R2 of both the ANN models is greater than 0.99 and fined compressive strength and Young’s modulus: a study on
MSE are of the order 1­ 0−4 or lower, hence the models Main Range granite. Bull Eng Geol Env 74(4):1301–1319.
have a great value in terms of performance. Further- https://​doi.​org/​10.​1007/​s10064-​014-​0687-4
more, the work shows that ANN model developed with Asuero AG, Sayago A, Gonzalez AG (2006) The correlation coeffi-
cient: an overview. Crit Rev Anal Chem 36(1):41–59. https://​
Adam as optimizer dominates the other structure of the doi.​org/​10.​1080/​10408​34050​05267​66
ANN with SGD as optimizer. Auden JB (1934) The geology of the Krol belt. Rec Geol Survey India
67(4):357–454
Balogun AL, Adebisi N (2021) Sea level prediction using ARIMA,
Acknowledgements  The authors would like to express their gratitude SVR and LSTM neural network: assessing the impact of ensem-
to Itasca Consulting Group, Inc, tensorflow.org, python.org and r-pro- ble ocean—atmospheric processes on models’ accuracy. Geomat
ject.org for the analysis tools. VHRP would like to acknowledge UGC Nat Hazard Risk 12(1):653–674
Non-NET Fellowship—BHU for the financial assistance. AK and VS Benbouras MA, KettabMitiche R, Zedira H, Petrisor AI, Mezouar
would thank BHU-IOE seed grant for the financial support. N, Debiche F (2019) A new approach to predict the compres-
sion index using artificial intelligence methods. Mar Georesour
Geotechnol 37(6):704–720. https://​doi.​org/​10.​1080/​10641​19X.​
Declarations  2018.​14845​33
Beyabanaki SA (2020) A comparison between using finite difference
Conflict of interest  The authors declare that they have no known com- and limit equilibrium methods for landslide analysis of slopes
peting financial interests or personal relationships that could have ap- containing a weak layer. Am J Eng Res 9(12):68–79
peared to influence the work reported in this paper. Bharati AK, Ray A, Khandelwal M, Rai R, Jaiswal A (2021) Stabil-
ity evaluation of dump slope using artificial neural network
and multiple regression. Eng Comput. https://​doi.​org/​10.​1007/​
s00366-​021-​01358-y
Bini SA (2018) Artificial intelligence, machine learning, deep learn-
References ing, and cognitive computing: what do these terms mean and
how will they impact health care? J Arthroplast 33(8):2358–
Abraham MT, Satyam N, Jain P, Pradhan B, Alamri A (2021) Effect 2361. https://​doi.​org/​10.​1016/j.​arth.​2018.​02.​067
of spatial resolution and data splitting on landslide susceptibil- Boob D, Dey SS, Lan G (2022) Complexity of training relu neu-
ity mapping using different machine learning algorithms. Geo- ral network. Discrete Optim 44(1):100620. https://​doi.​org/​10.​
mat Nat Hazard Risk 12(1):3381–3408. https://d​ oi.​org/​10.​1080/​ 1016/j.​disopt.​2020.​100620
19475​705.​2021.​20117​91 Bui XN, Muazu MA, Nguyen H (2020) Optimizing Levenberg–
Aiello LC (2016) The multifaceted impact of Ada Lovelace in the digi- Marquardt backpropagation technique in predicting factor
tal age. Artif Intell 235:58–62. https://d​ oi.​org/​10.​1016/j.​artint.​ of safety of slopes after two-dimensional OptumG2 analy-
2016.​02.​003 sis. Eng Computer 36:941–952. https://​d oi.​o rg/​1 0.​1 007/​
Alavi AH, Gandomi AH (2011) A robust data mining approach for s00366-​019-​00741-0
formulation of geotechnical engineering systems. Eng Comput Bushaev V (2018) Adam—latest trends in deep learning optimization.
Eng Compu 28(3):242–274. https://d​ oi.o​ rg/1​ 0.1​ 108/0​ 26444​ 0111​ Towards Data Science. https://​www.​towar​dsdat​ascie​nce.​com/​
11181​32 adam-​latest-​trends-​in-​deep-​learn​ing-​optim​izati​on-​6be9a​29137​
Andriani GF, Loiotine L (2020) Multidisciplinary approach for assess- 5c. Accessed 24 September 2021
ment of the factors affecting geohazard in karst valley: the case Cao Y, Yin K, Alexander DE, Zhou C (2016) Using an extreme learn-
study of Gravina di Petruscio (Apulia, South Italy). Environ ing machine to predict the displacement of step-like landslides in
Earth Sci 79:458. https://​doi.​org/​10.​1007/​s12665-​020-​09212-y relation to controlling factors. Landslides 13(4):725–736. https://​
Andriani GF, Parise M (2015) On the applicability of geomechani- doi.​org/​10.​1007/​s10346-​015-​0596-z
cal models for carbonate rock masses interested by karst pro- Cao C, Feng J, Tao Z (2021) 3D numerical simulation of landslides
cesses. Environ Earth Sci 74:7813–7821. https://d​ oi.o​ rg/1​ 0.1​ 007/​ for the full high waste dump using SPH method. Adv Civ Eng.
s12665-​015-​4596-z https://​doi.​org/​10.​1155/​2021/​88978​26
Andriani GF, Diprizio G, Pellegrini V (2015) Landslide susceptibility Chakraborty A, Goswami D (2017) Prediction of slope stability
of the La Catola Torrent catchment area (Daunia Apennines, using multiple linear regression (MLR) and artificial neu-
southern Italy): a new complex multi-step approach. In: Lollino ral network (ANN). Arab J Geosci. https://​doi.​org/​10.​1007/​
G, Manconi A, Guzzetti F, Culshaw M, Bobrowsky P, Luino s12517-​017-​3167-x
Chen XL, Liu CG, Chang ZF, Zhou Q (2016) The relationship
between the slope angle and the landslide size derived from limit

13
487 
Page 16 of 18 Environmental Earth Sciences (2022) 81:487

equilibrium simulations. Geomorphology 253:547–550. https://​ Feng X, Li S, Yuan C, Zeng P, Sun Y (2018) Prediction of slope stabil-
doi.​org/​10.​1016/j.​geomo​rph.​2015.​01.​036 ity using naive Bayes classifier. KSCE J Civ Eng 22(3):941–950.
Chen C, Xie MW, Jiang YJ, Jia BN, Du Y (2021a) A new method https://​doi.​org/​10.​1007/​s12205-​018-​1337-3
for quantitative identification of potential landslide. Soil Found Francioni M, Stead D, Sharma J, Clague JJ, Brideau MA (2021)
61(5):1475–1479. https://​doi.​org/​10.​1016/j.​sandf.​2021.​07.​004 An integrated InSAR-borehole Inclinometer-numerical mod-
Chen CY, Chen HW, Wu WC (2021b) Numerical modeling of eling approach to the assessment of a slow-moving landslide.
interactions of rainfall and earthquakes on slope stability Environ Eng Geosci 27(3):287–305. https://​doi.​org/​10.​2113/​
analysis. Environ Earth Sci 80:524. https://​doi.​org/​10.​1007/​ EEG-D-​20-​00109
s12665-​021-​09855-5 Frattini P, Crosta GB (2013) The role of material properties and land-
Chi Z, Jiang Z, Kamruzzaman MM, Hafshejani BA, Safarpour scape morphology on landslide size distributions. Earth Planet
M (2021) Adaptive momentum-based optimization to train Sci Lett 361:310–319. https://d​ oi.o​ rg/1​ 0.1​ 016/j.e​ psl.2​ 012.1​ 0.0​ 29
deep neural network for simulating the static stability of the Gelisli K, Kaya T, Babacan AE (2015) Assessing the factor of safety
composite structure. Eng Computer. https://​doi.​org/​10.​1007/​ using an artificial neural network: case studies on landslides in
s00366-​021-​01335-5 Giresun, Turkey. Environ Earth Sci 73:8639–8646. https://​doi.​
Choobbasti AJ, Farrokhzad F, Barari A (2009) Prediction of slope org/​10.​1007/​s12665-​015-​4027-1
stability using artificial neural network (case study: Noabad, Guo L, Chen G, Gong S, Sun H, Chantat K (2021) Analysis of rain-
Mazandaran, Iran). Arab J Geosci 2(4):311–319. https://​doi.​org/​ fall-induced landslide using the extended DDA by incorporating
10.​1007/​s12517-​009-​0035-3 matric suction. Comput Geotech 135:104145. https://​doi.​org/​10.​
Cui P, Guo CX, Zhou JW, Hao MH, Xu FG (2014) The mechanisms 1016/j.​compg​eo.​2021.​104145
behind shallow failures in slopes comprised of landslide deposits. Gupta D, Hazarika BB, Berlin M, Sharma UM, Mishra K (2021)
Eng Geol 180:34–44. https://​doi.​org/​10.​1016/j.​enggeo.​2014.​04.​ Artificial intelligence for suspended sediment load prediction:
009 a review. Environ Earth Sci 80(9):1–39. https://​doi.​org/​10.​1007/​
Cüneyt Aydin A, Tortum A, Yavuz M (2006) Prediction of concrete s12665-​021-​09625-3
elastic modulus using adaptive neuro-fuzzy inference system. Handin J (1969) On the Coulomb-Mohr failure criterion. J Geophys
Civ Eng Environ Syst 23(4):295–309. https://​doi.​org/​10.​1080/​ Res 74(22):5343–5348
10286​60060​07723​48 Hansen C (2019) Optimizers explained—adam, momentum and sto-
Daoud JI (2017) Multicollinearity and regression analysis. In: journal chastic gradient descent. Machine Learning from Scratch. https://​
of physics: conference series 949(1):012009. IOP publishing. www.​mlfro​mscra​tch.​com/​optim​izers-​expla​ined/#/. Accessed on
https://​www.​ui.​adsabs.​harva​rd.​edu/​link_​gatew​ay/​2017J​PhCS.​ 20 September 2021
949a2​009D/ https://​doi.​org/​10.​1088/​1742-​6596/​949/1/​012009 Harari YN (2014) Sapiens: a brief history of humankind. Harvill
Dar LA, Shah MY (2021) Deep-seated slope stability analysis and Secker, London
development of simplistic FOS evaluation models for stone col- Hazari S, Roy S, Ghosh S (2021) Stability analysis of layered
umn-supported embankments. Transp Infrastruct Geotechnol soil slopes using truncated pile with numerical solution.
8:203–227. https://​doi.​org/​10.​1007/​s40515-​020-​00134-7 Transp Infrastruct Geotechnol. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​
Di Leo G, Sardanelli F (2020) Statistical significance: p value, 0.05 s40515-​021-​00174-7
threshold, and applications to radiomics—reasons for a conserva- Hendrycks D, Gimpel K (2016) Gaussian error linear units (gelus).
tive approach. Eur Radiol Exp 4(1):1–8. https://​doi.​org/​10.​1186/​ arXiv e-prints, pp.arXiv-1606. https://​doi.​org/​10.​48550/​arXiv.​
s41747-​020-​0145-y 1606.​08415
Dou J, Yunus AP, Bui DT, Merghadi A, Sahana M, Zhu Z, Chen CW, Hill RD, Peart MR (1998) Land use, runoff, erosion and their control: a
Khosravi K, Yang Y, Pham BT (2019) Assessment of advanced review for southern China. Hydrol Process 12(13–14):2029–2042
random forest and decision tree algorithms for modeling rain- Hoang ND, Pham AD (2016) Hybrid artificial intelligence approach
fall-induced landslide susceptibility in the Izu-Oshima Volcanic based on metaheuristic and machine learning for slope stabil-
Island, Japan. Sci Total Environ 662:332–346. https://d​ oi.o​ rg/1​ 0.​ ity assessment: a multinational data analysis. Expert Syst Appl
1016/j.​scito​tenv.​2019.​01.​221 46:60–68. https://​doi.​org/​10.​1016/j.​eswa.​2015.​10.​020
Eberly LE (2007) Multiple linear regression. Top Biostat. https://​doi.​ Hong H, Pradhan B, Jebur MN, Bui DT, Xu C, Akgun A (2016) Spatial
org/​10.​1007/​978-1-​59745-​530-5_9 prediction of landslide hazard at the Luxi area (China) using sup-
Elekes F, Parteli EJ (2021) An expression for the angle of repose of dry port vector machines. Environ Earth Sci 75(1):1–4. https://​doi.​
cohesive granular materials on Earth and in planetary environ- org/​10.​1007/​s12665-​015-​4866-9
ments. Proc National Academy of Sciences. https://​doi.​org/​10.​ Igwe O, Mode W, Nnebedum O, Okonkwo I, Oha I (2014) The analysis
1073/​pnas.​21079​65118 of rainfall-induced slope failures at Iva Valley area of Enugu
Erickson BJ, Korfiatis P, Akkus Z, Kline T, Philbrick K (2017) Toolkits State. Niger Environ Earth Sci 71(5):2465–2480. https://​doi.​org/​
and libraries for deep learning. J Digit Imaging 30(4):400–405. 10.​1007/​s12665-​013-​2647-x
https://​doi.​org/​10.​1007/​s10278-​017-​9965-6 Jain A, Kumar AM (2007) Hybrid neural network models for hydro-
Fall M, Azzam R, Noubactep C (2006) A multi-method approach to logic time series forecasting. Appl Soft Comput 7(2):585–592.
study the stability of natural slopes and landslide susceptibility https://​doi.​org/​10.​1016/j.​asoc.​2006.​03.​002
mapping. Eng Geol 82(4):241–263. https://​doi.​org/​10.​1016/j.​ Jais IK, Ismail AR, Nisa SQ (2019) Adam optimization algorithm for
enggeo.​2005.​11.​007 wide and deep neural network. Knowl Eng Data Sci 2(1):41–46
Fazio N, Perrotti M, Andriani G, Mancini F, Rossi P, Castagnetti C, Japkowicz N (2001) Supervised versus unsupervised binary-learning
Lollino P (2019) A new methodological approach to assess the by feedforward neural networks. Mach Learn 42(1):97–122.
stability of discontinuous rocky cliffs using in-situ surveys sup- https://​doi.​org/​10.​1023/A:​10076​60820​062
ported by UAV-based techniques and 3-D finite element model: Kahyaoğlu MR, İmançlı G, Özden G, Kayalar AŞ (2017) Numerical
a case study. Eng Geol 260:105205. https://​doi.​org/​10.​1016/j.​ simulations of landslide-stabilizing piles: a remediation project
enggeo.​2019.​105205 in Söke. Turkey Environ Earth Sci 76(19):1–4. https://​doi.​org/​
10.​1007/​s12665-​017-​6989-7
Kainthola A, Sharma V, Pandey VHR, Jayal T, Singh M, Srivastav
A, Singh PK, Ray PKC, Singh TN (2021) Hill slope stability

13
Environmental Earth Sciences (2022) 81:487 Page 17 of 18  487

examination along Lower Tons valley, Garhwal Himalayas, India. southwest China. CATENA 90:53–62. https://​doi.​org/​10.​1016/j.​
Geomat Nat Hazard Risk 12(1):900–921. https://d​ oi.o​ rg/1​ 0.1​ 080/​ catena.​2011.​11.​001
19475​705.​2021.​19067​58 Pham BT, Prakash I, Jaafari A, Bui DT (2018) Spatial prediction of
Kathuria A (2018) Intro to optimization in deep learning: momentum, rainfall-induced landslides using aggregating one-dependence
RMSProp and Adam. PaperspaceBlog. https://​www.​blog.​paper​ estimators classifier. J Indian Soc Remote Sens 46(9):1457–1470.
space.​com/​intro-​to-​optim​izati​on-​momen​tum-​r mspr​op-​adam/. https://​doi.​org/​10.​1007/​s12524-​018-​0791-1
Accessed on 1 October 2021 Pradhan B (2011) (2011) Use of GIS-based fuzzy logic relations and
Khamkar D, Aher S, Gawali P, Mhaske S (2022) Investigating prob- its cross application to produce landslide susceptibility maps in
able causes for predicting catastrophic landslides along NH-60 three test areas in Malaysia. Environ Earth Sci 63(2):329–349.
excavated through semi-arid basaltic terrain of Chandanapuri https://​doi.​org/​10.​1007/​s12665-​010-​0705-1
Ghat, Maharashtra, India. Environ Dev Sustain 24(2):2362–2386. Pradhan B, Lee S (2009) Landslide risk analysis using artificial neural
https://​doi.​org/​10.​1007/​s10668-​021-​01537-3 network model focussing on different training sites. Int J Phys
Khandelwal M, Singh TN (2009) Prediction of blast-induced ground Sci 4(1):1–5. https://​doi.​org/​10.​5897/​IJPS.​90003​43
vibration using artificial neural network. Int J Rock Mech Min Pradhan B, Lee S (2010) Delineation of landslide hazard areas on Pen-
Sci 46(7):1214–1222. https://​doi.​org/​10.​1016/j.​ijrmms.​2009.​03.​ ang Island, Malaysia, by using frequency ratio, logistic regres-
004 sion, and artificial neural network models. Environ Earth Sci
Khurma RA, Aljarah I, Sharieh A, Mirjalili S (2020) Evolopy-fs: An 60(5):1037–1054. https://​doi.​org/​10.​1007/​s12665-​009-​0245-8
open-source nature-inspired optimization framework in python Qi C, Tang X (2018) Slope stability prediction using integrated metaheuris-
for feature selection. Evolutionary machine learning techniques. tic and machine learning approaches: a comparative study. Comput
Springer, Singapore, pp 131–173. https://​d oi.​o rg/​1 0.​1 007/​ Ind Eng 118:112–122. https://​doi.​org/​10.​1016/j.​cie.​2018.​02.​028
978-​981-​32-​9990-0_8 Qiu H, Cui P, Regmi AD, Hu S, Wang X, Zhang Y, He Y (2017) Influ-
Kingma DP, Ba J (2014) Adam: a method for stochastic optimization. ence of topography and volume on mobility of loess slides within
arXiv preprint https://​doi.​org/​10.​48550/​arXiv.​1412.​6980. different slip surfaces. CATENA 157:180–188. https://​doi.​org/​
Kouli M, Loupasakis C, Soupios P, Vallianatos F (2010) Landslide 10.​1016/j.​catena.​2017.​05.​026
hazard zonation in high risk areas of Rethymno Prefecture, Crete Rabczuk T, Belytschko T (2004) Cracking particles: a simplified mesh-
Island. Greece Nat Hazard 52(3):599–621. https://​doi.​org/​10.​ free method for arbitrary evolving cracks. Int J Numer Method
1007/​s11069-​009-​9403-2 Eng 61(13):2316–2343. https://​doi.​org/​10.​1002/​nme.​1151
Labuz JF, Zang A (2012) Mohr-Coulomb failure criterion. Rock Rabczuk T, Belytschko T (2007) A three-dimensional large deforma-
Mech Rock Eng 45(6):975–979. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​ tion meshfree method for arbitrary evolving cracks. Comput
s00603-​012-​0281-7 Method Appl Mech Eng 196(29–30):2777–2799. https://​doi.​
Lau ET, Sun L, Yang Q (2019) Modelling, prediction and classifica- org/​10.​1016/j.​cma.​2006.​06.​020
tion of student academic performance using artificial neural Rabczuk T, Areias PMA, Belytschko T (2007) A meshfree thin shell
networks. SN Appl Sci 1(9):1–10. https://​d oi.​o rg/​1 0.​1 007/​ method for nonlinear dynamic fracture. Int J Numer Method Eng
s42452-​019-​0884-7 720(5):524–548. https://​doi.​org/​10.​1002/​nme.​2013
Lawal AI, Kwon S (2020) Application of artificial intelligence to Rabczuk T, Zi G, Bordas S, Nguyen-Xuan H (2010) A simple and
rock mechanics: an overview. J Rock Mech Geotech Eng. robust three-dimensional cracking-particle method without
https://​doi.​org/​10.​1016/j.​jrmge.​2020.​05.​010 enrichment. Comput Method Appl Mech Eng 199(37–40):2437–
Li LC, Tang CA, Zhu WC, Liang ZZ (2009) Numerical analysis of 2455. https://​doi.​org/​10.​1016/j.​cma.​2010.​03.​031
slope stability based on the gravity increase method. Comput Raghuvanshi TK (2019) Plane failure in rock slopes–a review on sta-
Geotech 36(7):1246–1258. https://​doi.​org/​10.​1016/j.​compg​eo.​ bility analysis techniques. J King Saud Univ-Sci 31(1):101–109.
2009.​06.​004 https://​doi.​org/​10.​1016/j.​jksus.​2017.​06.​004
Lollino P, Andriani GF (2017) Role of brittle behaviour of soft cal- Rahul KM, Rai R, Shrivastva BK (2015) Evaluation of dump slope
carenites under low confinement: laboratory observations and stability of a coal mine using artificial neural network. Geomech
numerical investigation. Rock Mech Rock Eng 50:1863–1882. Geophys Geo-Energ Geo-Resour 1:69–77. https://​doi.​org/​10.​
https://​doi.​org/​10.​1007/​s00603-​017-​1188-0 1007/​s40948-​015-​0009-8
Lu P, Rosenbaum MS (2003) Artificial neural networks and grey sys- Ramakrishnan D, Singh TN, Verma AK, Gulati A, Tiwari KC (2013)
tems for the prediction of slope stability. Nat Hazard 30:383– Soft computing and GIS for landslide susceptibility assess-
398. https://​doi.​org/​10.​1023/B:​NHAZ.​00000​07168.​00673.​27 ment in Tawaghat area, Kumaon Himalaya. India Nat Hazard
Ma J, Tang H, Liu X, Hu X, Sun M, Song Y (2017) Establishment 65(1):315–330. https://​doi.​org/​10.​1007/​s11069-​012-​0365-4
of a deformation forecasting model for a step-like landslide Ray A, Kumar V, Kumar A, Rai R, Khandelwal M, Singh TN (2020)
based on decision tree C5. 0 and two-step cluster algorithms: a Stability prediction of Himalayan residual soil slope using artifi-
case study in the three gorges reservoir area China. Landslides cial neural network. Nat Hazards 103(3):3523–3540. https://​doi.​
14(3):1275–1281. https://​doi.​org/​10.​1007/​s10346-​017-​0804-0 org/​10.​1007/​s11069-​020-​04141-2
Meng X, Yin M, Ning L, Liu D, Xue X (2015) A threshold artificial Ren H, Zhuang X, Cai Y, Rabczuk T (2016) Dual-horizon peridynam-
neural network model for improving runoff prediction in a karst ics. Int J Numer Method Eng 108(12):1451–1476. https://​doi.​
watershed. Environ Earth Sci 74(6):5039–5048 org/​10.​1002/​nme.​5257
Namdev N, Agrawal S, Silkari S (2015) Recent advancement in Ren H, Zhuang X, Rabczuk T (2017) Dual-horizon peridynamics: a
machine learning based internet traffic classification. Proc Com- stable solution to varying horizons. Comput Method Appl Mech
put Sci 60:784–791. https://​doi.​org/​10.​1016/j.​procs.​2015.​08.​238 Eng 318:762–782. https://​doi.​org/​10.​1016/j.​cma.​2016.​12.​031
Parteli EJ, Schmidt J, Blümel C, Wirth KE, Peukert W, Pöschel T Rezania M, Javadi AA, Giustolisi O (2008) An evolutionary-based
(2014) Attractive particle interaction forces and packing density data mining technique for assessment of civil engineering sys-
of fine glass powders. Sci Rep 4(1):1–7. https://​doi.​org/​10.​1038/​ tems. Eng Comput 25(6):500–517. https://d​ oi.o​ rg/1​ 0.1​ 108/0​ 2644​
srep0​6227 40081​08915​26
Peng T, Wang SJ (2012) Effects of land use, land cover and rainfall Ribolzi O, Patin J, Bresson LM, Latsachack KO, Mouche E, Seng-
regimes on the surface runoff and soil loss on karst slopes in taheuanghoung O, Silvera N, Thiébaux JP, Valentin C (2011)
Impact of slope gradient on soil surface features and infiltration

13
487 
Page 18 of 18 Environmental Earth Sciences (2022) 81:487

on steep slopes in northern Laos. Geomorphology 127(1–2):53– moisture content of stablished soils. Transp Infrastruct Geotech-
63. https://​doi.​org/​10.​1016/j.​geomo​rph.​2010.​12.​004 nol 5:146–168. https://​doi.​org/​10.​1007/​s40515-​018-​0053-2
Ruiz E, Ferreño D, Cuartas M, López A, Arroyo V, Gutiérrez-Solana Thakur VC, Rawat BS (1992) Geological map of the western Himalaya.
F (2020) Machine learning algorithms for the prediction of the Published under the authority of the Surveyor General of India,
strength of steel rods: an example of data-driven manufactur- Printing Group of Survey of India, p 101
ing in steelmaking. Int J Comput Integr Manuf 33(9):880–894. Tiwari VN, Pandey VHR, Kainthola A, Singh PK, Singh KH, Singh
https://​doi.​org/​10.​1080/​09511​92X.​2020.​18035​05 TN (2020) Assessment of Karmi Landslide Zone, Bageshwar,
Rukhaiyar S, Alam MN, Samadhiya NK (2018) A PSO-ANN hybrid Uttarakhand, India. J Geol Soc India 96(4):385–393. https://​doi.​
model for predicting factor of safety of slope. Int J Geotech Eng org/​10.​1007/​s12594-​020-​1567-0
12(6):556–566. https://d​ oi.o​ rg/1​ 0.1​ 080/1​ 93863​ 62.2​ 017.1​ 30565​ 2 Turing AM (2009) Computing machinery and intelligence. Parsing
Saha S, Sarkar R, Thapa G, Roy J (2021) Modeling gully erosion sus- the turing test. Springer, Dordrecht, pp 23–65. https://​doi.​org/​
ceptibility in Phuentsholing, Bhutan using deep learning and 10.​1007/​978-1-​4020-​6710-5_3
basic machine learning algorithms. Environ Earth Sci 80(8):1–21 Valdiya KS (1980) Geology of the Kumaun Lesser Himalaya. Wadia
Sakellariou MG, Ferentinou MD (2005) A study of slope stability pre- Institute of Himalayan Geology, Dehradun, p 291
diction using neural networks. Geotech Geol Eng 23:419. https://​ Van Westen CJ, Rengers N, Terlien MT, Soeters R (1997) Prediction
doi.​org/​10.​1007/​s10706-​004-​8680-5 of the occurrence of slope instability phenomenal through GIS-
Samaniego E, Anitescu C, Goswami S, Nguyen-Thanh VM, Guo H, based hazard zonation. Geol Rundsch 86(2):404–414. https://d​ oi.​
Hamdia K, Zhuang X, Rabczuk T (2020) An energy approach org/​10.​1007/​s0053​10050​149
to the solution of partial differential equations in computational Verma AK, Singh TN, Chauhan NK, Sarkar K (2016) A hybrid FEM-
mechanics via machine learning: concepts, implementation and ANN approach for slope instability prediction. J Inst Eng (India)
applications. Comput Method Appl Mech Eng 362:112790. 97:171–180. https://​doi.​org/​10.​1007/​s40030-​016-​0168-9
https://​doi.​org/​10.​1016/j.​cma.​2019.​112790 Wang SC (2003) Artificial neural network. Interdisciplinary computing
Sardooi ER, Azareh A, Mesbahzadeh T, Sardoo FS, Parteli EJR, in java programming 2003. Springer, Boston, pp 81–100. https://​
Pradhan B (2021) A hybrid model using data mining and multi- doi.​org/​10.​1007/​978-1-​4615-​0377-4_5
criteria decision-making methods for landslide risk mapping at Zhang Z (2018) Improved adam optimizer for deep neural networks.
Golestan province. Iran Environ Earth Sci 80:487. https://​doi.​ In: 2018 IEEE/ACM 26th International Symposium on Quality
org/​10.​1007/​s12665-​021-​09788-z of Service (IWQoS), IEEE, pp. 1-2
Sarkar D, Bali R, Sharma T (2018) Practical machine learning with Zhang MS, Liu J (2010) Controlling factors of loess landslides in west-
Python. Apress, New York ern China. Environ Earth Sci 59(8):1671–1680. https://​doi.​org/​
Shariati M, Fereidooni D (2021) Rock slope stability evaluation using 10.​1007/​s12665-​009-​0149-7
kinematic and kinetic methods along the Kamyaran–Marivan Zhang W, Xiao D (2019) Numerical analysis of the effect of strength
road, west of Iran. J Mt Sci 18(3):779–793. https://​doi.​org/​10.​ parameters on the large-deformation flow process of earthquake-
1007/​s11629-​020-​6438-z induced landslides. Eng Geol. https://​doi.​org/​10.​1016/j.​enggeo.​
Siddique MA, Khan MM, Arif RB, Ashrafi Z (2018) Study and obser- 2019.​105239
vation of the variations of accuracies for handwritten digits Zhang N, Zhang J, Mu Q, Yang Z (2021) Numerical modeling of
recognition with various hidden layers and epochs using neural the Xinmo landslide from progressive movement to sudden
network algorithm. In: 4th international conference on electri- failure. Environ Earth Sci 80(9):1–5. https://​doi.​org/​10.​1007/​
cal engineering and information and communication technology s12665-​021-​09651-1
(iCEEiCT), IEEE, pp 118–123. Zhao Y, Xu H, Yang T, Wang S, Sun D (2021) A hybrid recogni-
Singaravel S, Suykens J, Geyer P (2018) Deep-learning neural-network tion model of microseismic signals for underground mining
architectures and methods: using component-based models in based on CNN and LSTM networks. Geomat Nat Hazard Risk
building-design energy prediction. Adv Eng Inform 38:81–90. 12(1):2803–2834
https://​doi.​org/​10.​1016/j.​aei.​2018.​06.​004 Zheng H, Tham LG, Liu D (2006) On two definitions of the factor of
Singh TN, Gulati A, Dontha L, Bhardwaj V (2008) Evaluating cut safety commonly used in the finite element slope stability analy-
slope failure by numerical analysis—a case study. Nat Hazard sis. Comput Geotech 33(3):188–195. https://​doi.​org/​10.​1016/j.​
47(2):263. https://​doi.​org/​10.​1007/​s11069-​008-​9219-5 compg​eo.​2006.​03.​007
Singh R, Kainthola A, Singh TN (2012) Estimation of elastic con- Zheng W, Zhuang X, Tannant DD, Cai Y, Nunoo S (2014) Unified
stants of rocks using an ANFIS approach. Appl Soft Comput continuum/discontinuum modeling framework for slope stabil-
12(1):40–45. https://​doi.​org/​10.​1016/j.​asoc.​2011.​09.​010 ity assessment. Eng Geol 179:90–101. https://​doi.​org/​10.​1016/j.​
Srivastav A, Pandey VH, Kainthola A, Singh PK, Dangwal V, Singh enggeo.​2014.​06.​014
TN (2021) Numerical analysis of a collapsed tunnel: a case study Zou JZ, Williams DJ, Xiong WL (1995) Search for critical slip surfaces
from NW Himalaya, India. Indian Geotech J 1:1–3. https://​doi.​ based on finite element method. Can Geotech J 32(2):233–246.
org/​10.​1007/​s40098-​021-​00567-y https://​doi.​org/​10.​1139/​t95-​026
Stančin I, Jović A (2019) An overview and comparison of free Python
libraries for data mining and big data analysis. 2019 42nd inter- Publisher's Note Springer Nature remains neutral with regard to
national convention on information and communication technol- jurisdictional claims in published maps and institutional affiliations.
ogy, electronics and microelectronics (MIPRO). IEEE, Opatija,
pp 977–982. https://​doi.​org/​10.​23919/​MIPRO.​2019.​87570​88 Springer Nature or its licensor holds exclusive rights to this article under
Stark TD, Choi H, McCone S (2005) Drained shear strength parameters a publishing agreement with the author(s) or other rightsholder(s);
for analysis of landslides. J Geotech Geoenviron Eng 131(5):575– author self-archiving of the accepted manuscript version of this article
588. https://d​ oi.o​ rg/1​ 0.1​ 061/(​ ASCE)1​ 090-0​ 241(2005)1​ 31:5​ (575) is solely governed by the terms of such publishing agreement and
Suman S, Khan SZ, Das SK, Chand SK (2016) Slope stability analysis applicable law.
using artificial intelligence techniques. Nat Hazard 84(2):727–
748. https://​doi.​org/​10.​1007/​s11069-​016-​2454-2
Taha OME, Majeed ZH, Ahmed SM (2018) Artificial neural net-
work prediction models for maximum dry density and optimum

13

You might also like