An optimal search for neural network parameters by Salp swarm optimization algorithm: a landslide application

Abstract

This study aims at investigating the balance between exploration and exploitation search capability of a newly developed Salp swarm optimization algorithm (SSA) for fine-tuning parameters of a three-hidden-layer neural network. The landslide study was selected as a thematic application, and a mountainous area of Vietnam was chosen as a case study. A training dataset with thirteen predictor variables and historical landslide occurrences from the study area were used to train and validate the model. The experiments showed an improvement in several statistic measurements such as Root mean square error = 0.3732, Overall accuracy = 79.35%, Mean absolute error = 0.3075, and Area under Receiver operating characteristic = 0.886 in comparison to conventional benchmark methods. Based on the results, the use of SSA would enhance the search efficiency and could be used as an alternative optimizer for a multiple hidden layer neural network for landslide application as well as for other natural hazard analysis.

Full text

Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=trsl20
Remote Sensing Letters
ISSN: 2150-704X (Print) 2150-7058 (Online) Journal homepage: https://www.tandfonline.com/loi/trsl20
An optimal search for neural network parameters
using the Salp swarm optimization algorithm: a
landslide application
Huu-Duy Nguyen, Vu-Dong Pham, Quoc-Huy Nguyen, Van-Manh Pham, Minh
Hai Pham, Van Manh Vu & Quang-Thanh Bui
To cite this article: Huu-Duy Nguyen, Vu-Dong Pham, Quoc-Huy Nguyen, Van-Manh Pham,
Minh Hai Pham, Van Manh Vu & Quang-Thanh Bui (2020) An optimal search for neural network
parameters using the Salp swarm optimization algorithm: a landslide application, Remote Sensing
Letters, 11:4, 353-362, DOI: 10.1080/2150704X.2020.1716409
To link to this article: https://doi.org/10.1080/2150704X.2020.1716409
Published online: 06 Feb 2020.
Submit your article to this journal
View related articles
View Crossmark data

An optimal search for neural network parameters using the
Salp swarm optimization algorithm: a landslide application
Huu-Duy Nguyen
a
, Vu-Dong Pham
b
, Quoc-Huy Nguyen
b
, Van-Manh Pham
a
,
Minh Hai Pham
c
, Van Manh Vu
d
and Quang-Thanh Bui
b
a
Faculty of Geography, VNU University of Science, Ha Noi, Viet Nam;
b
Center for Applied Research in Remote
Sensing and GIS (CarGIS), Faculty of Geography, VNU University of Science, Ha Noi, Viet Nam;
c
Vietnam
Institute of Geodesy and Cartography, Ha Noi, Viet Nam;
d
Faculty of Environmental Science, VNU University
of Science, Ha Noi, Viet Nam
ABSTRACT
This study aims at investigating the balance between exploration and
exploitation search capability of a newly developed Salp swarm opti-
mization algorithm (SSA) for fine-tuning parameters of a three-hidden-
layer neural network. The landslide study was selected as a thematic
application, and a mountainous areaofVietnamwaschosenasacase
study. A training dataset with thirteen predictor variables and historical
landslide occurrences from the study area were used to train and
validate the model. The experiments showed an improvement in
several statistic measurements such as Root mean square
error = 0.3732, Overall accuracy = 79.35%, Mean absolute error = 0.3075,
and Area under Receiver operating characteristic = 0.886 in compar-
ison to conventional benchmark methods. Based on the results, the
use of SSA would enhance the search efficiency and could be used as
an alternative optimizer for a multiple hidden layer neural network for
landslide application as well as for other natural hazard analysis.
ARTICLE HISTORY
Received 11 October 2019
Accepted 4 January 2020
1. Introduction
Landslides are the specific formof gravitational mass movement that is a geological disaster
caused by natural, human activities and occur mainly on slopes of mountainous areas
(Motagh et al. 2013). It is, therefore, necessary for local governments to map areas that
are susceptible to landslides for damage mitigation and rescue plans. The susceptibility
studies of landslides have been widely investigated by a variety of methods, among which
the uses of machine learning are found robust in improving mapping accuracy. Examples of
these methods can be referenced from the works of (Chen et al. 2019; Ghorbanzadeh et al.
2019;Heetal.2019)byusingartificial neural network, support vector machine, fuzzy weight
of evidence logistic regression, kernel logistic regression, logistic model trees, naïve Bayes,
neural-fuzzy and the application of meta-heuristic optimization algorithms in searching for
optimal parameters of classifiers (Pham et al. 2019; Nguyen et al. 2019)
The non-free-lunch theorem (Wolpert and Macready 1997) states that no model can
solve all problems because ofthe complexities of those, and the search for potential models
CONTACT Quang-Thanh Bui qthanh.bui@gmail.com Center for Applied Research in Remote Sensing and GIS
(CarGIS), Faculty of Geography, VNU University of Science, 334 Nguyen Trai, Ha Noi, Viet Nam
REMOTE SENSING LETTERS
2020, VOL. 11, NO. 4, 353–362
https://doi.org/10.1080/2150704X.2020.1716409
© 2020 Informa UK Limited, trading as Taylor & Francis Group

to address reality applications, or specifically to landslide problems, are therefore needed.
This study investigated a novel method named SSA-MLNN that combines the Salp swarm
optimization algorithm (SSA) for fine-tuning multiple hidden layer neural network (MLNN)
parameters for the landslide analysis. Sinho, a mountainous district in a mountainous
province (Lai Chau) in Vietnam (Figure 1), was selected as the case study as the landslides
in 2018 reportedly caused significant human loss and infrastructure damages.
2. Data and method
2.1. Historical landslides and predictor variables
The machine learning application for the evaluation and construction of the landslide map
requires knowledge of landslide in the past, including the type, place, and date of occurrence
Figure 1. Study area (a), historical landslides, and predictor variables in the remaining figures.
354 H.-D. NGUYEN ET AL.

(Jaafari et al. 2019). In this study, the historical locations of landslides were collected from field
surveys during the last several years, which were extractable from http://www.canhbaotruo
tlo.vn/. The landslide locations from the website are results from a national project which is
coordinated by Vietnam Ministry of Natural resources and Environment. These landslide
locations were verified/updated by the authors in fieldsurveysandbysatellitedata.The
detected locations were, in some cases, represented by points (for small areas) or by polygons
if the areas are large enough. The proposed model was based on point dataset, and therefore
for those polygons, the centre points were extracted and combined with other points to build
up historic landslide sets. Since susceptibility mapping is a binary application so that a similar
number of non-landslide was also randomly defined across the study area. Totally, 784 points,
including 392 historical landslides and 392 non-landslide, were used for training and valida-
tion of the proposed model.
In landslide analysis, it is essential to select suitable conditioning factors for the assessment.
Fromtheliteraturereview,threegroupsof variables were used for the analysis. The first
elevation-derived group includes the ASTER Digital elevation model (DEM), which was down-
loadable from https://earthexplorer.usgs.gov/,andDEM–derivable parameters such as slope,
aspect, curvature, Compound topographic index (CTI). The second set of variables relates to
the distribution of rivers or streams, including distance to river, river density, and Stream
power index (SPI). The third consists of satellite-derived indices such as Normalized difference
vegetation index (NDVI), Normalized difference build-up index (NDBI), Normalized difference
moisture index (NDMI), and Normalized difference water index (NDWI). These indices are
good indicators for land cover conditions, and they were calculated using Landsat 8
(Operational Land Imager instrument) imagery. The last predictor variable is rainfall, which
is one of the leading causes of landslides in the mountain region because precipitation can
weaken slope surfaces. Heavy rain may reactivate landslide movements that have occurred in
the past (Wang and Sassa 2006). The rainfall layer was interpolated from 8 national meteor-
ological stations in the study area, and it represents an accumulated rain during the rainy
season in 2018. These variables were preliminarily processed and converted into a similar data
format that is 30 m x 30 m raster in WGS84/UTM zone 48N.
2.2. Neural network optimized Salp swarm optimization algorithm
The neural network is a common machine method that has been widely used for both
classification and regression applications. The network learns from the training dataset and
remembers the contribution of each predictor variable after every iteration. Several previously
published works have verified the potential replacement of gradient descend by meta-
heuristic algorithms in fine-tuning parameters of the network. This study investigated the
potential use of a newly developed optimizer SSA to search for optimal weights of a MLNN
(Figure 2). From the trial and error process, the structure of this network was determined to
have one input layer, three hidden layers with 10, 9, 8 nodes (neurons), respectively, and one
output layer. Structurally, three hundred (300) weights connect the input layer to hidden
layers, between hidden layers, and between hidden layers to output. This network produces
a susceptible value Pibetween [0 1] range which is used with the observed value Oito
formulate the objective function (Root mean square error –RMSE) for the SSA algorithm.
The brief description is shown in (Figure 2)
REMOTE SENSING LETTERS 355

SSA is used as the optimizer for the proposed neural network. It is a bio-inspired
algorithm to mimic the swarming behaviours of the salp chain and was mathematically
investigated by the work of (Mirjalili et al. 2017). Like other swarm intelligence methods, n
artificial salp population (xi=[x1
i;x2
i... :xd
i,i¼1... n) are randomly initialized in a d-
dimensional search space between lower bound (liÞand upper bound (ui). The salps move
around, searching for the location of food (fÞin the search space. The swarm is led by
a leader (x1), and the others are considered as the followers who gradually (to avoid
stagnation) update their position according to their neighbour salps and consequently to
the leader (Equation 2). The leader updates itself towards the food sources by
(Equation 1), which is also altered (exploration step) during each iteration of the optimiza-
tion process by using the following equation:
x1¼fþc1uili
ðÞc2þli
ðÞ
fg
c30
x1¼fc1uili
ðÞc2þli
ðÞ
fg
c3<0 Equation1
xi¼1
2xiþxi1
ðÞ Equation2
Where is the population size, iindicates the i
th
salp, is used to balance the exploration and
exploitation search; c2;c3are random numbers; lis the current iteration, and Lis the
predefined maximum iteration number.
The best position of source food is remembered during the updating processes and is
used as the optimal solution of the objective function. In this study, SSA is used to search
for optimal parameters of MLNN (d= 300 connecting weights), which are encoded as the
300-dimensional search space in SSA. The initial steps initialize a swarm of 30 salps, each
of which is positioned in the search space as mentioned above. The algorithm iterates
searching for optimal positions of salps, and this process terminates until the iterations
reach maximum pre-defined times (1000 was selected in this case) or when the desirable
RMSE is achieved. The position of the optimal salp (300-dimensional values in SSA or 300
connecting weights in MLNN) is used to generate the landslide susceptible map for the
entire study area.
Figure 2. Neural network optimized Salp optimization algorithm. m is the number of validation data.
356 H.-D. NGUYEN ET AL.

3. Results and discussion
3.1. Conditioning factors using Mean Gini Accuracy and Mean Decrease Accuracy
Data exploration plays a vital role in evaluating the influences of predictor variables in
landslide analysis processes. From the literature review, currently, there is no known broad
guidance to help select appropriate landslide influence factors. In general, these factors are
selected from field observation, landslide type analysis, and data availability, which are
considered the most essential elements (Chen et al. 2018). This step is to evaluate the
relationships between the factors and the historic landslide locations. Chen et al. in 2017
(Chen et al. 2017;Buietal.2015) showed that the influences of conditional factors are not the
same for each model, in which some elements have significant contributions to the predic-
tion. However, some might deteriorate the overall accuracy and should be filtered out during
the preliminary step.
In this article, condition factors were ranked by using the Random Forest algorithm
with Mean Gini Accuracy and Mean Decrease Accuracy indicators. These techniques are
considered to be one of the best methods for prioritizing variable influence levels and
widely used by researchers. It assigns a weighting to each factor to distinguish the
prediction ability. Factors have higher points that are more important for models. While
the elements have the points equal to zero, which do not contribute to the models.
Among the 13 factors (Figure 3), the topographic-derived factors received the highest
rank values in reference to the location of landslide occurrences. The ranking order
continues with rain, river density, NDWI, CTI, distance to a river, NDVI, slope, aspect, SPI,
NDBI, NDMI, and curvature. The orders of the factors came along with data analysis results
from previous studies (Pham et al. 2016; Pham, Prakash, and Bui 2018), in which DEM,
Rainfall, NDVI are the most critical factors of landslide occurrence.
3.2. Model performance comparison
Usually, when RMSE is used as the objective function for optimization algorithms, over-
fitting issues might occur. This problem occurs when machine learning models produce
higher RMSE with the validation dataset, even though the model is trained to best fittothe
training dataset. Since it is hard to collect more landslide locations due to time, budget
Figure 3. Ranking of predictor variables (a) Mean decrease in accuracy coefficient (b) Mean decrease in
Gini coefficient.
REMOTE SENSING LETTERS 357

constraints, proper resampling methods should be applied to eliminate the problem of
over-fitting. 10-fold cross-validation has been used successfully used in previous works (Bui
et al. 2019a; Shahabi et al. 2014) by averaging results from all folds, and this method was also
used for resampling training data. Besides, the prescreening, normalization of the training
dataset, the selection of dropout rate at 20%, and the limitation of the searching boundary
were also applied before and during the training. The determination of RMSE desirable
value, number of iterations, neural network dropout rates, which were considered as
a manner to avoid over-fitting, were defined after several trials.
The performance of this model was evaluated and compared to several benchmarked
methods such as Random Forest (RF), Random Subspace (RS) and Bagging, which have been
commonly used in landslide analysis (He et al. 2019; Pham, Prakash, and Bui 2018). Table 1
shows several statistical measures to evaluate the performance of proposed models in
comparison to Random Forest (RF), Random Subspace (RS), and Bagging. The experiments
ended up with SSA-MLNN at RMSE =0.3732, Mean absolute error (MAE) = 0.3075, Overall
accuracy (OA) = 79.35%, Area under Receiver operating characteristic curve (AUC) = 0.886, and
RF (RMSE = 0.3879, MAE = 0.3294, OA = 78.61%, AUC =0.868), RS (RMSE = 0.4274, MAE =0.3930,
OA = 74.21%, AUC = 0.810) and Bagging (RMSE = 0.4216, MAE = 0.3566, OA = 74.32%,
AUC = 0,809). The results of SSA –MLNN are satisfactory in comparison to the latest works
of (Nguyen et al. 2019) with the use of multiboot based naïve Bayes trees (AUC = 0.824), (Pham
et al. 2019) by using ensemble methods (the best AUC = 0.836), and radial basis function
(AUC = 0.881) in the study of (He et al. 2019). Although these works were implemented with
different dataset, it could be seen that SSA –MLNN perform well over several methods in
referenced landslide susceptibility mapping studies.
The performance of the model was visualized by plotting the False positive rate on the
x-axis against the True positive rate on the y-axis for each method, as showed in Figure 4.
The AUC measures the perfection of the algorithms with considerable values range between
0.5 and 1. If the AUC equals one, which shows that the model is perfect. The results show
that the SSA-MLNN was more efficient than the other three models (AUC = 0.886), then RF
(AUC = 0.868), RS (AUC = 0.81), and the last was Bagging (AUC = 0.809). The most optimal
model could be used to generate landslide susceptibility for the entire study area.
Figure 4(a) visualizes the search mechanism of SSA, in which the x-axis shows iterative
order and y-axis plots the average of RMSE values (from 10 cross-validations) during the
search operation. From the graph, apart from a sudden jump (which is caused by
randomization of initial salp positions) from the beginning, RMSE gradually decreases.
This variation reflects the movement of the first salp, in which exploration and exploita-
tion process is balanced by a random c1as described in the previous section. It could be
noticed that the search mechanisms make meta-heuristic algorithms unique as men-
tioned in previous works of (Bui et al. 2019b,2019c,2019d), and the verification of new
algorithms in more diverse applications is therefore needed.
Table 1. Statistical measurements of the proposed method and benchmarked classifiers.
Classifier RMSE MAE OA (%) AUC
Random Forest 0.3879 0.3294 78.61 0.868
Random Subspace 0.4274 0.3930 74.21 0.810
Bagging 0.4216 0.3566 74.32 0.809
SSA-MLNN 0.3732 0.3075 79.35 0.886
358 H.-D. NGUYEN ET AL.

After being validated, the SSA-MLNN was used to produce the landslide susceptibility
map for the entire study area. The process was implemented by feeding the whole study
area (which was represented as a raster) with associated 13 variables through the SSA-
MLNN. The output values between [0–1] range were reclassified into five classes, namely
Very low, Low, Moderation, High and Very high as shown in Figure 5. Based on the analysis
of the landslide susceptibility map, more than 22.1% of the area is in the very low risk, 32% in
the low risk, 22.4% in the moderate risk, 13.1% in the high risk, and 10.4% in the very high
risk. The Southeast and Southwest area is highly sensitive by the landslide hazard because of
the deforestation process and the construction of the infrastructure. This susceptible map is
essential for assisting land use decision-makers and proposing risk management measures
for the protection of the population through landslide predictions in the future.
Figure 4. Performance of SSA-MLNN: (a) variation of RMSE in SSA –MLNN; (b) ROC curves and AUC
values of SSA-MLNN and benchmarked methods.
Figure 5. The final landslide-susceptibility map of the study area.
REMOTE SENSING LETTERS 359

3.3. Discussion
The prediction of the surface sensitivity to landslide hazard is essential for territorial
planning, especially in the mountainous regions. The combination of advanced geospatial
techniques and machine learning allows the production of landslide susceptibility maps
more accurate and rapid. Accuracies and rapidness are the ultimate objectives of the
previous and ongoing works on the analysis of natural hazards (Akgun et al. 2012; Bui
et al. 2019a,2020). The first thing to mention is the collection of predictor variables and
preliminary processing of those. DEM, Rain, NDWI, NDVI, River density, NDBI, CTI, Distance
to river, NDMI, SPI, Slope, Aspect, and Curvature were selected because they are free and
collectable from the global portal and that means the proposed model can be reprodu-
cible in other places. Moreover, the laying of the susceptibility map over these variables
would provide useful information on which value range of each variable is most vulner-
able to landslides. So that the inherent problem solution for landslide mitigation can be
coordinated.
Another concern should be focused on how the training dataset is normalized
because the original data are measured in various units (degree, percentages, mm).
Data can be regrouped into ordinal classes as in (Bui et al. 2019a), which causes loss
of data detail or can be normalized into [0 1] value range. The second method was
used in this study, with the argument that the distribution of original data
unchanged and the susceptible values will be estimated based on the numerical
values of input data. In this study, we compared four machine learning methods:
Random forest, Random subspace, Bagging, and SSA-MLNN with this normalized
dataset. SSA-MLNN outperformed the others with higher AUC (0.886), smaller RMSE
(0.3732). Statistically, the difference between results from SSA and the others were
significant by using Wilcoxon signed ranked test.
Even though there are numerous studies on landslide analysis, but they differ from each
other on algorithms, datasets, or even structures/configurations of specific methods .i.e.,
customized neural network to fitwithspecific problems. In this study, the structure of the
neural network (the number of hidden layers, the number of neurons in each layer) was
determined based on previous works and based on the trial-and-error process with
a training dataset from the study area. However, the ‘No Free Lunch’theorem states that
no model that fits all problems, and this also means that the neural network structure is also
subject to change under different input datasets. The future work might be a focus on both
the adaptive determination of neural network structure and fine-tuning of its parameters.
This approach will be more problem independent but will require a more substantial
computation capability.
Since RMSE is used as the objective function for SSA, or generally for meta-heuristic
optimization algorithms, the over-fitting issue should be taken into consideration. This
study applied several techniques (k-fold cross-validation, the setting of drop-out rate,
limitation of the search boundary) to minimize its impacts, but some other methods can
also be tried, such as the inclusion of more training data from different geographic
locations, data augmentation. In which the collection of more training data plays
a crucial role in examining the performance of any classification method. The more data
is tested, the more reliable machine learning models will be, and decision-makers have
more accurate data to rely on for their management activities.
360 H.-D. NGUYEN ET AL.

4. Conclusion
This study investigates the potential use of Salp swarm optimization algorithm in fine-tuning
neural network parameters for landslide susceptibility mapping in a mountainous area of
Vietnam. The experiment was successful when the proposed hybrid model outperformed
other benchmarked methods in all statistical measurements, specifically (RMSE = 0.3732,
MAE = 0.3075, OA = 79.35, AUC = 0.886). Since the Wilcoxon signed ranked test was also
conducted, that means the model is a more accurate method, and it can be used as an
alternative solution for landslide study, or a potential method for other natural hazard analysis.
The innovation of machine learning and geoinformation technology makes the extrac-
tion of knowledge from spatial data more accurate and rapid, and it is particularly useful
in natural hazard management and post-disaster responses. This study verified a newly
developed optimization algorithm for landslide susceptibility mapping, but the fast-
growing of new meta-heuristic optimization algorithms provide more opportunity for
the investigation of new classification methods in natural hazard analysis. However,
methods might not be widely applicable if training dataset is limited to specific areas.
In this regard, more experiments with diverse training data are, therefore, crucial for the
verification of the applicability of new methods in solving more extensive problems.
Funding
This research is funded by Asia Research Center, Vietnam National University - Hanoi and Korea
Foundation for Advanced Studies under grant number [CA.19.8A].
ORCID
Quang-Thanh Bui http://orcid.org/0000-0002-5059-9731
References
Akgun, A., E. A. Sezer, H. A. Nefeslioglu, C. Gokceoglu, and B. Pradhan. 2012.“An Easy-to-use MATLAB
Program (Mamland) for the Assessment of Landslide Susceptibility Using a Mamdani Fuzzy
Algorithm.”Computers & Geosciences 38 (1): 23–34. doi:10.1016/j.cageo.2011.04.012.
Bui, D. T., T. Tuan, H. Klempe, B. Pradhan, and I. Revhaug. 2015.“Spatial Prediction Models for
Shallow Landslide Hazards: A Comparative Assessment of the Efficacy of Support Vector
Machines, Artificial Neural Networks, Kernel Logistic Regression, and Logistic Model Tree.”
Landslides 1–18. doi:10.1007/s10346-015-0557-6.
Bui, Q.-T., Q.-H. Nguyen, X. L. Nguyen, V. D. Pham, H. D. Nguyen, and V.-M. Pham. 2020.“Verification of
Novel Integrations of Swarm Intelligence Algorithms into Deep Learning Neural Network for Flood
Susceptibility Mapping.”Journal of Hydrology 581: 124379. doi:10.1016/j.jhydrol.2019.124379.
Bui,Q.-T.,Q.-H.Nguyen,V.M.Pham,M.H.Pham,andA.T.Tran.2019a.“Understanding Spatial Variations
of Malaria in Vietnam Using Remotely Sensed Data Integrated into GIS and Machine Learning
Classifiers.”Geocarto International 34 (12): 1300–1314. doi:10.1080/10106049.2018.1478890.
Bui, Q.-T., Q.-H. Nguyen, V. M. Pham, V. D. Pham, M. H. Tran, T. T. H. Tran, H. D. Nguyen, X. L. Nguyen,
and H. M. Pham. 2019b.“A Novel Method for Multispectral Image Classification by Using Social
Spider Optimization Algorithm Integrated to Fuzzy C-Mean Clustering.”Canadian Journal of
Remote Sensing 45 (1): 42–53. doi:10.1080/07038992.2019.1610369.
Bui, Q.-T., M. V. Pham, Q.-H. Nguyen, L. X. Nguyen, and H. M. Pham. 2019c.“Whale Optimization
Algorithm and Adaptive Neuro-Fuzzy Inference System: A Hybrid Method for Feature Selection
REMOTE SENSING LETTERS 361

and Land Pattern Classification.”International Journal of Remote Sensing 40 (13): 5078–5093.
doi:10.1080/01431161.2019.1578000.
Bui, Q.-T., M. Pham Van, N. T. T. Hang, Q.-H. Nguyen, N. X. Linh, P. M. Hai, T. A. Tuan, and P. Van Cu.
2019d.“Hybrid Model to Optimize Object-based Land Cover Classification by Meta-heuristic
Algorithm: An Example for Supporting Urban Management in Ha Noi, Viet Nam.”International
Journal of Digital Earth 12 (10): 1118–1132. doi:10.1080/17538947.2018.1542039.
Chen, W., H. Shahabi, A. Shirzadi, H. Hong, A. Akgun, Y. Tian, J. Liu, A.-X. Zhu, and S. Li. 2019.“Novel
Hybrid Artificial Intelligence Approach of Bivariate Statistical-methods-based Kernel Logistic
Regression Classifier for Landslide Susceptibility Modeling.”Bulletin of Engineering Geology and
the Environment 78 (6): 4397–4419. doi:10.1007/s10064-018-1401-8.
Chen, W., H. Shahabi, S. Zhang, K. Khosravi, A. Shirzadi, K. Chapi, B. Pham, et al. 2018.“Landslide
Susceptibility Modeling Based on GIS and Novel Bagging-Based Kernel Logistic Regression.”
Applied Sciences 8: 2540. doi:10.3390/app8122540.
Chen, W., X. Xie, J. Peng, J. Wang, Z. Duan, and H. Haoyuan. 2017.“GIS-based Landslide Susceptibility
Modelling: A Comparative Assessment of Kernel Logistic Regression, Naïve-Bayes Tree, and
Alternating Decision Tree Models.”Geomatics, Natural Hazards and Risk 8: 950–973.
doi:10.1080/19475705.2017.1289250.
Ghorbanzadeh, O., T. Blaschke, K. Gholamnia, S. R. Meena, D. Tiede, and J. Aryal. 2019.“Evaluation of
Different Machine Learning Methods and Deep-learning Convolutional Neural Networks for
Landslide Detection.”Remote Sensing 11 (2): 196. doi:10.3390/rs11020196.
He, Q., H. Shahabi, A. Shirzadi, S. Li, W. Chen, N. Wang, H. Chai, et al. 2019.“Landslide Spatial
Modelling Using Novel Bivariate Statistical Based Naïve Bayes, RBF Classifier, and RBF Network
Machine Learning Algorithms.”Science of the Total Environment 663: 1–15. doi:10.1016/j.
scitotenv.2019.01.329.
Jaafari, A., M. Panahi, B. T. Pham, H. Shahabi, D. T. Bui, F. Rezaie, and S. Lee. 2019.“Meta Optimization
of an Adaptive Neuro-fuzzy Inference System with Grey Wolf Optimizer and Biogeography-based
Optimization Algorithms for Spatial Prediction of Landslide Susceptibility.”CATENA 175: 430–445.
doi:10.1016/j.catena.2018.12.033.
Mirjalili, S., A. H. Gandomi, S. Z. Mirjalili, S. Saremi, H. Faris, and S. M. Mirjalili. 2017.“Salp Swarm
Algorithm: A Bio-inspired Optimizer for Engineering Design Problems.”Advances in Engineering
Software 114: 163–191. doi:10.1016/j.advengsoft.2017.07.002.
Motagh, M., H.-U. Wetzel, S. Roessner, and H. Kaufmann. 2013.“A TerraSAR-X InSAR Study of
Landslides in Southern Kyrgyzstan, Central Asia.”Remote Sensing Letters 4 (7): 657–666.
doi:10.1080/2150704X.2013.782111.
Nguyen, P. T., T. Tuyen, A. Shirzadi, B. Pham, H. Shahabi, E. Omidvar, A. Amini, et al. 2019.
“Development of a Novel Hybrid Intelligence Approach for Landslide Spatial Prediction.”
Applied Sciences 9: 2824. doi:10.3390/app9142824.
Pham, B. T., B. Pradhan, D. T. Bui, I. Prakash, and M. B. Dholakia. 2016.“A Comparative Study of
Different Machine Learning Methods for Landslide Susceptibility Assessment: A Case Study of
Uttarakhand Area (India).”Environmental Modelling & Software 84: 240–250. doi:10.1016/j.
envsoft.2016.07.005.
Pham, B. T., I. Prakash, and D. T. Bui. 2018.“Spatial Prediction of Landslides Using a Hybrid Machine
Learning Approach Based on Random Subspace and Classification and Regression Trees.”
Geomorphology 303: 256–270. doi:10.1016/j.geomorph.2017.12.008.
Pham, B. T., A. Shirzadi, H. Shahabi, E. Omidvar, S. K. Singh, M. Sahana, D. T. Asl, B. B. Ahmad,
N. K. Quoc, and S. Lee. 2019.“Landslide Susceptibility Assessment by Novel Hybrid Machine
Learning Algorithms.”Sustainability 11 (16): 4386. doi:10.3390/su11164386.
Shahabi, H., S. Khezri, B. B. Ahmad, and M. Hashim. 2014.“Landslide Susceptibility Mapping at
Central Zab Basin, Iran: A Comparison between Analytical Hierarchy Process, Frequency Ratio and
Logistic Regression Models.”CATENA 115: 55–70. doi:10.1016/j.catena.2013.11.014.
Wang, H., and K. Sassa. 2006.“Rainfall-induced Landslide Hazard Assessment Using Artificial Neural
Networks.”Earth Surface Processes and Landforms 31: 235–247. doi:10.1002/esp.1236.
Wolpert, D. H., and W. G. Macready. 1997.“No Free Lunch Theorems for Optimization.”IEEE
Transactions on Evolutionary Computation 1 (1): 67–82. doi:10.1109/4235.585893.
362 H.-D. NGUYEN ET AL.