Catena 184 (2020) 104240

Contents lists available at ScienceDirect

Catena
journal homepage: www.elsevier.com/locate/catena

Artificial neural network ensembles applied to the mapping of landslide T

susceptibility
L. Bragagnolo, R.V. da Silva, J.M.V. Grzybowski*
Federal University of Fronteira Sul, Rodovia RS 135 km 72, Erechim, CEP 99.700-000, RS, Brazil

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

Keywords: This study proposes a comprehensive methodology to the application of an Artificial Neural Network Ensemble
Natural disasters (ANNE) for the mapping of landslide susceptibility. The identification of susceptible areas was performed on the
Landslide susceptibility assessment basis of landslide inventory databases and seven parameters from three classes: geomorphological (elevation,
Machine learning aspect, slope, topographic moisture index, profile curvature),geological (lithology) and environmental (land
GIS
use). Studies are presented for two major cities in Brazil, Porto Alegre and Rio de Janeiro. As the main result, we
show that the susceptibility maps generated by the ANNE feature higher accuracy than those published by
official organs such as the Brazilian Geological Survey and Geotechnical Institute Foundation (Geo-Rio). This
indicates that the proposed methodology can be an effective tool to assist the development of reliable landslide
susceptibility maps in an efficient and agile manner.

1. Introduction The susceptibility to landslides can be defined as the probability of

the spatial occurrence of a landslide event based on the relationships
Landslides are regarded as natural processes resulting from the in- between the occurrence distribution and a set of predisposing factors in
teraction of the most varied environmental factors, which include a given area, such as geoenvironmental thematic variables (Guzzetti
phenomena of weathering, earthquake, precipitation and snowmelt, as et al., 2005). In turn, the prevention of damages resulting from land-
well as anthropic activities (Cubito et al., 2005). These events pose slides is based on mapping susceptibility, monitoring environmental
serious risks to life and infrastructure in different regions of the world variables and studying the dynamics and properties of chemical, phy-
(Cendrero and Dramis, 1996; Kawabata and Bandibas, 2009). The risks sical and biological agents involved. In summary, the methods used in
associated with landslides include injuries, loss of lives, direct and in- the study of landslide susceptibility can be classified into two main
direct damage to natural resources and infrastructure, such as roads, categories: direct and indirect mapping (Casale and Margottini, 1999;
bridges and communication lines (Saha et al., 2002). For this reason, Thiery et al., 2013). The direct mapping seeks to determine the degree
the zoning of landslide susceptibility is fundamental to allow the de- of susceptibility on the basis of a qualitative approach in which ex-
velopment of preventive measures and the planning of strategies of perience, knowledge about the terrain conditions and evaluation of
evacuation, in order to avoid catastrophic damages and casualties. preexisting terrain maps are greatly accounted for. This approach in-
Recent examples that illustrate the devastating effects of landslides cludes different classification processes based on the weighting of fac-
include the events recorded in China (2017), in which landslides trig- tors of influence (Martinović et al., 2016) and, according to Thiery et al.
gered by heavy rains resulted in casualties and disappearances (2013), it has two main limitations: the results are subjective, since they
(Osborne, 2017); Colombia (2017) that resulted in the loss of over 200 are linked to the experience and knowledge of the personnel; and the
lives (Dewan et al., 2017); Indonesia (2017), where at least two dozen difficulty in transcribing phenomena due to the occurrence of subtle
people were buried and at least a dozen deaths were reported (Burke, environmental changes in specific areas. These limitations make direct
2017); Mexico (2016), which caused at least 40 deaths (Trucco and methods impracticable in most developing countries due to the severe
Lopez, 2016); Guatemala (2015), with at least 130 deaths (Brumfield, lack of specialized personnel and resources. In turn, the indirect map-
2015); Brazil (2015), where more than 900 deaths were reported in ping is based on quantitative models, such as the statistical and de-
seven cities located in the mountainous region of Rio de Janeiro state terministic ones, which seek to find numerical correlations based on the
due to flooding and landslides (CEPED, 2015). relationship between the factors that favor the occurrence of landslides

*
Corresponding author.
E-mail address: jose.grzybowski@uffs.edu.br (J.M.V. Grzybowski).

https://doi.org/10.1016/j.catena.2019.104240
Received 12 January 2019; Received in revised form 4 August 2019; Accepted 23 August 2019
0341-8162/ © 2019 Elsevier B.V. All rights reserved.
L. Bragagnolo, et al. Catena 184 (2020) 104240

and the distribution of these events on the terrestrial surface different themes (Chen et al., 2017b; Huang and Gao, 2017; Lee and
(Martinović et al., 2016; Van Westen et al., 2003). However, some Kang, 2016; Pourghasemi et al., 2017; Santin et al., 2016; Zhang and
limitations may be cited, such as the sensitivity of the results to the Zimba, 2017), but its application to landslide susceptibility mapping is
quality and accuracy of the thematic data layers, and the specificity of still to be explored in more depth. The authors in Chen et al. (2017b)
certain susceptible regions that may end up not being taken into ac- proposed the use of ANN ensembles, maximum entropy (MaxEnt) and
count in some cases (Thiery et al., 2007, 2013). support vector machine (SVM) models to the identification of landslide
Several methods based on indirect mapping have been proposed and susceptible areas. The ANN-SVM was reported to feature the highest
employed in the specialized literature. For instance, one can cite area under the curve (AUC) of all (0.826), followed by ANN-MaxEnt-
probabilistic methods (Ering and Babu, 2016; Lari et al., 2014; Mondini SVM (0.811), ANN-MaxEnt (0.803), and SVM-MaxEnt (0.792). Pham
and tsung Chang, 2014; Pandey and Sharma, 2017; Park et al., 2013; et al. (2017) evaluated and compared the performance of aggregation
Robbins, 2016; Zêzere et al., 2008), logistic regression, (Atkinson and methods techniques (AdaBoost, Bagging, Dagging, MultiBoost, Rotation
Massari, 2011; Ayalew and Yamagishi, 2005; Bai et al., 2010; Colkesen Forest, and Random SubSpace) with ANNs MLP-based approach. It was
et al., 2016; Gorsevski et al., 2006; Lin et al., 2017; Ozdemir and found that the ensemble techniques significantly improved the ANN
Altural, 2013; Tsangaratos and Ilia, 2016; Wang et al., 2013), support performance.
vector machines (Huang and Zhao, 2018), fuzzy logic (Abedi In this research, we propose a framework based on ANNEs and
Gheshlaghi and Feizizadeh, 2017; Akgun et al., 2012; Aksoy and bootsrapping to improve the accuracy of landslide susceptibility map-
Ercanoglu, 2012; Ercanoglu and Gokceoglu, 2002; Ilanloo, 2011; ping. The training and validation are performed in two main Brazilian
Leonardi et al., 2016; Tien Bui et al., 2012), decision trees capitals (Rio de Janeiro and Porto Alegre) to obtain over thirty thou-
(Althuwaynee et al., 2014; Poudyal, 2013; Pradhan, 2013; Saito et al., sand ANNs from which the ones with best performance and lowest error
2009; Yeon et al., 2010) and artificial neural networks correlation are selected to compose the ANNE. The ANNE was devel-
(Alimohammadlou et al., 2014; Arnone et al., 2016; Pradhan and Saro, oped in Python programming language and applied to generate land-
2007; Chen et al., 2017b; Conforti et al., 2014; Ermini et al., 2005; slide susceptibility maps, which were then compared to those provided
Gómez and Kavzoglu, 2005; Kawabata and Bandibas, 2009; Lee et al., by the Geologic Survey of Brazil and Geotechnical Institute Foundation
2004; Melchiorre et al., 2008; Neaupane and Achet, 2004; Nefeslioglu (Geo-Rio). The manuscript is organized as follows: Section 2 presents
et al., 2008; Pham et al., 2017; Pradhan and Lee, 2010; Tsangaratos and the methods and characterizes the study areas; Section 3 presents the
Benardos, 2014; Yesilnacar and Topal, 2005; Zare et al., 2013; Kanungo results, which include the susceptibility maps and ANNE performance
et al., 2006). From those, artificial neural networks (ANNs) stand out graphs; Section 4 discusses the results and prospects further research in
for several reasons. Among them, one can cite the independence from the field; finally, Section 5 gives final remarks and summarizes the
the statistical distribution of the data (Lee et al., 2004), the in- contributions of this research.
dependence from the scale of the data, the possibility of using quali-
tative variables along with quantitative ones (Kawabata and Bandibas, 2. Materials and methods
2009); the ability to detect complex nonlinear relations among the
variables (Tu, 1996). Because ANNs are tailored to problem-solving and 2.1. Study areas
are data-driven, a great variety of existing data and available database
formats can be effectively taken as input data (Melchiorre et al., 2008; The assessment of susceptibility to landslides in this study was
Tsangaratos and Benardos, 2014; Conforti et al., 2014). carried out in two municipalities in Brazil. The first one, Porto Alegre,
The application of ANNs in the assessment of susceptibility to the capital of Rio Grande do Sul state, located in the South region
landslides was addressed by Ermini et al. (2005). Two different ap- (Fig. 1). The second, the municipality of Rio de Janeiro, capital of Rio
proaches to neural networks were applied: one was based on Multilayer de Janeiro state and located in the Southeast region (Fig. 2). These
Perceptron (MLP) and another on Probabilistic Neural Network (PNN). cities were chosen due to the availability of landslide inventory data-
The authors report that the MLP-based approach performed slightly bases and of geomorphological, geological and environmental input
better than the PNN-based approach, featuring correct classification of parameters for training, validation and testing of the ANNE.
73% of the landslides mapped, while for the PNN this number was 68%.
More recently, Tsangaratos and Benardos (2014) proposed the use of 2.1.1. Porto Alegre municipality
the Mahalanobis metric to improve the selection of non-landslide points Porto Alegre has a total area of 496.682 km² and 1,409,351 in-
and thus enrich databases used in ANN training. The product of the habitants (IBGE. Insituto Brasileiro de Geografia e Estatística, 2017a).
study was a system capable of calculating the index of landslide sus- According to De Aguiar and Hackbart (2008), there are records of more
ceptibility, where a total of 29 true positives, 29 true negatives, 3 false than 50 mm of rainfall in less than one hour and, between the years of
positives and 4 false negatives were reported, thus achieving an accu- 2005 and 2007, rainfall of 100 mm in one hour were observed. The
racy of 89.2%, from which the authors concluded that the ANN was episodes of rainfall with such characteristics favor the triggering of
able to give a satisfactory account of landslide susceptibility in the landslides and, therefore, this area is a proper case study. Indeed, being
study region. The authors in Conforti et al. (2014) also applied ANNs to considered a priority area, Porto Alegre was studied by the Geological
map landslide susceptibility in the Turbolo River basin in Calabria, Survey of Brazil in order to have flood and landslide susceptibility
Italy. By applying an ANN with three layers and 10 inputs, the authors mappings published (CPRM. Companhia de Pesquisa de
reported an overall accuracy of 87% with 91% hit in the test set. RecursosMinerais, 2015). Also, Porto Alegre has Decree n° 16.931/2011
Although the performance of ANNs is regarded as satisfactory in the (Porto Alegre, 2011) whose purpose is to monitor and to prevent dis-
mapping of landslide susceptibility and many other fields, recent stu- orderly and irregular occupation of landslide-prone areas.
dies indicate that the performance of ANNs can be further improved by
the implementation of Artificial Neural Network Ensembles (ANNEs) 2.1.2. Rio de Janeiro municipality
(Pham et al., 2017). The main reason for the potential of ANNEs is that, Rio de Janeiro covers an area of 1200.177 km² and has 6,320,446
in principle, the outputs errors of an ensemble of independently trained inhabitants (IBGE. Insituto Brasileiro de Geografia e Estatística, 2017b).
artificial neural networks (and therefore with distinct knowledge According to de Amorim et al. (2014), the municipality of Rio de Ja-
backgrounds) tend to cancel each other and consequently highlight the neiro is characterized by geographic factors that favor the occurrence of
identity of the real properties and functional relations of the phe- high-intensity rainfall. As an aggravating factor, the city presents
nomena under study (Santin et al., 2016). The application of ensembles massive irregular occupations in areas regarded as susceptible to
or committees of ANNs has been successfully applied in the study of landslides, such as hills and slopes. . These conditions contribute to the

2
L. Bragagnolo, et al. Catena 184 (2020) 104240

Fig. 1. Location map of the municipality of Porto Alegre.

triggering of erosive processes. probability of occurrence of landslides from the rainfall monitoring
Rio de Janeiro has an alert system for heavy rains and landslides through a network of rain gauges. There is also the Community Alert
named Alerta Rio(Alert Rio). This system reports four levels of and Alarm System, a project based on the Alerta Rio system that

Fig. 2. Location map of the municipality of Rio de Janeiro.

3
L. Bragagnolo, et al. Catena 184 (2020) 104240

includes SMS alerts as well as alarm networks by sirens, which are in- and aspect information were extracted from the digital elevation
stalled in 103 communities in the city. When rainfall volumes mon- models (DEM), produced by the TOPODATA (INPE. Instituto de
itored by rain gauges installed in different regions of the city reach pre- Pesquisas Espaciais, 2011) project, with a spatial resolution of 30 m.
established levels, alerts are issued and alarms are triggered, indicating
the risk of landslides and requiring residents to move to safe support 2.2.1.4. Profile curvature. The profile curvature refers to the curvature
points (Prefeitura do Rio de Janeiro, 2013b). The municipal govern- towards a slope. This parameter is considered of influence in the sliding
ment has also developed a landslide susceptibility map, which classifies processes since it guides the flow of water in the slope, making it
the regions according to susceptibility to landslide (high, medium, or convergent or divergent (Nefeslioglu et al., 2008; Yesilnacar and Topal,
low) (Prefeitura do Rio de Janeiro, 2013a). 2005). The profile curvature is usually classified into three classes,
being: concave, convex and rectilinear. Positive values reflect a concave
2.2. Datasets slope, whereas negative values indicate a convex slope. Values close to
zero represent a rectilinear structure, without inflection. The profile
2.2.1. Production of the thematic data layers curvature information for both municipalities studied was obtained
The thematic data layers consist of the input parameters for the from the products derived from the DEM TOPODATA (INPE. Instituto
training, validation, testing and application of the ANNE. Such para- de Pesquisas Espaciais).
meters exert, jointly, either directly or indirectly, some influence on the
mechanism of occurrence of landslide events. Therefore, based on the 2.2.1.5. Topographic wetness index (TWI). The topographic wetness
study by Pascale et al. (2013), the thematic geoenvironmental para- index is defined as a theoretical measure of flow accumulation and
meters to be used were defined as: lithology, soil use, aspect, elevation, hence soil moisture at any point within a basin. From this, landslide
profile curvature, slope and topographic wetness index (TWI). Each of susceptibility can be estimated as a function of the relationship of the
them is described in the following. topographic effects on the hydrological response of an area (Gómez and
Kavzoglu, 2005; Tien Bui et al., 2017). As pore pressure increases, soil
2.2.1.1. Lithology. According to Pascale et al. (2013); Akgun et al. resistance will be reduced, thus favoring the occurrence of slope faults
(2012); Chen et al. (2017a), landslides are strongly influenced by the (Ray and Jacobs, 2007; Tien Bui et al., 2017; Chen et al., 2017a; Gómez
lithological properties of the Earth's surface. Different kinds of lithology and Kavzoglu, 2005). As the topographic wetness index is intended to
influence the different levels of susceptibility (Chen et al., 2017a). This represent the topographic control of soil moisture, in this study it was
is due to the fact that this parameter is responsible for the definition of considered as an indirect measure of soil water content. In a seminal
the shear strength of the materials that constitute the slopes, thus work, Beven and Kirkby (1979) proposed the Eq. 1 for the calculation of
controlling its stability condition (Zêzere et al., 2017). The vector layers the topographic wetness index, assuming static conditions:
for the lithology parameter, for both municipalities, were extracted
from the information bank of the Geological Survey of Brazil (CPRM. A
TWI = ln ⎛⎜ s ⎟⎞
Companhia de Pesquisa de Recursos Minerais, 2017). ⎝ tan β⎠ (1)

2.2.1.2. Land-use coverage. According to Gómez and Kavzoglu (2005), where AS is the upslope contributing area per unit contour length
parameters such as cohesion, angle of friction, soil unit weight and and tan β is the local slope. Thus, the index reflects the tendency of soil
pore-pressure tend to undergo changes when subjected to different saturation developing at any point in a basin (in terms of As), as well as
types of coverage. Also, the authors in Pisano et al. (2017) observed the the tendency of the gravitational forces to provide the flow, expressed
influence of soil cover on the evaluation of areas susceptible to in terms of tan β as an approximation of the hydraulic gradient
landslides from the evaluation of the changes occurred over the (Pourghasemi et al., 2012).
years. In this sense, the soil cover is an important factor to the The TWI raster maps for each city were produced using the r.topidx
modelling of landslides. The data were obtained from the information module, available in GRASS GIS 7.2.2 software (GRASS Development
provided in the Environmental Diagnosis of the Municipality of Porto Team, 2017). The data used for the application of the selected ANNEs
Alegre (Hasenack et al., 2008), and from the vector layer made were obtained from the delimitation of the region of interest in the GIS
available on the Data Rio platform (Rio de Janeiro, 2016). software, from which the information of the rasters was extracted for
each of the previously presented parameters. The input data layers for
2.2.1.3. Elevation, slope and aspect. Although the mechanisms that links evaluation of the susceptibility to landslides are presented in Figs. 3, 4
the occurrence of landslides to the altitudes of a region are still not and 7, for the municipality of Porto Alegre and Figs. 5, 6 and 8, for the
clear, it is known that the elevation exerts influence on a considerable municipality of Rio de Janeiro.
number of biophysical parameters, as well as on the human activities
that will be developed in a certain area (Gómez and Kavzoglu, 2005), a 2.2.2. Training, validation and test data
fact that could affect soil mass stability. In turn, the slope is The landslide inventory from Rio de Janeiro is available at the
characterized as one of the main parameters that influence the website of the Geological Survey of Brazil through the SCDI (Cadastre
triggering of landslides since it is related to the velocity of the surface System of Landslides and Floods) (CPRM. Companhia de Pesquisa de
and groundwater flow and, consequently, to the moisture content of the Recursos Minerais, 2012). The landslide inventory from Porto Alegre
soil, capillary pressure and distribution of saturated and non-saturated was collected from the de Brito (2014).
zones (Gómez and Kavzoglu, 2005; Zêzere et al., 2017). Just as the role The databases for Rio de Janeiro and Porto Alegre feature 885 and
of elevation in triggering slope faults is not clearly understood, the 247 landslide records, respectively. In order to compose the training,
relationships between aspect and these events are not completely validation and test sets for each municipality, 50 landslide records of
understood (Capitani et al., 2013). However, some studies indicated the respective database were randomly extracted by means of random
the influence of this parameter on the occurrence of landslides. In sampling. Another 50 points representing non-landslide sites were ob-
general, studies report that the aspect exerts an indirect influence on tained with the aid of GRASS GIS by random spatial distribution within
the stability of a soil mass as a function of the precipitation, wind and the computational region of each municipality. Regarding the sampling
solar radiation that reach the face of the respective slope, and that, procedure, random sampling is a standard sampling method that offers
consequently, results in different conditions of humidity, weathering simplicity and low computational cost, although it may result in subpar
and predominant vegetation of the terrain (Chen et al., 2017a; Ding training and test sets and therefore cause the performance of machine
et al., 2017; Tien Bui et al., 2017). For this study, the elevation, slope learning models to deteriorate. In this case, other more elaborate

4
L. Bragagnolo, et al. Catena 184 (2020) 104240

Fig. 3. Input data layers for the municipality of Porto Alegre (a) Aspect, (b) Slope, (c) Elevation, (d) Profile curvature, (e) TWI, and (f) Satellite image of the area.

techniques are available that can handle the sampling process in a were predominantly located in areas of eluvial and colluvial deposits.
systematic way (Marjanović et al., 2011; Hong et al., 2018; Goetz et al., As for Rio de Janeiro, the prevailing land-use for non-landslide points
2011). was in urbanized regions.
The selection of a smaller set of records is justified by the un- From the vector layers and the landslide and non-landslide events,
certainties associated with the definition of non-landslides points. A the respective information was extracted for each point using the
high number of records would require the definition of a greater GRASS GIS software. Landslide event characteristics were identified
number of negative examples, which could harm the learning of the with value 1 and, non-landslide, with value 0. Thus, two input matrices
ANNs. On the other hand, in order to make it possible for the network to of 100 × 7 were generated for each city, with 100 being the number of
learn the existing patterns in places where there were no landslides records and 7 the number of parameters used.
reported, 50 points were randomly selected for each municipality to The 100 records were randomly separated into 3 distinct sets, in the
serve as records of “ non-landslide” sites. following proportions: 70% of records for training, 15% for validation
Since it is not feasible from the available data to distinguish one and 15% for testing. The first two were randomly re-sampled for the
type of landslide from another, this research looks for all kinds of development of individual neural networks, as will be detailed in
landslides. It must be mentioned that the datasets used in this study do Section 2.3.1. Data were normalized to the interval [0,1] to standardize
not provide any information other than the landslide location and the the scale of the stimuli supplied to the network.
year of occurrence. In Brazil, not only landslide datasets are very scarce
in number, but also they commonly lack basic information such as the 2.3. Artificial neural network ensemble (ANNE)
date of the event. Thus, the point is to search for an efficient and agile
methodology that can provide reliable estimates even in this scenario of The ANN was developed in the Python language based on the
scarce and poorly detailed data. principle of multilayer perceptron (MLPs) (Haykin, 1999). This form of
The spatial distribution of the landslides events can be seen in neural network consists of a structure that features a set of sensory
Figs. 7 and 8. The statistical characterization of the landslides and non- neurons, constituted by the input layer, in which the parameters that
landslides data for Porto Alegre and Rio de Janeiro is presented in the are used for the training, validation and test are inserted; another set of
Tables 1 and 2, respectively. neurons located in the so-called intermediate or hidden layer, in addi-
Regarding the nominal qualitative parameters lithology and land- tion to a layer of output neurons, which in this case will result in a
use, it was verified that the lithological unit with the highest occurrence measure of landslide susceptibility of a given pixel, on a scale ranging
of landslides in the municipality of Rio de Janeiro refers to the Rio from 0 (low susceptibility) to 1 (high susceptibility). In the hidden
Negro complex, while the non-landslide points were concentrated in the layer, the variables from the input layer are processed allowing the
region that encompasses fluvial-lagoon deposits. Meanwhile, the pre- network to learn complex tasks in order to progressively extract sig-
dominant land-use both for landslide and non-landslide points occurred nificant characteristics of the input patterns. The early stopping pro-
in urban areas. For Porto Alegre, events were concentrated in areas of cedure was applied during the training in order to avoid overfitting, i.e.,
Facies Serra do Herval lithologic unit, and the non-landslide points excessive fitting of the ANN to the training data (Haykin, 1999). A

5
L. Bragagnolo, et al. Catena 184 (2020) 104240

maximum number of epochs of 200 was allowed and the learning rate
was set to α = 0.6. Fig. 9 shows the structure of a multi-layered neural
network and the ANNE.

2.3.1. Bootstrapping
The ANNE was implemented based on the bootstrapping aggrega-
tion method (Koch et al., 2013), in which different sets of training and
validation data were randomly generated, as well as initial synaptic
weights and network architectures, in order to provide diverse and
varied learning stimuli. For the definition of the maximum limit of
neurons in the intermediate layer, equation proposed by Netch-Nielsen
(1987) was applied:
NH ≤ 2NI + 1 (2)

where NH is the number of neurons in the intermediate layer and NI is

the number of input parameters. Considering 7 input parameters, the
maximum number of neurons in the layer adopted was 15, according to
Eq. (2). Thus, the individual networks were created from 100 resamples
of the training and validation sets and 30 sets of initial conditions for
each architecture (number of neurons in the intermediate layer), with a
minimum of 2 and maximum of 15 neurons, adding up to the total of
42,000 ANNs.

2.3.2. Selection and test

The ANNs that obtained the best performances in the test set and,
additionally, low-correlated output errors were selected to compose the
ANNE. The low correlation of output errors is aimed at ensuring that
the networks not only give accurate results but also that they have
distinct learning backgrounds, such that they established themselves in
different local minimums in the parameter space of synaptic weights. Of
the total number of ANNs, 6000 individual networks were selected in
the first stage to performance. In the second stage, the networks se-
lected in the first stage were tested for correlation of output errors by
means of the Pearson correlation coefficient. Then, those with corre-
lation p > 0,5 related to previously selected ANNs were discarded from
Fig. 4. Land use for the municipality of Porto Alegre. the ensemble. The output of the ANNE (OE) is obtained by means of the
application of Eq. (3), in which the mean of the responses of each in-
dividual ANN (Oj) that belongs to the ensemble is computed. As the
composition of the ANNE was completed, it was applied to the eva-
luation of the areas of interest.

Fig. 5. Input data layers for the municipality of Rio de Janeiro (a) Aspect, (b) Slope, (c) Elevation, (d) Profile curvature, (e) TWI, and (f) Satellite image of the area.

6
L. Bragagnolo, et al. Catena 184 (2020) 104240

Fig. 6. Land use for the municipality of Rio de Janeiro. .

n
1
OE =
n
∑ Oj
j=1 (3)

2.3.3. Sensitivity assessment

The evaluation of parameter sensitivity for each input parameter of
the network was performed for the ANNE by monitoring changes in the
output due to perturbations in one input parameter at a time, while the
others were maintained constant in their baseline values. All evalua-
tions were performed with normalized values in the interval [0,1].

2.3.4. Performance measures

In order to measure the performance of the trained neural network,
the accuracy, precision and recall metrics were calculated according to
the following equations:
Tp + Tn
Accuracy =
Tp + Tn + Fp + Fn (4)

Tp
Precision =
Tp + Fp (5)

Tp
Recall =
Tp + Fn (6)
where Tp is the true positives, Tn is the true negatives, Fp is the false
positives and Fn is the false negatives. Values closer to 1 indicate better
performance.

3. Results

Fig. 7. Location of landslide and non-landslide points and lithology map for the 3.1. Training, validation, test and sensitivity assessment
municipality of Porto Alegre.
The training, validation, test and sensitivity analysis of the ANNs
and the composition of the ANNE for each municipality took about 12 h
to complete on a 64-bit Intel® Core™ i5-5200U @ 2.2 GHz processor
computer with 6 GB of RAM.

7
L. Bragagnolo, et al. Catena 184 (2020) 104240

Fig. 8. Location of landslide and non-landslide points and lithology map for the municipality of Rio de Janeiro.

3.1.1. Porto Alegre 0.0113°/m to 0.1552°/m, which represents the prevalence of concave
From the individual ANNs with the best performances it was pos- slopes (Fig. 10).
sible to choose two uncorrelated networks (Pearson correlation coeffi-
cient p < 0,5) to compose the ANNE. The characteristics of the selected 3.1.1.2. Sensitivity assessment. The parameters that presented the
networks are presented in Table 3. ANNs that obtained highly corre- greatest influence on the susceptibility results according to the
lated with the ones selected were discarded since high correlation in- Table 4 were the profile curvature, followed by the TWI and slope.
dicates similar learning patterns which resulted in similar error pat-
terns. Therefore, such ANNs were not adequate to compose the ANNE. 3.1.2. Rio de Janeiro
Recall that the main objective of the application of an ensemble instead From the selected networks, it was possible, as in the case of Porto
of an individual network is the cancellation of part of the output errors Alegre, to choose 2 uncorrelated ANNs, with Pearson correlation coef-
as a means of providing more accurate results. ficient p < 0.5. The characteristics of the selected networks for the
The test results for Porto Alegre resulted in 100% correct classifi- composition of the ensemble are presented in Table 3. As in the case of
cations, with no false-positive or false-negative results. Because of this, Porto Alegre, the ANNE was able to provide 100% correct classifica-
all performance measures calculated resulted in 1. For purposes of tions in the test set. No false positives or false negatives were observed.
classifying the outputs of the ANNE, which is a decimal number in the Thus, all performance measures resulted equal to 1.
interval [0,1], the threshold 0.5 was applied applied on the basis of
study of the Area Under Curve (AUC), such that a given point was 3.1.2.1. Landslides and conditioning factors. It was observed that the
classified as a landslide point whenever the ANNE output give values areas with highest susceptibility (considering values above 0.8) are
equal to or greater than 0.5, and non-landslide point, otherwise. mostly found in the lithological unit of Rio Negro Complex, in
residential areas composed of regions with slopes between 0 and
3.1.1.1. Landslides and conditioning factors. It was verified that the 30.41%, slopes oriented between 143.40° and 215.60°, elevations
areas with higher susceptibility (considering a susceptibility greater from 0 to 233.79 m, TWI varying between 4.35 and 6.91 and profile
than 0.8) were found in the lithological unit Fácies Serra do Herval, in curvatures from − 0.3733° to − 0.01174°, representing convex slopes
residential areas, regions with slopes between 16.78% and 33.54%, (Fig. 11).
slopes oriented between 0° and 71.20°, elevations of 62.05 m to
126.63 m, TWI varying between 5.95–8.23 and profile curvature from 3.1.2.2. Sensitivity assessment. For the case of Rio de Janeiro, the

Table 1
Statistical analysis of the dataset used for ANNE training, validation and testing referring to the municipality of Porto Alegre.
Slope (%) Aspect (°) Elevation (m) TWI Profile curvature (°/m)

Landslide points Mean 26.04 214.20 120.62 7.10 − 4.6 × 10−3

Variance 126.42 13,763.12 3330.05 1.27 5.8 × 10−4
Non-landslide points Mean 7.81 189.23 48.35 7.94 − 2.2 × 10−3
Variance 51.99 9847.5 2393.70 2.13 2.6 × 10−4

8
L. Bragagnolo, et al. Catena 184 (2020) 104240

Table 2
Statistical analysis of the dataset used for ANNE training, validation and testing referring to the municipality of Rio de Janeiro.
Slope (%) Aspect (°) Elevation (m) TWI Profile curvature (°/m)

Landslide points Mean 20.77 184.87 55.92 7.35 − 1.8 × 10−2

Variance 212.11 10,800.17 2522.11 2.23 1.6 × 10−3
Non-landslide points Mean 12.09 220.82 32.67 8.07 − 9.83 × 10−3
Variance 202.27 11,009.36 1048.85 2.69 4.99 × 10−4

3.2. Landslide susceptibility assessment

From the ANNEs composed on the basis of the proposed metho-

dology, it is noticeable that they were able to identify landslide events
with satisfactory performance within the test set. In the following, the
results obtained by means of the ANNE will be compared to those ob-
tained by means of classic approaches, as it is the case of the landslide
susceptibility maps for Porto Alegre and Rio de Janeiro. At first glance,
it is noticeable that the maps of susceptibility generated by the ANNE
feature a high level of similarity with the maps from official organs.
This indicates that the ANNE was able to extract patterns and features
from the mapping between input and output to make an evaluation of
landslide susceptiblity which is consistent with those obtained by
means of traditional methods. This can be seen in Figs. 12 and 13. The
ANNE output values for the Porto Alegre and Rio de Janeiro ranged
from 0 to 0.9999, where values near 1 indicate greater probability of
occurrence of a landslide, while low values such as the ones near 0
show a low probability. The values of the thresholds for delimitation of
landslide susceptibility classes can be set as follows. First, the objective
meaning of high, medium and low susceptibility has to be defined in
terms of numbers. As for our case studies, we classified as high sus-
ceptibility class the set of points that would encompass at least 65% of
the landslide events in the database, while for the medium suscept-
ibility class this number would be 25%. Second, on the basis of such
choice, the values of the threshold were computed using the histogram
of the ANNE outputs. As such, the threshold values are site-specific,
since the ANNE outputs depend on data and training. Further, the
choice of such percentages allows for the generation of more or less
conservative susceptibility maps. Following this procedure, the
threshold values can be directly computed and the corresponding sus-
ceptibility maps can be given an objective interpretation. Needless to
say, one can delimit susceptibility in a larger or smaller number of
classes, according to necessity. Following this procedure, the threshold
values obtained for Porto Alegre were 0.48 and 0.99 and for Rio de
Janeiro were 0.50 and 0.90. It shall be remarked that the results depend
on the definition of optimal limits for classification, since more con-
servative thresholds would result in less false negatives, i.e., un-
Fig. 9. Structuring an ANNE. (a) Architecture of an individual RNA 7-5-1 recognized landslide-prone areas, but in a greater number of false po-
(seven input parameters, five neurons in the hidden layer and a response sitives, in which safe areas may be classified as risky. This type of
neuron) and (b) ANNE composition: a set of individual networks is selected and
approach may be appropriate for territorial planning tasks or in pre-
from application of an H(O) function, the final answer is generated.
liminary phases of risk analysis. However, less conservative approaches
would might lead to reduction of false-positives and increase of false-
Table 3 negatives, which are undesirable in most cases (Ermini et al., 2005).
Characteristics of the selected networks for the ANNE composition for the Fig. 14a and b show the classification of the unseen records (land-
municipalities of Porto Alegre and Rio de Janeiro. slide points not sen by the ANNE during the training, validation and test
Porto Alegre Rio de Janeiro steps) superimposed on the map with landslide susceptibility classes. It
can be observed that the number of landslide events located in the high
Test error norm 0.00529 0.02518 0.6442 1.159
susceptibility class is higher for the susceptibility map generated by the
Neurons in the hidden layer 12 12 9 11
ANNEs than for the maps used as reference, in both cases. This indicates
that the ANNE performed better at classifying the areas into the classes
parameters that presented the greatest influence on the susceptibility of landslide susceptibility classes.
results, according to Table 4, were profile curvature, slope, TWI and From the susceptibility maps generated by the ANNE, a total area of
land-use. 395.32 km2 (84.74%) was identified for Porto Alegre, located in areas
of low susceptibility, 59.42 km2 (12.74%) in areas of medium sus-
ceptibility and 11.79 km2 (2.53%) in regions of high susceptibility. As
for the municipality of Rio de Janeiro, the areas of low, medium and

9
L. Bragagnolo, et al. Catena 184 (2020) 104240

Fig. 10. Relation between regions susceptible to landslides and conditioning factors for the municipality of Porto Alegre.

high susceptibility represent, respectively, 532.31 km2 (46.64%), of problem (Ering and Babu, 2016; Atkinson and Massari, 2011; Akgun
436.92 km2 (38.28%) and 172.11 km2 (15.08%). These results give et al., 2012; Althuwaynee et al., 2014; Zêzere et al., 2008), it is ne-
clear indications of hazardous areas that should be avoided or regarded cessary to develop them more to the point that they become trust-
with care in what concerns human occupation or construction of in- worthy and largely applied in real-world situations, such that tragedies
frastructure. such as those cited in the introduction to this paper can be avoided or
mitigated to the most. One step towards that is to reduce the limitations
of current approaches, to deal adequately with the diversity of para-
4. Discussion meters employed (Gómez and Kavzoglu, 2005), with the lack of ade-
quate databases and to pursue effectiveness and agility in the genera-
Landslide susceptible areas can be identified using different tech- tion of susceptibility maps for large regions of uncovered areas, as
niques, with the use of different factors and methods (Zare et al., 2013). found in developing countries. According to Tsangaratos and Benardos
Despite of the existence of several methodologies applicable to this type

10
L. Bragagnolo, et al. Catena 184 (2020) 104240

Table 4 (2014), the lack of complete understanding of the complex physical

Sensitivity of the model for the input parameters for the Porto Alegre and Rio de behaviors of some geoenvironmental phenomena often results in
Janeiro cities. oversimplifications or inadvertent adoption of dubious assumptions
Variation of the final result into models. This is likely to produce distortions and errors in results
and, as a consequence, inadequate predictions from mathematical
Parameter Porto Alegre Rio de Janeiro models. In this sense, the use of non-parametric techniques, such as
ANNEs, can be regarded as promising, not only due to their accuracy,
Aspect 9.0 × 10−5 4.0 × 10−4
Elevation 6.0 × 10−5 1.1 × 10−3 but especially because they are data-driven, i.e., they work directly
Land use 5.0 × 10−5 1.6 × 10−2 with true parameters and data. In addition, as this type of methodology
Lithology < 1.0 × 10−5 < 1.0 × 10−4 applies learning algorithms, it is the ANNs themselves that define their
Profile curvature 6.7 × 10−1 7.3 × 10−1
own architecture from the best quantitative description of the phe-
Slope 2.1 × 10−3 4.9 × 10−2
TWI 2.3 × 10−3 2.9 × 10−2
nomenon under study. This shall be taken into account in the choice of
responsive and supportive methodologies to deal with environmental

Fig. 11. Relation between regions susceptible to landslides and conditioning factors for the municipality of Rio de Janeiro.

11
L. Bragagnolo, et al. Catena 184 (2020) 104240

Fig. 12. Landslide susceptibility maps (a) generated from ANNE and showing the points of events of landslides that occurred in the municipality of Porto Alegre and
(b) landslide susceptibility map produced by the Brazilian Geological Survey (CPRM. Companhia de Pesquisa de RecursosMinerais, 2015).

Fig. 13. Landslide susceptibility maps (a) generated from ANNE and showing the points of events of landslides that occurred in the municipality of Rio de Janeiro
and (b) landslide susceptibility map produced by the Rio de Janeiro Geotechnical Institute Foundation (Geo-Rio) (Prefeitura do Rio deJaneiro, 2013a).

problems with a large number of parameters and complex physical being a black-box modelling approach, ANNs do not require deep a
description. understanding of the underlying physical mechanisms of the phenom-
Since ANNs can be tailored for parallel computing platforms, the enon under study, while for the same reason it is by far not as insightful
methodology can greatly benefit from the rapid development of General as white-box modelling approaches.
Purpose Graphics Processing Units (GPGPUs), in which hundreds or The use of ANNEs allow the identification of areas of susceptibility
even thousands of processors run training kernels in parallel. Further, in an efficient and agile way. As illustrated by the results, areas con-
ANNs are greatly versatile in the sense that they can be trained over a sidered as ones with high susceptibility can be promptly identified by a
varied range of databases with different input parameters and still properly trained algorithm. Although the results from the ANNEs
provide satisfactory results, being the only requirement that the avail- should not be regarded as definitive, they provide a prompt and valu-
able data are sufficiently informative. Additionally, ANNs allow expert able starting point to be complemented and evaluated by experts. In
knowledge to be straightforwardly regarded in the learning process. this sense, the ability of machine learning techniques to rapidly find a
This feature is by far not as straightforward in methodologies based on way through a considerable amount of data and promptly provide di-
logistic regression, for example (Ozdemir and Altural, 2013; Wang agnostics can greatly improve the process of generating knowledge that
et al., 2013; Bai et al., 2010; Hosmer and Lemeshow, 2000). Finally, is important to the environmental safety of cities, allowing preventive

12
L. Bragagnolo, et al. Catena 184 (2020) 104240

and possibly other developing countries due to the lack of experts and
the precariousness of civil defense organs, which are severely limited by
the lack of personnel and infrastructure. In addition, according to
Ermini et al. (2005) the responses of artificial neural networks can be
hampered by noises in the input parameter data, especially lithology
and land-use, as well as in the output data, the latter particularly re-
lated to survey errors.

4.1. ANNEs performance

It was verified that the ANNEs perform well in the identification of

areas susceptible to landslides, as observed in the results of the test
stage for both municipalities. The ANNE was able to classify all the
records from the test set correctly, thus no false positives or false ne-
gatives were observed. However, one of the limitations was the the
small number of ANNs with low correlations, which severely limited
the number of networks available to compose the ensemble. In part, this
impairs the evaluation of the performance of the ensemble.
Nevertheless, the ensembles present a greater background of knowl-
edge, as can be observed in the variation of test errors, which are
smaller for the ANNE. Anyway, even with a low number of selected
networks, the ANNE presented satisfactory results, being able to ef-
fectively learn and generalize. This fact is still corroborated by the high
quantitative agreement of the location of the unseen landslide points
(those not used in the training of the networks) in which 85.5% of the
records for the municipality of Porto Alegre and 92.1% for the muni-
cipality of Rio de Janeiro were located in areas of medium and high
susceptibility. The comparison between the location of the landslide
events for the maps produced in the present study and the references
from CPRM and Rio de Janeiro municipal government showed that
ANNE was more accurate in the sense that most of the unseen records
Fig. 14. Landslide records not used in the training, test and validation steps
are located in areas classified as highly susceptible to landslides.
according to the classification low (1), average (2) and high (3) susceptibility in
Meanwhile, for the reference maps, such accuracy was not observed, as
comparison between the maps generated by the ANNEs and the maps used as
reference. can be seen in the corresponding histograms (Fig. 14a and b).

4.2. Landslides, conditioning factors and sensitivity assessment

or even emergency measures to be taken timely. The main advantages
are the reduced time required to identify susceptible areas, the possi- Both in Porto Alegre and Rio de Janeiro, residential areas were
bility of using freely available data and different types and combina- those that prevailed as regions with high susceptibility to landslides.
tions of input variables. Besides, according to Santin et al. (2016) and The geomorphological changes promoted by anthropic action from, for
Tsangaratos and Benardos (2014), as the learning of the networks is example, slope cuts and increase of the load on them, result in the
given from the relations between input and output data, simplifications, decrease of the resistance of the soil, which favors the triggering of
assumptions and approximations end up being avoided. Likewise, it is landslides. Most susceptible areas also prevailed in slopes between
easy to compile, normalize, resize, reprocess and overlap input and 16.78% and 33.54% in Porto Alegre, and from 0% to 30.41% in Rio de
output data from any Geographic Information System (GIS) software Janeiro. According to Zare et al. (2013), as the slope increases, the
(Tsangaratos and Benardos, 2014; Openshaw and Openshaw, 1997). In shear stress in the soil or other unconsolidated material usually also
addition, ANNEs are able to handle complex data patterns, identify increases. It is expected that the smooth slopes will have a low fre-
subtle patterns present in training input and solve problems with un- quency of mass movements, due to the generally smaller shear stresses.
known patterns present in the input data (Gómez and Kavzoglu, 2005; Regarding the aspect parameter, slopes oriented to the North and
Conforti et al., 2014). Northeast directions presented a higher probability of occurrence of
Landslide susceptibility mapping is an important source of in- mass movements, for the municipality of Porto Alegre; meanwhile,
formation to planning entities, public organs and engineers in choosing Southeast and South presented a higher probability of occurrence of
appropriate areas for the deployment of infrastructure, thus enhancing mass movements for the municipality of Rio de Janeiro. This is directly
land-use and mitigating the effects of inadequate land use on the trig- related to the local conditions of each area, being influenced by the
gering of landslides with potential risks to human lives. The application wind direction, solar radiation received, prevailing direction of pre-
of ANNEs, as shown in the results, can help promote a more adequate cipitation events, causing a greater or lesser degree of weathering on
and efficient planning of land-use, at least as a preliminary assessment these areas (Ozdemir and Altural, 2013; Pham et al., 2017; Varnes,
of large areas, as showed in the case studies; whereas, on a local scale, it 1984; Pourghasemi et al., 2012; Ayalew and Yamagishi, 2005).
is considered necessary to follow specialists from this area of knowl- The authors in Pachauri and Pant (1992) reported that greater
edge, not because the methodology itself is inappropriate, but because elevations result in a higher degree of susceptibility to landslides.
the level of detail of satellite imagery from which the thematic layers However, other studies have observed the tendency that landslide
are obtained do not generally provide enough spatial resolution for a susceptibility decreases as elevations increases (Gorsevski et al., 2006;
detailed account on a local level. Zhang et al., 2012; Meena et al., 2019). In Brazil, regions with greater
Although the application of ANNEs was shown to have advantages, elevations have a more intense vegetation cover since they are regarded
it also has limitations, such as the dependence on the existence of ap- as permanent preservation areas (PPA), in such way that this behavior
propriate landslide databases. This is especially problematic in Brazil may be expected in the study areas. Indeed, the most susceptible areas

13
L. Bragagnolo, et al. Catena 184 (2020) 104240

fit between altitudes of 62.05 m and 126.63 m for Porto Alegre and 0 to References
233.79 m for Rio de Janeiro, although there are much greater eleva-
tions in the study regions. On the other hand, according to the land-use Abedi Gheshlaghi, H., Feizizadeh, B., 2017. An integrated approach of analytical network
map, the areas with lower elevations had the greatest interference from process and fuzzy based spatial decision making systems applied to landslide risk
mapping. J. Afr. Earth Sci. 133, 15–24. https://doi.org/10.1016/j.jafrearsci.2017.05.
anthropic activities (Ding et al., 2017). 007.
Regarding the TWI, similar results were observed by Hong et al. Akgun, A., Sezer, E.A., Nefeslioglu, H.A., Gokceoglu, C., Pradhan, B., 2012. An easy-to-use
(2015), where it was found that values below 5 resulted in a higher MATLAB program (MamLand) for the assessment of landslide susceptibility using a
Mamdani fuzzy algorithm. Comput. Geosci. 38 (1), 23–34. https://doi.org/10.1016/
frequency of landslides. In general, the highest slopes have the lowest j.cageo.2011.04.012.
index values. Likewise, regions with larger drainage areas and smaller Aksoy, B., Ercanoglu, M., 2012. Landslide identification and classification by object-based
slopes will produce higher values of the TWI (Gómez and Kavzoglu, image analysis and fuzzy logic: an example from the Azdavay region (Kastamonu,
Turkey). Comput. Geosci. 38 (1), 87–98. https://doi.org/10.1016/j.cageo.2011.05.
2005). Besides, according to Lamb (1996), shallow landslides can occur 010.
on slopes when precipitation water infiltrates the soil and eliminates Alimohammadlou, Y., Najafi, A., Gokceoglu, C., 2014. Estimation of rainfall-induced
suction, thereby reducing apparent cohesion. landslides using ANN and fuzzy clustering methods: a case study in Saeen Slope,
Azerbaijan province, Iran. Catena 120, 149–162.
In the case of profile curvature, more susceptible areas tend to be
Althuwaynee, O.F., Pradhan, B., Ahmad, N., 2014. Landslide susceptibility mapping using
those with a profile tending to the concave shape for Porto Alegre, decision-tree based CHi-squared automatic interaction detection (CHAID) and
while the opposite behavior was observed for Rio de Janeiro. According Logistic regression (LR) integration. IOP Conf. Ser.: Earth Environ. Sci. 20, 012032.
to Chen et al. (2017b), landslides are expected to occur on slopes that http://stacks.iop.org/1755-1315/20/i=1/a=012032?key=crossref.
aa9a3e1fe54a9d463f8afb7b2af06e13%5Cnhttp://iopscience.iop.org/1755-1315/
have concave curvature, since this feature is capable of resulting in the 20/1/012032.
overloading of a slope by means of the volume of concentrated water Arnone, E., Francipane, A., Scarbaci, A., Puglisi, C., Noto, L.V., 2016. Effect of raster
and the resulting shear stress. However, according to Ohlmacher resolution and polygon-conversion algorithm on landslide susceptibility mapping.
Environ. Model. Softw. 84, 467–481. https://doi.org/10.1016/j.envsoft.2016.07.
(2007), results that differ from the theoretical view can be given as a 016.
result of the resistive forces between the soil particles arising from the Atkinson, P.M., Massari, R., 2011. Autologistic modelling of susceptibility to landsliding
convergence of materials. in the Central Apennines, Italy. Geomorphology 130 (1-2), 55–64.
Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regression for
Finally, it is emphasized that the parameters used as inputs to the landslide susceptibility mapping in the Kakuda-Yahiko Mountains, Central Japan.
ANNE should not be evaluated separately, considering the complexity Geomorphology 65 (1-2), 15–31.
of the relationships between them to trigger a landslide event. As for the Bai, S.B., Wang, J., Lü, G.N., Zhou, P.G., Hou, S.S., Xu, S.N., 2010. GIS-based logistic
regression for landslide susceptibility mapping of the Zhongxian segment in the Three
sensitivity evaluation, it was verified that the parameters that had the Gorges area, China. Geomorphology 115 (1-2), 23–31. https://doi.org/10.1016/j.
greatest influence on the prediction of landslides were profile curva- geomorph.2009.09.025.
ture, slope, TWI and land-use, the latter only for of Rio de Janeiro. Such Beven, K.J., Kirkby, M.J., 1979. A physically based, variable contributing area model of
basin hydrology. Hydrol. Sci. Bull. 24 (1), 43–69.
parameters are often applied in the different methods for the study of
Brumfield, B., 2015. Death Toll After Guatemala Landslide Tops 130. http://edition.cnn.
areas susceptible to these events (Ozdemir, 2011; Pascale et al., 2013; com/2015/10/04/americas/guatemala-landslide-dead-missing/index.html?sr=
Ilia and Tsangaratos, 2016; Ozdemir and Altural, 2013; Chen et al., fb10415guatemalalandslidedeadmissing759aVODtopLink.
2017a; Pourghasemi et al., 2013; Wang et al., 2013). In addition, a Burke, D., 2017. Dozens Feared Dead as Landslide Hists Homes in Indonesia. http://www.
dailymail.co.uk/news/article-4370468/Dozens-feared-dead-Indonesian-landslide.
study by Pradhan and Saro (2007) also found that the two most im- html.
portant parameters on the prediction mechanism for landslide sus- Capitani, M., Ribolini, A., Bini, M., 2013. The slope aspect: a predisposing factor for
ceptibility were slope and curvature. The difference in the order of landsliding? Compt. Rendus Geosci. 345 (11-12), 427–438. https://doi.org/10.1016/
j.crte.2013.11.002.
relevance of these parameters between the study areas (Table 4) can be Casale, R., Margottini, C., 1999. Floods and Landslides: Integrated Risk Assessment.
justified by the different patterns of geological, geomorphological and Springer Berlin Heidelberg, Heidelberg.
environmental parameters in each area. Cendrero, A., Dramis, F., 1996. The contribution of landslide to landscape evolution in
Europe. Geomorphology (15), 191–211.
CEPED, 2015. 2011 – Inundações e Deslizamento na Região Serrana do Rio de Janeiro.
5. Final remarks http://www.ceped.ufsc.br/2011-inundacoes-e-deslizamento-na-regiao-serrana-do-
rio-de-janeiro/.
Chen, W., Pourghasemi, H.R., Panahi, M., Kornejady, A., Wang, J., Xie, X., Cao, S., 2017a.
This research employed an Artificial Neural Network Ensemble
Spatial prediction of landslide susceptibility using an adaptive neuro-fuzzy inference
(ANNE) to generate landslide susceptibility maps and presented two system combined with frequency ratio, generalized additive model, and support
case studies to illustrate the potential of the methodology. The results vector machine techniques. Geomorphology 297, 69–85. https://doi.org/10.1016/j.
geomorph.2017.09.007.
indicate that the ANNEs were able to identify the relationships between
Chen, W., Reza, H., Kornejady, A., Zhang, N., 2017b. Landslide spatial modeling: in-
the conditioning factors and could generate consistent susceptibility troducing new ensembles of ANN, MaxEnt, and SVM machine learning techniques.
maps which agree with those obtained by means of traditional meth- Geoderma 305, 314–327. https://doi.org/10.1016/j.geoderma.2017.06.020. May.
odologies. The sensitivity of landslide susceptibility to seven predis- Colkesen, I., Sahin, E.K., Kavzoglu, T., 2016. Susceptibility mapping of shallow landslides
using kernel-based Gaussian process, support vector machines and logistic regression.
posing factors were assessed and it was observed from the case studies J. Afr. Earth Sci. 118, 53–64. https://doi.org/10.1016/j.jafrearsci.2016.02.019.
that the ones with the greatest influence were profile curvature, topo- Conforti, M., Pascale, S., Robustelli, G., Sdao, F., 2014. Evaluation of prediction capability
graphic wetness index, land use and slope. of the artificial neural networks for mapping landslide susceptibility in the Turbolo
River catchment. Catena 113, 236–250. https://doi.org/10.1016/j.catena.2013.08.
The results indicate that ANNEs can provide accurate and efficient 006.
identification of landslide susceptible areas. For future studies it is CPRM. Companhia de Pesquisa de Recursos Minerais, 2012. Sistema de Cadastro de
suggested the application of other environmental input parameters in Deslizamentos e Inundações. http://www.cprm.gov.br/scdi/movimentos/.
CPRM. Companhia de Pesquisa de Recursos Minerais, 2015. Carta de suscetibilidade a
order to identify other variables that exert influence on the identifica- movimentos gravitacionais de massa e inundações. Município de Porto Alegre - RS.
tion of susceptible areas. http://rigeo.cprm.gov.br/xmlui/bitstream/handle/doc/15106/CS-PortoAlegre.pdf?
A recommendation to public organs and managers, particularly to sequence=1.
CPRM. Companhia de Pesquisa de Recursos Minerais, 2017. GeoSGB. Unidades
municipal civil defense organs from developing countries, is to create
litoestratigráficas. http://geowebapp.cprm.gov.br/ViewerWEB/.
and maintain online landslide inventories, which would facilitate the Cubito, A., Ferrara, V., Pappalardo, G., 2005. Landslide hazard in the Nebrodi Mountains
employment of available methodologies to identify susceptible areas, as (Northeastern Sicily). Geomorphology 66 (1-4 SPEC. ISS.), 359–372.
De Aguiar, A.A., Hackbart, E., 2008. Primeira metade do século XX teve anos muito secos.
well as increase the reliability of existing results and maps. Finally,
de Amorim, M.F., Quelhas, O.L.G., da Motta, A.L.T.S., 2014. A resiliência das cidades
through the integration of the tool developed in this research and frente a chuvas torrenciais: estudo de caso do plano de contingência dacidade do Rio
landslide inventories, it becomes viable to develop real-time monitoring de Janeiro. Soc. Nat. 26 (3), 519–534. http://www.scielo.br/scielo.php?script=sci_
and warning systems that can help enhance the agility, efficiency end arttext&pid=S1982-45132014000300519&lng=pt&tlng=pt.
de Brito, M.M., 2014. Geoprocessamento aplicado ao mapeamento da suscetibilidade a
effectiveness of response measures.

14
L. Bragagnolo, et al. Catena 184 (2020) 104240

escorregamentos no município de Porto Alegre, RS (Ph.D. thesis). Universidade rainfall-induced landslides using improved self-organizing linear output map, support
Federal do Rio Grande do Sul. http://www.lume.ufrgs.br/handle/10183/101208. vector machine, and logistic regression. Eng. Geol. 224, 62–74. https://doi.org/10.
Dewan, A., Deaton, J., Castillo, M., 2017. More than 200 Dead, Many Missing in Colombia 1016/j.enggeo.2017.05.009. April.
Mudslides. http://edition.cnn.com/2017/04/02/americas/colombia-mudslide/. Marjanović, M., Kovačević, M., Bajat, B., Voženílek, V., 2011. Landslide susceptibility
Ding, Q., Chen, W., Hong, H., 2017. Application of frequency ratio, weights of evidence assessment using SVM machine learning algorithm. Eng. Geol. 123 (3), 225–234.
and evidential belief function models in landslide susceptibility mapping. Geocarto Martinović, K., Gavin, K., Reale, C., 2016. Development of a landslide susceptibility as-
Int. 32 (6), 619–639. sessment for a rail network. Eng. Geol. 215, 1–9.
Ercanoglu, M., Gokceoglu, C., 2002. Assessment of landslide susceptibility for a landslide- Meena, S., Ghorbanzadeh, O., Blaschke, T., 2019. A comparative study of statistics-based
prone area (north of Yenice, NW Turkey) by fuzzy approach. Environ. Geol. 41 (6), landslide susceptibility models: a case study of the region affected by the Gorkha
720–730. Earthquake in Nepal. ISPRS Int. J. Geo Inf. 8 (2), 94.
Ering, P., Babu, G.L., 2016. Probabilistic back analysis of rainfall induced landslide - a Melchiorre, C., Matteucci, M., Azzoni, A., Zanchi, A., 2008. Artificial neural networks and
case study of Malin landslide, India. Eng. Geol. 208, 154–164. https://doi.org/10. cluster analysis in landslide susceptibility zonation. Geomorphology 94, 379–400.
1016/j.enggeo.2016.05.002. Mondini, A.C., tsung Chang, K., 2014. Combining spectral and geoenvironmental in-
Ermini, L., Catani, F., Casagli, N., 2005. Artificial Neural Networks applied to landslide formation for probabilistic event landslide mapping. Geomorphology (213),
susceptibility assessment. Geomorphology (66), 327–343. 183–189. https://doi.org/10.1016/j.geomorph.2014.01.007.
Goetz, J.N., Guthrie, R.H., Brenning, A., 2011. Integrating physical and empirical land- Neaupane, K.M., Achet, S.H., 2004. Use of backpropagation neural network for landslide
slide susceptibility models using generalized additive models. Geomorphology 129 monitoring: a case study in the higher Himalaya. Eng. Geol. 74 (3-4), 213–226.
(3-4), 376–386. https://doi.org/10.1016/j.geomorph.2011.03.001. Nefeslioglu, H.A., Gokceoglu, C., Sonmez, H., 2008. An assessment on the use of logistic
Gómez, H., Kavzoglu, T., 2005. Assessment of shallow landslide susceptibility using ar- regression and artificial neural networks with different sampling strategies for the
tificial neural networks in Jabonosa River Basin, Venezuela. Eng. Geol. 78 (1-2), preparation of landslide susceptibility maps. Eng. Geol. 97 (3-4), 171–191.
11–27. Netch-Nielsen, R., 1987. Kolmogorov's mapping neural network existence theorem. In:
Gorsevski, P.V., Gessler, P.E., Foltz, R.B., Elliot, W.J., 2006. Spatial prediction of landslide First IEEE International Joint Conference on Neural Networks, pp. 11–14.
hazard using logistic regression and ROC analysis. Trans. GIS 10 (3), 395–415. Ohlmacher, G.C., 2007. Plan curvature and landslide probability in regions dominated by
GRASS Development Team, 2017. Geographic Resources Analysis Support System earth flows and earth slides. Eng. Geol. 91 (2-4), 117–134.
(GRASS GIS) Software, Version 7.2. Open Source Geospatial Foundation. http:// Openshaw, S., Openshaw, C., 1997. Artificial Intelligence in Geography. Wiley,
grass.osgeo.org. Chichester.
Guzzetti, F., Reichenbach, P., Cardinali, M., Galli, M., Ardizzone, F., 2005. Probabilistic Osborne, S., 2017. China Landslide: One Dead and 37 Missing After Rockfall in South of
landslide hazard assessment at the basin scale. Geomorphology 72, 272–299. Country. https://ind.pn/2JDrAyM.
Hasenack, H., Weber, E., Marcuzzo, S., 2008. Diagnóstico Ambiental de Porto Alegre: Ozdemir, A., 2011. Landslide susceptibility mapping using Bayesian approach in the
Geologia, Solos, Drenagem, Vegetação e Ocupação. Secretaria Municipal do Meio Sultan Mountains (Akşehir, Turkey). Nat. Hazards 59 (3), 1573–1607.
Ambiente, Porto Alegre 2008. Ozdemir, A., Altural, T., 2013. A comparative study of frequency ratio, weights of evi-
Haykin, S., 1999. Neural Networks: A Comprehensive Foundation, 2nd edition. Prentice- dence and logistic regression methods for landslide susceptibility mapping: Sultan
Hall, Inc., New Jersey. mountains, SW Turkey. J. Asian Earth Sci. 64, 180–197. https://doi.org/10.1016/j.
Hong, H., Pradhan, B., Sameen, M.I., Kalantar, B., Zhu, A., Chen, W., 2018. Improving the jseaes.2012.12.014.
accuracy of landslide susceptibility model using a novel region-partitioning ap- Pachauri, A., Pant, M., 1992. Landslide hazard mapping based on geological attributes.
proach. Landslides 15 (4), 753–772. Eng. Geol. 32 (1-2), 81–100. http://www.sciencedirect.com/science/article/pii/
Hong, H., Pradhan, B., Xu, C., Tien Bui, D., 2015. Spatial prediction of landslide hazard at 001379529290020Y.
the Yihuang area (China) using two-class kernel logistic regression, alternating de- Pandey, V.K., Sharma, M.C., 2017. Probabilistic landslide susceptibility mapping along
cision tree and support vector machines. Catena 133, 266–281. https://doi.org/10. Tipri to Ghuttu highway corridor, Garhwal Himalaya (India). Remote Sens. Appl.:
1016/j.catena.2015.05.019. October. Soc. Environ. 8, 1–11.
Hosmer, D.W., Lemeshow, S., 2000. Applied Logistic Regression, 2nd edition. John Wiley Park, H.J., Lee, J.H., Woo, I., 2013. Assessment of rainfall-induced shallow landslide
& Sons, New York, USA. susceptibility using a GIS-based probabilistic approach. Eng. Geol. 161, 1–15.
Huang, J., Gao, J., 2017. An ensemble simulation approach for artificial neural network: https://doi.org/10.1016/j.enggeo.2013.04.011.
an example from chlorophyll a simulation in Lake Poyang, China. Eco. Inform. 37, Pascale, S., Parisi, S., Mancini, A., Schiattarella, M., Conforti, M., Sole, A., Murgante, B.,
52–58. https://doi.org/10.1016/j.ecoinf.2016.11.012. Sdao, F., 2013. Landslide susceptibility mapping using artificial neural network in the
Huang, Y., Zhao, L., 2018. Review on landslide susceptibility mapping using support urban area of Senise and San Costantino Albanese (Basilicata, Southern Italy). In:
vector machines. Catena 165, 520–529. https://doi.org/10.1016/j.catena.2018.03. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial
003. January. Intelligence and Lecture Notes in Bioinformatics) 7974 LNCS PART 4, pp. 473–488.
IBGE. Insituto Brasileiro de Geografia e Estatística, 2017a. Porto Alegre. https://cidades. Pham, B.T., Bui, D.T., Prakash, I., Dholakia, M.B., 2017. Hybrid integration of multilayer
ibge.gov.br/brasil/rs/porto-alegre/panorama. perceptron neural networks and machine learning ensembles for landslide suscept-
IBGE. Insituto Brasileiro de Geografia e Estatística, 2017b. Rio de Janeiro. https:// ibility assessment at Himalayan area (India) using GIS. Catena 149, 52–63. https://
cidades.ibge.gov.br/brasil/rj/rio-de-janeiro/panorama. doi.org/10.1016/j.catena.2016.09.007.
Ilanloo, M., 2011. A comparative study of fuzzy logic approach for landslide susceptibility Pisano, L., Zumpano, V., Malek, Rosskopf, C.M., Parise, M., 2017. Variations in the sus-
mapping using GIS: an experience of Karaj dam basin in Iran. Procedia. Soc. Behav. ceptibility to landslides, as a consequence of land cover changes: a look to the past,
Sci. 19, 668–676. https://doi.org/10.1016/j.sbspro.2011.05.184. and another towards the future. Sci. Total Environ. 601-602, 1147–1159. https://doi.
Ilia, I., Tsangaratos, P., 2016. Applying weight of evidence method and sensitivity ana- org/10.1016/j.scitotenv.2017.05.231.
lysis to produce a landslide susceptibility map. Landslides 13 (2), 379–397. https:// Porto Alegre, 2011. DECRETO N° 16.931, de 26 janeiro de 2011. Institui o Programa de
doi.org/10.1007/s10346-015-0576-3. Fiscalização e Monitoramento Urbano Ambiental do Município de Porto alegre e
INPE. Instituto de Pesquisas Espaciais, 2011. TOPODATA. Banco de dados revoga o Decreto n° 15.243, de 4 de julho de 2006. http://www2.portoalegre.rs.gov.
geomorfométricos do Brasil. http://www.dsr.inpe.br/topodata/acesso.php. br/cgi-bin/nph-brs?s1=000031523.DOCN.&l=20&u=%2Fnetahtml%2Fsirel
Kanungo, D.P., Arora, M.K., Sarkar, S., Gupta, R.P., 2006. A comparative study of con- %2Fsimples.html&p=1&r=1&f=G&d=atos&SECT1=TEXT.
ventional, ANN black box, fuzzy and combined neural and fuzzy weighting proce- Poudyal, C.P., 2013. Landslide susceptibility analysis using decision tree method, Phidim,
dures for landslide susceptibility zonation in Darjeeling Himalayas. Eng. Geol. 85 (3- Eastern Nepal. Bull. Dep. Geol. 15 (0), 69–76.
4), 347–366. Pourghasemi, H., Yousefi, S., Kornejady, A., Cerdà, A., 2017. Performance assessment of
Kawabata, D., Bandibas, J., 2009. Geomorphology Landslide susceptibility mapping using individual and ensemble data-mining techniques for gully erosion modeling. Sci.
geological data, a DEM from ASTER images and an Artificial Neural Network (ANN). Total Environ. 609, 764–775.
Geomorphology 113 (1-2), 97–109. https://doi.org/10.1016/j.geomorph.2009.06. Pourghasemi, H.R., Mohammady, M., Pradhan, B., 2012. Landslide susceptibility map-
006. ping using index of entropy and conditional probability models in GIS: Safarood
Koch, W., Meng, Y., Shah, M., Chang, W., Yu, X., 2013. Predicting nerve guidance conduit Basin, Iran. Catena 97, 71–84. https://doi.org/10.1016/j.catena.2012.05.005.
performance for peripheral nerve regeneration using bootstrap aggregated neural Pourghasemi, H.R., Pradhan, B., Gokceoglu, C., Mohammadi, M., Moradi, H.R., 2013.
networks. Proc. Int. Joint Conf. Neural Netw. Application of weights-of-evidence and certainty factor models and their comparison
Lamb, P.C., 1996. Residual soils. In: Landslides: investigation and mitigation, special in landslide susceptibility mapping at Haraz watershed, Iran. Arab. J. Geosci. 6 (7),
report. National Academic Press, Washington, DC, pp. 507–524. 2351–2365.
Lari, S., Frattini, P., Crosta, G.B., 2014. A probabilistic approach for landslide hazard Pradhan, B., 2013. A comparative study on the predictive ability of the decision tree,
analysis. Eng. Geol. (182), 3–14. https://doi.org/10.1016/j.enggeo.2014.07.015. support vector machine and neuro-fuzzy models in landslide susceptibility mapping
Lee, D.-H., Kang, D.-S., 2016. The application of the artificial neural network ensemble using GIS. Comput. Geosci. 51, 350–365. https://doi.org/10.1016/j.cageo.2012.08.
model for simulating streamflow. Proc. Eng. 154, 1217–1224. https://doi.org/10. 023.
1016/j.proeng.2016.07.434. Pradhan, B., Lee, S., 2010. Landslide susceptibility assessment and factor effect analysis:
Lee, S., Ryu, J.-H., Won, J.-S., Park, H.-J., 2004. Determination and application of the backpropagation artificial neural networks and their comparison with frequency ratio
weights for landslide susceptibility mapping using an artificial neural network. Eng. and bivariate logistic regression modelling. Environ. Model. Softw. 25 (6), 747–759.
Geol. 71 (3-4), 289–302. http://www.sciencedirect.com/science/article/pii/ https://doi.org/10.1016/j.envsoft.2009.10.016.
S001379520300142X. Pradhan, B., Saro, L., 2007. Utilization of optical remote sensing data and GIS tools for
Leonardi, G., Palamara, R., Cirianni, F., 2016. Landslide susceptibility mapping using a regional landslide hazard analysis using an artificial neural network model. Earth Sci.
fuzzy approach. Proc. Eng. 161, 380–387. https://doi.org/10.1016/j.proeng.2016. Front. 14 (6), 143–151.
08.578. Prefeitura do Rio de Janeiro, 2013a. Geo Rio. Mapa de susceptibilidade ao escorrega-
Lin, G.F., Chang, M.J., Huang, Y.C., Ho, J.Y., 2017. Assessment of susceptibility to mento - Rio de Janeiro. http://pcrj.maps.arcgis.com/apps/webappviewer/index.

15
L. Bragagnolo, et al. Catena 184 (2020) 104240

html?id=7606de2873ee441fae431368967ef56b. artificial neural network classifier. Nat. Hazards 74 (3), 1489–1516.

Prefeitura do Rio de Janeiro, 2013b. Rio de janeiro em busca da resiliência frente chuvas Tsangaratos, P., Ilia, I., 2016. Comparison of a logistic regression and Naive Bayes clas-
fortes. sifier in landslide susceptibility assessments: the influence of models complexity and
Ray, R.L., Jacobs, J.M., 2007. Relationships among remotely sensed soil moisture, pre- training dataset size. Catena 145, 164–179. https://doi.org/10.1016/j.catena.2016.
cipitation and landslide events. Nat. Hazards 43 (2), 211–222. 06.004.
Rio de Janeiro, 2016. Cobertura vegetal e uso da terra 2016. http://www.data.rio/ Tu, J.V., 1996. Advantages and disadvantages of using artificial neural networks versus
datasets/a22768dc645043f5bdfd1affab888ded_0?selectedAttribute=CLASSE. logistic regression for predicting medical outcomes. J. Clin. Epidemiol. 49 (11),
Robbins, J.C., 2016. A probabilistic approach for assessing landslide-triggering event 1225–1231.
rainfall in Papua New Guinea, using TRMM satellite precipitation estimates. J. Van Westen, C.J., Rengers, N., Soeters, R., 2003. Use of geomorphological information in
Hydrol. 541, 296–309. https://doi.org/10.1016/j.jhydrol.2016.06.052. indirect landslide susceptibility assessment. Nat. Hazards (30), 399–419.
Saha, A.K., Gupta, R.P., Arora, M.K., 2002. GIS-based landslide hazard zonation in the Varnes, D.J., 1984. Landslide Hazard Zonation: A Review of Principles and Practice.
Bhagirathi (Ganga) Valley, Himalayas. Int. J. Remote Sens. 23 (2), 357–369. Wang, L.-J., Sawada, K., Moriguchi, S., 2013. Landslide susceptibility analysis with lo-
Saito, H., Nakayama, D., Matsuyama, H., 2009. Comparison of landslide susceptibility gistic regression model based on FCM sampling strategy. Comput. Geosci. 57, 81–92.
based on a decision-tree model and actual landslide occurrence: the Akaishi https://doi.org/10.1016/j.cageo.2013.04.006.
Mountains, Japan. Geomorphology 109 (3-4), 108–121. https://doi.org/10.1016/j. Yeon, Y.-K., Han, J.-G., Ryu, K.H., 2010. Landslide susceptibility mapping in Injae, Korea,
geomorph.2009.02.026. using a decision tree. Eng. Geol. 116 (3-4), 274–283. https://doi.org/10.1016/j.
Santin, F.M., da Silva, R.V., Grzybowski, J.M., 2016. Artificial neural network ensembles enggeo.2010.09.009.
and the design of performance-oriented riparian buffer strips for the filtering of ni- Yesilnacar, E., Topal, T., 2005. Landslide susceptibility mapping: a comparison of logistic
trogen in agricultural catchments. Ecol. Eng. 94, 493–502. https://doi.org/10.1016/ regression and neural networks methods in a medium scale study, Hendek region
j.ecoleng.2016.06.008. (Turkey). Eng. Geol. 79 (3-4), 251–266.
Thiery, Y., Malet, J.-p., Sterlacchini, S., Puissant, A., Maquaire, O., 2007. Landslide sus- Zare, M., Pourghasemi, H.R., Vafakhah, M., Pradhan, B., 2013. Landslide susceptibility
ceptibility assessment by bivariate methods at large scales: application to a complex mapping at Vaz Watershed (Iran) using an artificial neural network model: a com-
mountainous environment. Geomorphology 92, 38–59. parison between multilayer perceptron (MLP) and radial basic function (RBF) algo-
Thiery, Y., Maquaire, O., Fressard, M., 2013. Application of expert rules in indirect ap- rithms. Arab. J. Geosci. 6 (8), 2873–2888.
proaches for landslide susceptibility assessment. Landslides 11 (3), 411–424. Zêzere, J.L., Garcia, R.A.C., Oliveira, S.C., Reis, E., 2008. Probabilistic landslide risk
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012. Spatial prediction of analysis considering direct costs in the area north of Lisbon (Portugal).
landslide hazards in Hoa Binh province (Vietnam): a comparative assessment of the Geomorphology 94 (3-4), 467–495.
efficacy of evidential belief functions and fuzzy logic models. Catena 96, 28–40. Zêzere, J.L., Pereira, S., Melo, R., Oliveira, S.C., Garcia, R.A.C., 2017. Science of the total
https://doi.org/10.1016/j.catena.2012.04.001. environment mapping landslide susceptibility using data-driven methods. Sci. Total
Tien Bui, D., Tuan, T.A., Hoang, N.-D., Thanh, N.Q., Nguyen, D.B., Van Liem, N., Pradhan, Environ. 589, 250–267. https://doi.org/10.1016/j.scitotenv.2017.02.188.
B., 2017. Spatial prediction of rainfall-induced landslides for the Lao Cai area Zhang, F., Chen, W., Liu, G., Liang, S., Kang, C., He, F., 2012. Relationships between
(Vietnam) using a hybrid intelligent approach of least squares support vector ma- landslide types and topographic attributes in a loess catchment, China. J. Mt. Sci. 9
chines inference model and artificial bee colony optimization. Landslides 14 (2), (6), 742–751.
447–458. http://link.springer.com/10.1007/s10346-016-0711-9. Zhang, H., Zimba, P.V., 2017. Analyzing the effects of estuarine freshwater fluxes on fish
Trucco, F., Lopez, E., 2016. Mexico Landslides Kill at Least 40. http://edition.cnn.com/ abundance using artificial neural network ensembles. Ecol. Model. 359, 103–116.
2016/08/07/world/mexico-landslides-deaths. https://doi.org/10.1016/j.ecolmodel.2017.05.010.
Tsangaratos, P., Benardos, A., 2014. Estimating landslide susceptibility through a

16