ACKNOWLEDGEMENT

It is not a fair task to acknowledge all the people who made this [Link] thesis possible with
a few words. However, I will try to do my best to extend my great appreciation to everyone
who helped me scientifically and emotionally throughout this study.

I shall begin with God the almighty: without His will I would have never
found the right path. His mercy was with me throughout my life and ever more in this study. I
thank Him for enlightening my soul with the respected love and compassion for the other
humans and allowing me to enter a field where I could practice this desire.

Then, I would like to express my sincere gratitude to my advisor Dr. Varun Dutt for the
continuous support for my [Link] study and related research, for his patience, motivation,
and immense knowledge. He provided me an opportunity to join his team as intern, and who
gave access to the laboratory and research facilities. Without his precious support it would
not be possible to conduct this research. His guidance helped me in all the time of research
and writing of this thesis. I could not have imagined having a better advisor and mentor for
my study. I also remain indebted for his understanding and support during the times when I
was really down and depressed due to personal family problems.

My sincere gratitude is reserved for Dr. Venkata Uday Kala and their invaluable insights
and suggestions. We had lots of fun with uday sir.

I am sincerely thankful to Pratik Chaturvedi sir from DRDO, who has helped me during my
internship. How can I forgot you guys Praveen kumar (“neta giri nahi”), Priyanka sihag
(“aaj sir gusse me to nahi?”), Ankush pathania(“Yooo man”) , for making my life easy in
IIT Mandi campus, a big “Thank” to you guys.

Words cannot express the feelings I have for my parents for their constant unconditional
support, patience, understanding, and encouragement in every possible way. Their endless
love priceless, perpetual, indispensable help, support and everything made all this possible.

SHUBHAM SHROTI

1
2
 ABSTRACT

Landslides plague the Himalayan region, and landslide occurrence is

widespread in hilly areas. A landslide, i.e., collapse of a mass of earth or rock
from a mountain or cliff, is a common phenomenon in hills. Due to landslides,
average losses in the Himalayas cost more than INR 550 Crores per year and
more than 200 deaths per year. Thus, forecasting these events locally will
improve the preparedness of society in facing landslide disasters. There is an
urgent need to develop Landslide Monitoring and Warning / Prediction System,
which help in monitoring and predicting landslide activity in a specific area of
interest; and, alerting people when landslide risks exceed predefined thresholds.
In India, several research organizations are working on using wired and wireless
sensors for monitoring various parameters related to landslides. However, the
cost of these monitoring systems is very high, predictions are not available at
daily-scale and hourly-scale, and their predictions are limited to the parameters
sensed at only a single site. These limitations restrict the large-scale deployment
of landslide monitoring systems over a large geographical area. This project
enumerates the idea of using low-cost micro-electro-mechanical systems
(MEMS) - based sensors for local site-specific monitoring of landslides for
every 10-minute interval. A prototype of the low-cost system is currently
installed on a hill in Manali road (Hanogi, Pandoh, Thalot, Deode), and
Jogindernagar road (Kotropi, Gumma, Narla)- Himachal Pradesh. Initially, the
historical rainfall records, soil properties and geology information of the
selected site were analysed to understand the behaviour of the landslide area.
Then the site-specific low-cost MEMS sensors were installed to collect data on
rainfall, temperature, pressure, relative humidity, light intensity, soil moisture,
and soil movement at the selected site. Local site-specific data is stored in the

3
internet-based cloud database where historical rainfall has been identified as a
significant causal factor for debris movements and associated landslides. Thus,
rainfall could be used for predicting debris movements and associated landslide
events before their actual occurrence. A recent approach to predicting debris
movements is to use machine-learning techniques. In machine-learning
literature, both conventional and deep methods have been proposed. The
random forest algorithm, which is a conventional machine-learning technique
was compared to a deep machine-learning technique, long short-term memory
(LSTM) in their ability to predict the intensity of daily debris movements at the
Tangni area in Chamoli, Uttarakhand. Furthermore, the system is capable of
sending SMS alerts to registered mobile users in case there is irregular land
movement or if the probability of landslide exceeds predefined thresholds. We
have compared and evaluated a number of machine-learning algorithms that
improved the accuracy of the system. This chapter proposes a MEMS system-
based IoT framework for sensing landslides and predict the rainfall at daily
scale.

4
 TABLE OF CONTENT

ACKNOWLEDGMENT i
ABSTRACT iii
TABLE OF CONTENTS v
LIST OF TABLES vii
LIST OF FIGURE ix
1. INTORDUCTION 1
1.1 Introduction………………………………………………………… 1
1.2 System architecture………………………………………………… 11
2. IOT Gateway Protocol and Data Acquisition Tools 14
2.1 IOT Gateway Protocol……………………………………………... 14
2.1.1 Hypertext transfer protocol………………………………… 14
2.1.2 MQTT Protocol…………………………………………….. 16
2.2 Data Acquisition Tool……………………………………………… 19
2.2.1 Thingspeak IOT……………………………………………. 19
2.2.2 Weka Tool………………………………………………….. 21
3. Data Mining 20
3.1 Data Mining………………………………………………………… 20
3.1.1 Predictive Approach………………………………………... 22
3.1.2 Descriptive Approach………………………………………. 23
3.2 Data Mining Steps………………………………………………….. 23
3.3 Data Pre-processing………………………………………………… 24
3.4 Machine Learning Algorithm………………………………………. 25
3.4.1 Naïve base…………………………………………………... 27
3.4.2 Random Forest……………………………………………… 29
4. Predictive Model Design with LSTM 32
4.1 The Problem of the Long Short Term Dependencies……………….. 33
4.2 LSTM Network……………………………………………………… 35
4.2.1 The core idea behind LSTM………………………………… 37
4.3 Evaluation…………………………………………………………… 38
4.3.1 Mean error…………………………………………………… 39
4.3.2 Mean absolute error…………………………………………. 39
4.3.3 Root mean square error……………………………………… 40
4.3.4 Evaluation schemes………………………………………….. 40
5. Implementation Results, Conclusion and Future Scope 42
5.1 Data Collected……………………………………………………….. 42
5.2 Apply normalisation Results………………………………………… 44
5.3 Split Data for Training………………………………………………. 45
5.4 LSTM Design Model Accuracy……………………………………... 45
5.5 LSTM model Accuracy at the Different Parameter…………………. 46
Conclusion 49
Future Scope 51

5
References 52

6
 LIST OF FIGURES

Figure Name of Figure Page

Number Numbe
r
Figure 1 Satellite techniques for landslide hazard assessment 4
(a): Digital Elevation Models (DEM) and (b): Synthetic-
Aperture Radar (SAR)
Figure 1a Close Range Digital Photogrammetry Techniques Applied 4
for Landslides Monitoring
Figure 1b Conventional ground-based geodetic techniques for 5
deformation monitoring
(a) Traditional transit and (b) Modern total station
Figure 1c Global Positioning System for geotechnical evaluations 6
(a) Fast static GPS techniques and (b) Real time Kinematics
GPS techniques
Figure 1d inclinometer 7

Figure 1e extensometer 7

Figure 1f piezometer 8

Figure 1g Geophones 9

Figure 1h Tiltmeters 12

Figure 2.1 HTTP establishing TCP connection between Client and 16

Server (WIKIBooks, 2015))
Figure 2.2 MQTT data transmission architecture with a broker 18

Figure 2.3. interface of thingspeak channels 21

Figure Structure of Naive Bayes 28

3.4.1
Figure 4.1 Recurrent Neural Networks have loops 32

Figure 4.1 An unrolled recurrent neural network 33

Figure 4.3 The repeating module in a standard RNN contains a single 36

layer
Figure 4.4. The repeating module in an LSTM contains four interacting 36
layers

Figure 5.2 Normalization Function 45

7
Figure 5.2a is Series to supervised learning for chamoli data. 45

Figure 5.3 Split Data for Training 46

Figure 5.4 LSTM Model Design Accuracy 46

Figure 5.5a graph between actual and predicted values 47

Figure 5.5 LSTM model accuracy on tagni dataset 48

b
Figure 5.5c graph between actual and predicted values 48

8
 LIST OF TABLES
Table Name of Table Page
Number Numbe
r
Table 2 A summary of IoT Application Layer protocols 14

Table sample of csv file of chamoli rainfall data 42

5.1.

Table sample file of chamoli rainfall data 43

5.1a

9
 CHAPTER 1

Introduction

1.1 Introduction

Landslides are uncertain geological events, and they pose great

dangers to life and infrastructure (Parkash, 2011). It causes a lot of
damages to life and property, block roads, and disrupt the
transportation of goods and services. For places at very high altitudes,
where everything from food to clothing is imported from cities,
blocking roads is a very big issue. These reasons become very
supporting evidence for combating the problem at the earliest. In
India, especially in the Himalayan region, landslides are more frequent
than any other geological phenomena causing more than 200 deaths
and, on average, INR 550 Crores in damages to infrastructure yearly
(Chaturvedi, Arora, & Dutt, 2017; Chaturvedi & Dutt, 2015). Because
of large costs and many deaths due to landslides, there is a need to
design and develop frameworks that monitor landslides and alert
people before they occur. To be effective, those frameworks should
possess the following features: sense soil properties and soil
movement at landslide-prone sites; log sensed data at a remote site
via a cloud infrastructure; allow analyses of logged data; and, alert
people via mobile applications before landslides occur.

1
Several technologies have been proposed for monitoring landslides
(Guzzetti et al., 2012; Niethammer et al., 2012; McKean & Roering,
2004; Zhang et al., 2004). Monitoring of deformation of structures and
ground surface displacements during landslides can be accomplished
by using different types of systems and techniques. These techniques
and instrumentation can be classified as:

 Remote sensing or satellite techniques with space-derived

information.
 Photogrammetric techniques
 Ground-based conventional surveying techniques
 GPS techniques
 Geotechnical techniques

1. Remote sensing or satellite techniques

Remote sensing in the form of the photographic, scanning, and
processing system have significant potential for landslide hazard
assessment and improved understanding of landslide processes.
Satellite images in the optical region with high spatial resolution are
used for producing landslide inventory maps and for mapping factors
related to the occurrence of landslides such as surface morphology,
structural and lithological properties, land cover, and temporal
changes of these factors.

Synthetic-Aperture Radar (SAR) is a form of radar that is used to

create two- or three- dimensional images of objects, such as

2
landscapes (see Figure 1a). SAR uses the motion of the radar antenna
over a target region to provide finer spatial resolution than
conventional beam-scanning radars.

The SAR works similar to a phased array. However, contrary to a

large number of the parallel antenna elements of a phased array, SAR
uses one antenna in time multiplexing. The different geometric
positions of the antenna elements are result of the moving platform.
The SAR-processor stores all the radar returned signals, as amplitudes
and phases, for the time period T from position A to D. Now it is
possible to reconstruct the signal which would have been obtained by
an antenna of length v · T, where v is the platform speed. As the line
of sight direction changes along the radar platform trajectory, a
synthetic aperture is produced by signal processing that has the effect
of lengthening the antenna. Making T larger makes the synthetic
aperture large and hence a higher resolution can be achieved.

Digital Elevation Models (DEM): Another important application

of Earth observation data is the generation of Digital Elevation Models
(DEMs), which provide basic information for the characterization of
slopes surfaces (see Figure 1b). The DEMs provide a specialized
database that represents the relief of a surface between points of
known elevation. By interpolating known elevation data from sources
such as ground surveys and photogrammetric data capture, a
rectangular digital elevation model grid can be created. GIS software

3
can use digital elevation models for 3D surface visualization,
generating contours, and performing viewshed visibility analysis.

Fig 1: Satellite techniques for landslide hazard assessment

(a): Digital Elevation Models (DEM) and (b): Synthetic-Aperture Radar (SAR)

2. Photogrammetric techniques
Photogrammetry is the science of making measurements from
photographs. The output of photogrammetry is typically a map,
drawing, measurement, or a 3D model of some real-world object or
scene.

Figure 1a: Close Range Digital Photogrammetry Techniques Applied for Landslides Monitoring

3. Ground-based conventional surveying techniques

Conventional ground-based geodetic techniques have been used for
deformation monitoring of structures and landslides.

Two basic methods for the design of a deformation survey can be

used:

4
 A horizontal or vertical control network is established in the area
under investigation with control points located in the deforming
region.
 Total station instruments are used to measure angles and distances
to target-prisms located on the moving mass.

In both cases, the aim is the computation of target point coordinates

and/or heights for each measurement period. From the comparison
of these coordinates/heights, after all, proper statistical and reliability
tests, the horizontal and/or vertical displacement vectors of each
control point can be determined.

Figure 1b: Conventional ground-based geodetic techniques for deformation monitoring

(a) Traditional transit and (b) Modern total station

4. GPS techniques
The Global Positioning System (GPS) can be used as an alternative
surveying tool to assist in geotechnical evaluations of steep slopes by
providing 3D coordinate time series of displacements at discrete
points on the sliding surface. GPS positioning is based on measuring
the transit time of radio signals emitted by orbiting satellites. GPS
offers advantages over conventional terrestrial methods.

5
 GPS provides specially coded satellite signals that can be
processed with a GPS receiver, enabling the receiver to compute
position, velocity and time. A minimum of four GPS satellite signals is
required to compute positions in three dimensions and the time offset
in the receiver clock. Accuracy and precision of data increase with
more satellites. A GPS receiver uses trilateration (a more complex
version of triangulation) to determine its position on the surface of the
earth by timing signals from three satellites in the Global Positioning
System.

Figure 1c: Global Positioning System for geotechnical evaluations

(a) Fast static GPS techniques and (b) Real time Kinematics GPS techniques

5. Geotechnical techniques
Geotechnical sensors are used extensively in the monitoring of
structures. These sensors are often placed within the structure and
out of sight. The main geotechnical sensors used for deformation
monitoring include; extensometers, inclinometers, piezometers,
strain meters, pressure cells, geophones, tiltmeters, crack meters, etc.
Geotechnical sensors provide measurements that are often essential

6
in deformation monitoring. The principal techniques/instrumentation
will be briefly discussed below:

5.1 Inclinometers This instrument

installed in boreholes drilled within the
landslide mass. They measure the
curvature of initially straight boreholes,
thus detecting any change in inclination
of the borehole casing.

Fig 1d: Inclinometer

5.2 Extensometers
Measure the axial displacement between
a number of reference points in the same
measurement’s axis. They can be
installed within a borehole or on the
slope surface. The wire extensometer is
Fig 1e: Extensometers widely used typically measuring
baselines of up to 80 m in length with
an accuracy of ± 0.3 mm per 30 m. The actual accuracy depends on the
temperature corrections and the quality of the installation of the
extensometer. Maintaining a constant tension throughout the use of the
wire extensometer is very important.

5.3 Piezometers

7
 Many landslides are triggered by slope
saturation following heavy rainfall.
Measurement of pore water pressures
and piezometric levels form an
Fig 1f: Piezometers important part of slope stability
analysis. Threshold levels can be defined to provide early warning of
conditions that may lead to catastrophic failure.

5.5 Geophones
They are devices that can measure vibration associated with the
movement. They can detect landslides on the basis of frequency
composition, amplitude, and duration of the vibration signal.

Fig 1g: Geophones

5.4 Tiltmeters
They are instruments that can measure degrees of rotation. They are
based on electrolytic level sensors. They can be more sensitive than
and are useful on extremely slow-moving rotational failures.
Tiltmeters are used to determine the direction of movement, to
delimit the areas of deformation and to determine the mechanism of
movement (e.g., slumping or slope creep). They can also provide
advance warning of accelerated slide movement and quantify the
effectiveness of landslide repairs.

8
 Fig 1h: Tiltmeters

Although these techniques provide effective landslide monitoring

technologies, they only scan the terrain’s surface, and they are
expensive to use. For example, in the city of Portland, where DEMs
were used for landslide monitoring, the mapping cost was between
USD 400 and USD 600 per square mile (Weinstein, 2010). This high
cost for landslide monitoring has been a major reason why the city
was unable to afford a large-scale DEM implementation. Similarly,
high costs and operating constraints (McKinnon, 2015) associated
with UAVs limited their usage to a wider area for landslide monitoring.
These limitations restrict the large-scale deployment of such kind of
landslide monitoring systems.

IoT frameworks provide a foundation for connecting sensors,

actuators, and other smart technologies and thus they help to
improve automation and control in various operations (Borgia, 2014).
These IoT frameworks could provide alternate solutions for
monitoring landslides. In the past few years, researchers have
designed and developed certain IoT frameworks for monitoring
landslides and alerting people (Arnhardt & Neussner, 2013; Aziz &
Aziz, 2011; Ramesh, 2009). These IoT frameworks involve data
collection, transmission, pre-processing, machine-learning, decision-

9
making, and information dissemination via mobile applications (Khan
et al., 2012). Thus, these IoT frameworks go beyond the typical
Wireless Sensor Networks (WSNs), which consist of only sensors that
collect data and transmit it wirelessly (Othman & Shazali, 2012).
However, these frameworks use costly sensing components like
geophones, pore-pressure transducers, and tiltmeters (Arnhardt &
Neussner, 2013; Aziz & Aziz, 2011; Ramesh, 2009). The high cost of
sensing landslides via existing IoT frameworks limits their large-scale
deployment across several landslide-prone areas in the world.

One way of reducing cost is by using MicroElectroMechanicalSystems

(MEMS) based sensors (Gupta & Ahmad, 2007) as a part of existing
landslide-sensing IoT frameworks. MEMS is a technology that uses
advances in fabrication techniques to embed an electro-mechanical
system on a single chip (Mehergany & Roy, 1999). In MEMS-based
sensors, mechanical elements like gears and beams are added on top
of electrical systems, where this electromechanical combination
allows one to sense the physical world (Mehergany & Roy, 1999). Over
last four-to-five decades, due to low costs, MEMS-based sensors have
been increasingly used for a wide range of applications in healthcare,
automotive, defense, and communications (PFT, 2002). The potential
of using MEMS-based sensors across these applications has also
drawn the attention of researchers to exploit these sensors for
monitoring natural disasters (Arnhardt & Neussner, 2013; Manconi &

10
Giordan, 2016). However, applications of low-cost MEMS-based
sensors for landslide monitoring is non-existent in India. Therefore, we
had developed a low-cost MEMS-sensors based framework for risk-
assessment and risk-communication against landslides. However,
sensors used for landslide monitoring in the real world need to be first
calibrated and tested at the lab scale. This framework provides
features like data logging, thresholding, and communication for
alerting people living in landslide-prone areas. The system will initially
generate landslide advisory once daily; and later, with data-logging
and Global Climate Models’.

1.2 System Architecture

Based upon a geophysical survey, drill at 2 to 3 points certain distance
apart at the selected site. Each of these drill points have a standalone
sensing system. This sensing system have wired sensors going
subsurface (certain sensors for measuring rain, light intensity,
temperature, relative humidity, and barometric pressure may exist on
the survey). Please note that the transmission of the data from each
drill point will be done wirelessly to the remote location (for data
logging). Thus, our system is not a centralized wired sensor network
where data is logged at a central location on site and then transmitted
to the remote site; rather, it is a decentralized wireless sensing
system, where each drill point is a standalone broadcasting station
sending its data to the remote server. This decentralized architecture

11
helps install sensors at a drill point without any dependency on other
drill points. Also, this decentralized system is fault tolerant, where the
failure of sensors at one drill point do not influence sensors at other
drill points. Figure 6 shows diagram of data flow and from sensed site
to the remote location, where data will be logged at the remote
location to showcase the process of data flow and model building in
the project.

Figure 6: LMS model architecture

(GCMs’) downscaling, provide this advisory at a shorter time interval

of 10 minutes. In addition to this, we presented different machine-
learning techniques to train on the dataset generated by the MEMS-
based sensors for predicting landslides on an hourly scale. Prior
research has emphasized on predicting landslides in a daily-scale.
There has been research in machine learning that aims to predict

12
landslides based on the statistical analysis of historical landslide
events and its triggering factors. However, prediction of landslides
suffers from a class imbalance problem as landslides and land-
movement are very rare events. In this project, we apply state of-the-
art techniques to correct the class imbalance in landslide datasets.
More specifically, to overcome the class-imbalance problem, we use
different synthetic and oversampling techniques to a real-world
landslide data collected from the bamboo block in IIT Mandi campus
as well as Chandigarh – Manali highway. Also, we apply several
machine learning algorithms to the landslide data set for predicting
landslides and evaluating our algorithms. Different algorithms have
been assessed using techniques like the True Positive rate, False
Positive rate, Area under the ROC curve (AUC) and sensitivity index
(d’). Results suggested that Decision tree algorithm performed better
compared to other classification techniques like neural networks,
logistic regression, support vector machines, and random forest. We
highlight the implications of our results and methods for predicting
landslides in the real world. As we had done a noble work to monitor
and predict landslides based on soil-movements with the low cost
MEMS based sensor, which is really cheap. We filled the patent
application to secure this technology.

13
 CHAPTER 2

IOT Gateway Protocol and Data Acquisition Tools

In this chapter, we will discuss the data acquisition tool and some data
transmission protocol for sending the data to the web server and
acquire the data from the database, applying the same tools for mine
the data and use the classifier.

2.1 IOT Gateway Protocol

To start with, like in any other form of communication between
Human-to-Human (H2H) or D2D (Device-to-Device) there should be a
protocol that promotes or helps individuals or devices to understand
each other. In the case of the IoT communication between D2D or
(Machine-to-Machine) M2M and the cloud, there is a broad set of
protocols that facilitate communications.

Transport WAN Compute

Protocol Protocol Messaging (2G, 3G, Power Resources Security
4G)
HTTP/ REST TCP Rqst/Rspnse Excellent Fair 100Ks/RAM Low-
Flash Optimal
MQTT TCP Pub/Subsrb Excellent Good 10Ks/RAM Medium-
Rqst/Rspnse Flash Optimal
Table 1: A summary of IoT Application Layer protocols

The following sections of the chapter will be a presentation of the

details of the above summarized IoT protocols.

2.1.1 Hypertext transfer protocol (HTTP):

14
HTTP (Fielding & Reschke, 2014) is the most widely and popularly
adapted Application Layer protocol on the World Wide Web. The
standardization of HTTP has been done by the Internet Engineering
Task Force (IETF) in collaboration with the World Wide Web
Consortium (W3C) (MIT). HTTP works on a Client-Server messaging
technology where the client requests for a Hypertext Markup
Language (HTML) page from a server and the server also responses
with an HTML page. As illustrated in table 1 HTTP relies on the TCP as
a transport protocol, which uses sockets to transfer data. The
connection between the client and server begins with the client via a
socket connection on the port 80, which is the assigned port number
for HTTP to the Server. When the connection is established, it means
that the server accepts the request of the client, which is in an HTML
page form, and other objects.

However, upon the connection establishment, the HTML pages and

the objects are then exchanged between the client browser and the
web server. After the completion of the request, the TCP terminates
the connection between the client and the server and also clears the
memory so that previous requests from the client are removed. With
the HTTP, requests, such as GET, PUT, POST and DELETE, are the four
methods mostly used. The GET request displays a web page and its
objects upon a request to the user. The PUT and POST request

15
methods are used to modify server resources and the DELETE request
removes resources that are not needed.

Moreover, there are two HTTP connection types

that can be established with the TCP. These are the Non-Persistent
(HTTP/1.0) and Persistent (HTTP/1.1) connections. The main
difference between the Non-Persistent (HTTP/1.0) and Persistent
(HTTP/1.1) depends on the number of TCP connections needed to
transmit a Uniform Resource Locator (URL) of a web page and its
objects. Figure 4 shows an HTTP connection scheme between a client
and a server.

Figure 2.1: HTTP establishing TCP connection between Client and Server (WIKIBooks, 2015))

2.1.2 MQTT PROTOCOL

The MQTT (ISO/IEC 20922) (Richard J Coppen, 2016) is a machine-to-

machine (M2M) IoT connectivity transport protocol suitable for Low-
Power and Lossy Networks. MQTT is designed as a Client-to-Server
and it employs a publish/subscribe messaging protocol paradigm. Its
implementation is based on the TCP/IP protocol, which is
characterized as reliable, ordered and error-checked protocol. It is
extremely light weight, open, simple and easy to implement. It is
16
designed to provide connectivity to embedded devices, to enable
communication within constrained environments, i.e. communication
in M2M and IoT devices and applications where a small Code footprint
is required or the Network bandwidth is limited.

MQTT uses the publish/subscribe architecture which consists of the

Publisher (Client), Subscribers, a Broker (Server), Sessions and Topics.
The publish/subscribe paradigm is a communication protocol
between a client and server/subscriber which requires a central
MQTT Broker to manage and route data among MQTT networks
nodes or subscribers. The publishers are lightweight sensors that
connect to a Broker to send data. Subscribers are devices or
applications that are logically attached to a client who is interested in
a sensor data and they are connected to the Broker to be informed
whenever new data is received. The Broker classifies sensor data into
topics and sends them to the subscribers interested in the topics.
Technically, topics are message queues that support the
publish/subscribe pattern for clients and logically, topics allow clients
to exchange information with defined semantics.

Finally, a session identifies an attachment of a client to a server. All

communication between client and server takes place as part of a
session. Figure 10 illustrates the MQTT data transmission architecture
with a Broker, which serves a data server directing all data to their
appropriate destinations.

17
 Figure 2.2 MQTT data transmission architecture with a broker

As depicted in figure 2.2 , it can be deduced that the publish/subscribe

message pattern provides a one-to-many messaging and that the
Broker controls the distribution of information between the publisher
client (the source of data) and the subscriber client (the destination
of the data). The Broker stores, forwards, filters and prioritizes
published requests from the publisher client to the subscriber clients.
With the MQTT Broker system, clients can switch between the
publisher and subscriber roles depending on their objectives at a
particular instance. Also, within the Broker there are the MQTT

18
Quality of Service (QoS) levels. The QoS levels are of 0, 1 and 2 that
describe the increasing levels of the guaranteed message delivery.

2.2 Data Acquisition Tools

A. Data Acquisition - ThingSpeak IoT: ThingSpeak is an open

source Internet of Things application and API to store and
retrieve data from things using the HTTP protocol over the
Internet or via a Local Area Network. There are three
fundamental areas ThingSpeak focus on are gathering
information from devices, dissecting and visualizing data with
the help of integrated MATLAB and setting off a response
modified by users. The information is stored online and could be
downloaded for further research. Up until now, ThingSpeak is
the most widely recognized decision for information logging.

a. Benefits of ThingSpeak: According to ThingSpeak, making a

free account will offer access to the platform with up to 3
million messages for every year and the information is logged
every 15 seconds. For higher classes, for example, Student,
Home, Academic and Standard, ThingSpeak is accessible to be
obtained with better features: 33 million messages for every
year and the information is logged each second. ThingSpeak
is appropriate for ventures, institutions, business and even
government exercises.
19
b. Set-up: ThingSpeak stage is clear and straightforward for
clients. A new channel is made to store information and it
incorporates fields for up to 8 unique sensors/reading at the
same time. Each channel has a unique Application
Programming Interface (API) key and Channel ID. These
qualities will be utilized to impart amongst devices and
ThingSpeak. Tweaking the channel by changing the settings:
Name, Description, Field#, Metadata, Tags, Latitude,
Longitude, Elevation, Link, and Video URL if necessary. The
channel could likewise be set to be open or private and each
field could be classified with unit, title and number of reading
appeared. The most astonishing thing about ThingSpeak is its
corporation with MATLAB to make Visualization and Analysis.
These capacities will enable users to access to devices without
owning MATLAB, which is genuinely costly. In the event that
users are more into engineering mathematics and would
prefer toward having just the information, the "Data Export"
will download everything into .CSV format for additionally
further research. Figure 2.3 shows the interface of
ThingSpeak channels

20
 Figure 2.3: interface of thingspeak channels

B. Weka Tool: Weka is a collection of machine learning algorithms

for data mining tasks. The algorithms can either be applied

directly to a dataset or called from your own Java code. Weka
contains tools for data pre-processing, classification, regression,
clustering, association rules, and visualization.

21
 CHAPTER 3

DATA MINING
In this day and age, information is being produced in the form of data
by the most trivial of everyday human actions and interactions. Such
data are more often than not being stored and kept for further
processing. Each of them holds the possibility of containing hidden
knowledge, that may not be explicitly stored inside the data
structures, but can be derived from real world relations and processes.
For extraction of such hidden information, data mining is used. “The
abundance of data, coupled with the need for powerful data analysis
tools, has been described as a data rich but information poor situation.

3.1 Approaches to Data Mining

3.1.1 Predictive Approach
The aim of predictive approach to data mining is to acquire the ability
to predict values of attributes, without those being explicitly observed
prior to evaluation [2]. This is achieved by employing the usage of
machine learning (further referred to as ML) algorithms, that are
explained in detail in chapter 4.
A set of training data is provided to these algorithms. It contains
examples, and labels associated with them. If sufficient number of

22
examples has been provided, the algorithm is expected to acquire the
ability to predict the label.

3.1.2 Descriptive Approach

Descriptive approach to data mining strives to find characteristic
attributes of data, and discover relationships between them. These
are often returned in the form of data sets, that contain attributes
deemed to be the most characteristic given the mined data set, visual
representation of the data set, or association rules that are explained
in further detail in this work.

3.2 Data Mining Steps

Data mining is a process that involves a number of steps:
Data Cleaning
In the first step, data that contain corrupted or empty records are
removed.
Data Integration
In order to proceed with data mining, data need to be collected and
integrated into a single formatted structure. However, different
sources of data usually do not provide uniform structures and
interpretations of data, therefore integration into a single format
needs to take place.
Data Selection

23
For data mining to provide results of high quality, data of high quality
need to be supplied. Not all of the data collected are needed though.
Data selection allows for choosing only such data, that are relevant to
the task to be performed.
Data Transformation
The data that have passed the cleaning step are still not ready for data
mining purposes, for they still need to be transformed into format
accepted by the data mining algorithm. This is achieved via application
of techniques such as smoothing, aggregation or normalisation.
Data Mining
In this step, various algorithms may be applied on the data in order to
discover potential knowledge hidden within the data. Some of the
algorithms applied might be classification, clustering or association
analysis.

Pattern Evaluation
The importance of results provided by data mining needs to be
evaluated, for not all of the findings may be of interest to the inquiry.
Some results may display lower levels of prediction precision, or may
be hard to interpret to a human being. Redundant patterns are
therefore removed.
Knowledge Representation

24
Results that appear to be the most important undergo transformation
and visualisation in order to be presented in the most understandable
form.

3.3 Data Pre-processing

Data pre-processing is a key element to improve the accuracy of data

mining algorithms. Not all data that are collected for the purposes of
data mining are necessarily suited for the task. Wrong selection of
data may heavily influence results of data mining, and therefore
certain treatments need to be imposed on the data before they are
supplied to the algorithms. Data cleaning This method of data pre-
processing deals with removal of imperfect data, such as noise for
continuous attributes. Another of this method’s responsibilities is
replacing the missing data records. Feature selection and feature
reduction Feature selection and feature reduction are data pre-
processing methods that aim to improve the performance of the
learning model by eliminating redundant features and keeping the
important ones. Reduction of the number of features in the model
helps alleviate the curse of dimensionality, which is one of the major
problems in regards to data used for machine learning and data
mining. The increase in dimensionality of an object results in major
increase in its volume, that causes the available data to become
sparse. This sparsity is rather problematic for any method that
requires statistical significance, for the sheer amount of data that

25
needs to be supplied in order for that significance to be achieved
grows exponentially with each dimension of the data. By removal of
irrelevant features, dimensionality is decreased and performance of
the model is further augmented in enhancing its generalisation
capability, speeding up of the learning process, improve of the model's
interpretability and helping the user in acquiring a better
understanding of their data by highlighting the important features and
their relations.

Data Transformation
Some data may need normalization or discretization [8] to be usable
by data mining algorithms. Normalization is a process that is applied
on quantitative attributes, in order to eliminate the effect of having
different scale measures. Discretization is used to transform
quantitative attributes to qualitative ones. Additionally, data
transformation might assign weights to attributes in order to highlight
significant ones.
3.4 Machine Learning Algorithms
Machine Learning is a field that has evolved from the study of Artificial
Intelligence. In simplified terms, the goal of ML is to mimic the
principles of human cognitive processes by machines. Human brain
determines the identity of an object based on a set of it's characteristic
attributes. This principle applies to ML as well, as cluster analysis
works on a similar principle. However, the most important question is,
how to make machines able to learn. In the context of ML, the process
26
of learning can be understood as an inductive inference, where the
machine is presented with examples that contain incomplete
information about some statistical phenomenon. From here on, either
supervised, or unsupervised learning takes place. Supervised learning
usually takes the predictive approach, while unsupervised learning
tends to focus on descriptive approach.

3.4.1 NAÏVE BAYES:

The Naive Bayes classifier is a simple classifier which uses probability
to make predictions. It is mainly built on Bayes theorem. The
presumptions it makes are strongly naive but still it has been proven
to perform quite well in many real-world applications. The classifier is
also referred to as Idiot Bayes, Naive Bayes or Simple Bayes. A more
descriptive term for the underlying probability model would be
independent feature model. In simple terms, a Naive Bayes classifier
assumes that the presence or absence of a particular feature of a class
is unrelated to the presence or absence of any other feature. For
instance, an object may be considered to be a ball if it is round,
bouncy, about 4 inches to 8 inches in diameter. Even if these features
depend on each other or upon the presence of other features, a Naive
Bayes classifier considers all of these properties to independently
contribute to the probability that the object is a ball.
Depending on the precise nature of the probability model,
Naive Bayes classifiers can be trained very efficiently in a supervised

27
learning setting. In many practical applications, parameter estimation
for Naive Bayes models uses the method of maximum likelihood in
other words, one can work with the Naive Bayes model without
believing in Bayesian probability or using any Bayesian methods.
In spite of their naive design and apparently over-simplified
assumptions, Naive Bayes classifiers have worked brilliantly in solving
many complex problems. It is one of the most effective and efficient
learning algorithms for data mining and machine learning.

FIGURE 3.4.1. Structure of Naive Bayes

a. Probabilistic Model for Naïve Bayes:

The discussion so far has derived the independent

feature model, that is, the naive Bayes probability
model. The naive Bayes classifier combines this model
with a decision rule. One common rule is to pick the
hypothesis that is most probable; this is known as MAP
decision rule.

28
3.4.2. Random forest
Random Forest is a flexible, easy to use machine learning
algorithm that produces, even without hyper-parameter
tuning, a great result most of the time. It is also one of the
most used algorithms, because it’s simplicity and the fact
that it can be used for both classification and regression
tasks. In this post, you are going to learn, how the random

29
forest algorithm works and several other important things
about it.
ho

How it works: Random Forest is a supervised learning

algorithm. Like you can already see from its name, it creates
a forest and makes it somehow random. The „forest “it
builds, is an ensemble of Decision Trees, most of the time
trained with the “bagging” method. The general idea of the
bagging method is that a combination of learning models
increases the overall result. To say it in simple words:
Random forest builds multiple decision trees and merges
them together to get a more accurate and stable prediction.

One big advantage of random forest is, that

it can be used for both classification and regression

30
problems, which form the majority of current machine
learning systems. I will talk about random forest in
classification, since classification is sometimes considered
the building block of machine learning. Below you can see
how a random forest would look like with two trees:

31
 CHAPTER 4

Predictive Model Design with Long Short-

Term Memory
Introduction:
In this chapter, I am going to talk about the how to design the neural
network for prediction analysis with Long short-term memory (LSTM)
, LSTM units are units of a recurrent neural network (RNN), Recurrent
neural networks address this issue. They are networks with loops in
them, allowing information to persist.

Figure 4.1: Recurrent Neural Networks have loops

In the above diagram, a chunk of neural network, A, looks at some

input xt and outputs a value ht. A loop allows information to be passed
from one step of the network to the next. These loops make recurrent
neural networks seem kind of mysterious. However, if you think a bit
32
more, it turns out that they aren’t all that different than a normal
neural network. A recurrent neural network can be thought of as
multiple copies of the same network, each passing a message to a
successor. Consider what happens if we unroll the loop:

Figure 4.2: An unrolled recurrent neural network

This chain-like nature reveals that recurrent neural networks are

intimately related to sequences and lists. They’re the natural
architecture of neural network to use for such data. And they certainly
are used! In the last few years, there have been incredible success
applying RNNs to a variety of problems: speech recognition, language
modelling, translation, image captioning.

Essential to these successes is the use of

“LSTMs,” a very special kind of recurrent neural network which works,
for many tasks, much much better than the standard version. Almost
all exciting results based on recurrent neural networks are achieved
with them. It’s these LSTMs that this essay will explore.

4.1 The Problem of Long-Term Dependencies:

33
One of the appeals of RNNs is the idea that they might be able to
connect previous information to the present task, such as using
previous video frames might inform the understanding of the present
frame. If RNNs could do this, they’d be extremely useful. But can they?
It depends.
Sometimes, we only need to look at recent
information to perform the present task. For example, consider a
language model trying to predict the next word based on the previous
ones. If we are trying to predict the last word in “the clouds are in the
sky,” we don’t need any further context – it’s pretty obvious the next
word is going to be sky. In such cases, where the gap between the
relevant information and the place that it’s needed is small, RNNs can
learn to use the past information.

But there are also cases where we need more context. Consider trying
to predict the last word in the text “I grew up in France… I speak fluent
French.” Recent information suggests that the next word is probably

34
the name of a language, but if we want to narrow down which
language, we need the context of France, from further back. It’s
entirely possible for the gap between the relevant information and the
point where it is needed to become very large.
Unfortunately, as that gap grows, RNNs become unable to learn to
connect the information.

4.2 LSTM Networks:

Long Short-Term Memory networks – usually just called “LSTMs” – are
a special kind of RNN, capable of learning long-term dependencies.
They were introduced by Hochreiter & Schmidhuber (1997), and were
refined and popularized by many people in following work.1 They
work tremendously well on a large variety of problems, and are now
widely used.
LSTMs are explicitly designed to avoid the long-
term dependency problem. Remembering information for long
periods of time is practically their default behaviour, not something
they struggle to learn!

35
All recurrent neural networks have the form of a chain of repeating
modules of neural network. In standard RNNs, this repeating module
will have a very simple structure, such as a single tanh layer.

4.3: The repeating module in a standard RNN contains a single layer.

LSTMs also have this chain like structure, but the repeating module
has a different structure. Instead of having a single neural network
layer, there are four, interacting in a very special way.

4.4: The repeating module in an LSTM contains four interacting layers

let’s just try to get comfortable with the notation we’ll be using.

36
In the above diagram, each line carries an entire vector, from the
output of one node to the inputs of others. The pink circles represent
pointwise operations, like vector addition, while the yellow boxes are
learned neural network layers. Lines merging denote concatenation,
while a line forking denotes its content being copied and the copies
going to different locations.

4.2.1 The Core Idea Behind LSTMs

The key to LSTMs is the cell state, the horizontal line running through
the top of the diagram. The cell state is kind of like a conveyor belt. It
runs straight down the entire chain, with only some minor linear
interactions. It’s very easy for information to just flow along it
unchanged.

The LSTM does have the ability to remove or add information to the
cell state, carefully regulated by structures called gates.

Gates are a way to optionally let information through. They are

composed out of a sigmoid neural net layer and a pointwise
multiplication operation.

37
The sigmoid layer outputs numbers between zero and one, describing
how much of each component should be let through. A value of zero
means “let nothing through,” while a value of one means “let
everything through!”

4.3 Evaluation

The model performance is based on their predictive ability. To

evaluate the models the mean error (ME), the mean absolute error
(MAE), the root mean squared error (RMSE) and the continuous
ranked probability score (CRPS) are used.

4.3.1 Mean error (ME)

The mean error is calculated by

ME=1/n∑𝑛𝑖=1 yi − yˆi ,

where yi is the ith observation, yˆi is the prediction and n is the

number of observations. A positive value of the mean error implies
that the predictions tends to be underestimated, and the opposite if
the mean error is negative.

4.3.2 Mean absolute error (MAE)

38
The mean absolute error (MAE) measures the average of the absolute
differences between predictions and observations. It is defined by

MAE=1/n∑𝑛𝑖=1 . |yi − yˆi|

It is a quantity used to measure how close predictions are to the true

observations, without considering the direction of the errors. The
MAE is a linear score which means that all the individual differences
are weighted equally in the average. The error is measured in the
same unit as the original data and the perfects score is 0.

4.3.3 Root mean square error (RMSE)

Another way of quantifying the error is the root mean square error.
RMSE measures the square root of the average squared difference
between the predicted and observed values. It is defined by

Since the errors are squared before they are averaged, the RMSE gives
a relatively higher weight to large errors than to small errors. RMSE is
preferred when large errors are not desirable. The error is measured
in the same unit as the original data and the perfect score is 0.

4.3.4 Weighted scores

39
In some locations there are more snow measurements than in other
locations. If the mean score in each area are found based on all
scoreType equation here.s in the belonging locations, all areas contribute
equally to the final score in MAE, RMSE and CRPS than other areas.
The weighted score account for the difference in number of
measurements in the locations. The weighted score is the weighted
combinations of all the scores in a set of data. For example, the
weighted mean of a dataset {y1, y2,...,yn} with non-negative weights
{w1, w2,...,wn} is

∑𝑛
𝑖=1 𝑤𝑖 𝑦𝑖
𝑦=
∑𝑛
𝑖=1 𝑤𝑖

4.3.5 Evaluation schemes

Cross-validation in model selection

In order to get a more realistic prediction error, we would like to have

a test dataset that is separated from our training dataset. A training
set is the set of data that is used to estimate the model parameters,
and a test dataset is the data used to assess the strength of the
predictive model by using the evaluation criteria. In Table 6.1 a list of
the training and test period is given for each area. Some areas have
more than one period of training and test set. This is because one area
can use multiple weather stations with different years of data, or
because the various locations in one area do not have the same length
in years of historical snow data. The training datasets in Table 6.1 are

40
used to find model parameters both for the model with and without
random effects.

Cross-validation in models with random effects

In the analysis of models with random effects, cross-validation is used.

The data are partitioned k times into subsets, one training set and one
test set. The training set is interpreted as the "measurements of the
year". The data in the training set are randomly sampled from all the
measurements of the specific year. The MCMC analysis is performed
on the training subset, to estimate the random effect. The other test
dataset is retained to evaluate the models by making predictions of
the data. The average of the mean absolute values of the five subsets
are used as prediction error.

CHAPTER 5
41
Implementation Results, conclusion and future
scope
In this chapter we are going to show various results and conclusion
that we have achieved in our implementation.

5.1 Data Collected:

I have two data set Chamoli Rain-fall Data and tagni collected with the
help of the DRDO(DTRL) scientist Mr. Pratik Chaturvedi. When we ran
the script shown in Figure 5.1 .csv file is generated. A simple file for
chamoli is shown in Figure 5.1 a. figure 5.2b .csv file of tagni

Table 5.1: sample of csv file of chamoli rainfall data

42
Table 5.1a: sample file of chamoli rainfall data

Table 5.1b: sample file of tagni sensor data

43
5.2 Apply Normalization Result:

Normalization is a technique often applied as part of data preparation

for machine learning. The goal of normalization is to change the
values of numeric columns in the dataset to use a common scale,
without distorting differences in the ranges of values or losing
information.

When we applied the Normalization Function shown in Figure 5.2, it

will normalize the chamoli rain-fall data and figure 5.2a is Series to
supervised learning for chamoli data.

Figure 5.2

44
 Figure 5.2a

5.3 Split Data for Training: Rain-fall Dataset divided in to two part
training set—a subset to train a model.
test set—a subset to test the trained model.
Scaling for the training rainfall data set is 70 percent and test data is
30 percent and next section I will try to change in the test and training
dataset for prediction LSTM model from I can good prediction.
We ran the python script for splitting data set shown figure in figure
5.4.

Figure: 5.3

5.4 LSTM Model Accuracy: Once we ran the script shown in Figure
5.4, we get the graph for rain-fall reviews training and testing data.

45
 Figure 5.4
5.5 LSTM Model Accuracy at the Different Parameter:
We ran the network design script and fit the designed model for
prediction so we can predict the next value of Rainfall, so I tried to
change the hyperparameter, so I get the best prediction accuracy.
shown figure in 5.5 for accuracy of the prediction model, graph
between actual and predicted values shown figure in 5.5a and
changed in hyper parameter figure 5.5b.

Figure 5.5: best accuracy of the rain-fall Prediction

46
Figure 5.5a: graph between actual and predicted values

Table 5.5b

47
I am trying to get the best accuracy so that I will do working on the smote dataset to get the
best accuracy. Same as I have done work on the tagni data set I got the best accuracy shown
a figure in 5.6a, 5.6b and 5.6c.

Figure 5.6a

Figure 5.6b: LSTM model accuracy on tagni dataset

Figure 5.6c: graph between actual and predicted values

48
 Conclusion
A low-cost lab-scale IoT framework was proposed for monitoring and
predictive landslides. MEMS-based sensors were used in this
framework to measure rain-fall and soil moisture and soil movement
on a ramp. Sensors’ threshold values, derived from the experiment,
were compared to real-time sensor data, and SMS alerts were sent to
the registered mobile phones just when sensor values exceeded their
critical threshold levels.
The results show the potential of using MEMS-based sensors for
landslide monitoring in the real world. However, before deployment
the LEWS framework for monitoring landslides in the real world, we
overcome some challenges, and some of the challenges need to be
overcome. Challenges that we overcame includes making a cavity for
burying sensors on a hill with steep slopes, requiring a self-sustaining
power source, 24-hour monitoring of landslides, and packaging of
sensors so that they can withstand the rugged hilly terrain.
Real challenges come with our future system that requires a series of
sensors in depth of 15 meters range on a hill. For this drilling on steep
slopes requires state-of-the-art drilling equipment and
interdisciplinary expertise in civil and geotechnical areas. For this
purpose, it is encouraged to engage appropriate equipment and an
interdisciplinary team of researchers. A self-sustaining and stable
power source will be needed for real-world landslide monitoring. One
way to provide such a power source is by using solar energy like we
are providing now. However, if landslide monitoring system stays in
an “on” state for 24 hours a day, then the power requirements of such
a system would be very large. Overall, this would increase the
monitoring cost due to the investment in several solar panels and
batteries.
One way to keep the monitoring cost low is to go for a passive
monitoring system, which sleeps for 10 minutes, wakes-up to record
sensor values, and then sleeps again for 10 minutes. If rain or any
movement occurs when the system is sleeping, then the interrupt

49
wakes the system from sleep via a hardware interrupt. As revealed in
the experiments, rain (or water) caused an increase in soil moisture
over time. However, accumulation of moisture in the soil may not
necessarily trigger a landslide, where the landslide trigger is also
dependent upon the type of soil and its properties. Although only one
soil type was tested in the experiment reported above, both soil
properties and soil moisture were considered as causal factors in the
experiment. The experiment’s methodology could be easily adapted
to monitor other soil types.
machine-learning (ML) and deep learning techniques had been used
to predict landslides on a daily scale (Goetz et al. 2015; Catani et al.
2013). our primary goal was to try different ML algorithms to make
minute-scale predictions for rain-fall at a landslide site at Kamand,
Himachal Pradesh. We compared and evaluated the performances of
different ML algorithms that have proven to work well in prior
research prediction. We observed that 5.5 long short-term memory
algorithm outperformed other machine-learning techniques in
predicting soil-movements at the daily scale. This result agrees with
prior literature, 5.5 LSTM had performed accurately for daily-scale
prediction of landslides.
In conclusion, the proposed IoT landslides monitoring and prediction
framework helps to reduce the cost of monitoring landslides in the
real world. In fact, once scaled-up to a real hill, the framework is
expected to cost less than 4% of the cost of conventional systems that
are currently used for monitoring landslides. Given that there are 200
lives lost per year in the Himalayas alone due to landslides (Parkash,
2011), the cost-effectiveness of this IoT framework is expected to be
only INR 72 per head. This low cost is expected to make the framework
an affordable landslide monitoring system in the future – a system
that is deployable at several landslide-prone sites across the world.

50
Future Scope:
Some of future scopes that can be included in our research work are:
> we would like to emphasize on evaluating time-series forecasting
techniques like the Auto-Regressive Moving Average (ARIMA) model
(Khashei & Bijari, 2011).
> we would also like to use more sophisticated performance measures
for model comparison as part of our future research. Measures like
Area under the Receiver Operator Characteristics (ROC) curve
(Japkowicz & Stephen, 2002) and sensitivity-index (d’) (Macmillan &
Creelman, 2010) may provide alternate performance measures for
comparing the performance of different ML techniques.

51
 REFERENCES
1. S. L. Gariano, F. Guzzetti, 2016, Landslides in a changing climate, Earth-Science Reviews, 162, pp
227-252.

2. B. T. Pham, D. T. Bui, H. R. Pourghasemi, P. Indra, M. B. Dholakia, 2017, Landslide susceptibility

assessment in the Uttarakhand area (India) using GIS: a comparison study of prediction capability of
naïve Bayes, multilayer perceptron neural networks, and functional trees methods, Theoretical and
Applied Climatology, 128(1-2), pp 255-273.

3. J. N. Goetz, A. Brenning, H. Petschko, P. Leopold, 2015, Evaluating machine learning and

statistical prediction techniques for landslide susceptibility modeling, Computers & Geosciences, 81,
pp 1-11.

4. J. Mathew, D. G. Babu, S. Kundu, K.V. Kumar, and C. C. Pant, 2014, Integrating intensity–
duration-based rainfall threshold and antecedent rainfall-based probability estimate towards
generating an early warning for rainfall-induced landslides in parts of the Garhwal Himalaya, India,
Landslides, 11(4), pp 575-588.

5. A. Brenning, 2005, Spatial prediction models for landslide hazards: review, comparison, and
evaluation, Natural Hazards and Earth System Science,9 vol. 5, no. 6, pp. 853-862.

6. J. Quinlan, 1986, Induction of decision trees, Machine Learning, vol. 1, no. 1, pp. 81-106.

7. L. Breiman, 2001, Random Forests, Machine Learning, vol. 45, no. 1, pp. 5-32.

8. Tien Bui, B. Pradhan, O. Lofman, I. Revhaug, 2012, Landslide Susceptibility Assessment in

Vietnam Using Support Vector Machines, Decision Tree, and Naïve Bayes Models, Mathematical
Problems in Engineering, vol. 2012, pp. 1-26.

9. V. Vapnik, 1998, Statistical learning theory. New York: J. Wiley.

10. G. E. Batista, R. Prati, M. Monard, 2004, A study of the behavior of several methods for balancing
machine learning training data, ACM SIGKDD Explorations Newsletter, vol. 6, no. 1, p. 20.

11. T. Mikolov, M. Karafiát, L. Burget, J. Cernocký, S. Khudanpur, 2010, Recurrent neural network
based language model, In Interspeech, 2, pp 3.

12. M. Khashei, M. Bijari, 2011, A novel hybridization of artificial neural networks and ARIMA
models for time series forecasting. Applied Soft Computing, 11(2), pp 2664-2675.
49

52
13. F. Catani, D. Lagomarsino, S. Segoni, V. Tofani, 2013, Landslide susceptibility estimation by
random forests technique: sensitivity and scaling issues, Natural Hazards and Earth System Science,
vol. 13, no. 11, pp. 2815-2831.

14. L. Podolski, 2014, Stereoscopic analysis of landslides and landslide susceptibility on the southern
slopes of the Medvednica Mt, Retrieved from [Link]

15. U. Niethammer, M. R. James, S. Rothmund, 2012, UAV-based remote sensing of the Super-Sauze
landslide: Evaluation and results, Engineering Geology, 128, 2-11.

16. D. Zhang, S. H. I. Bin, X. Hong-Zhong, 2004, Experimental study on the deformation monitoring
of reinforced concrete T-beam using BOTDR, Journal of Southeast University (Natural Science
Edition), 4, 12.

17. J. McKean, J. Roering, 2004, Objective landslide detection and surface morphology mapping
using high-resolution airborne laser altimetry, Geomorphology, 57(3), 331-351.

18. P. Chaturvedi, S. Shrivastava, P. Kaur, 2017, Landslide Early Warning System Development
using Statistical Analysis of Sensors’ Data at Tangni Landslide, Uttarakhand, India, Advances in
Intelligent Systems and Computing, Springer International Publishing, 547

19. D. P. Kanungo, A. K. Maletha, M. Singh, N. Sharma, 2017, Ground-Based Wireless

Instrumentation and Real-Time Monitoring of Pakhi Landslide, Garhwal Himalayas, Uttarakhand
(India), In Workshop on World Landslide Forum, pp. 293-300

20. M. V. Ramesh, 2014, Design, development, and deployment of a wireless sensor network for
detection of landslides. Ad Hoc Networks, 13, 2-18.

21. E. Frank, M. Hall, and I. Witten, 2016, Data Mining: Practical Machine Learning Tools and
Techniques, 4th ed. Morgan Kaufmann.

22. R. Duda, 2004, Pattern Classification 2nd Edition with Computer Manual 2nd Edition Set. John
Wiley & Sons.

23. N. Macmillan, C. Creelman, 2010, Detection theory. New York, NJ [u.a.]: Psychology (In Press).

24. N. Mali, P. Chaturvedi, V. Dutt, V. U. Kala, 2017, Training of Sensors for Early-Warning System
of Rainfall-Induced Landslides, Proceedings of 19th International Conference on Soil-Mechanics and
Geotechnical Engineering, Sydney, Dec. 2017 (In Press).

25. S. Kaushik, A. Choudhury, N. Dasgupta, L. Pickett, V. Dutt, 2017, A study of Statistical

and Predictive Analysis of US Pain Medications. In International Conference on Machine-
Learning and Data Science, IEEE Conference (In Press).

26. K. Agrawal, Y. Baweja, D. Dwivedi, R. Saha, P. Prasad, S. Agrawal, S. Kapoor, P. Chaturvedi, N.

Mali, V.U. Kala, V. Dutt, 2017, A Comparison of Class Imbalance Techniques for Real-World
Landslide Predictions. In International Conference on Machine-Learning and Data Science, IEEE
Conference (In Press).

53