Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

9964 IEEE SENSORS JOURNAL, VOL. 19, NO.

21, NOVEMBER 1, 2019

Effective and Accelerated Forewarning of


Landslides Using Wireless Sensor Networks
and Machine Learning
T. Hemalatha, Maneesha Vinodini Ramesh, Senior Member, IEEE, and Venkat P. Rangan, Senior Member, IEEE

Abstract— In this paper, we have reported an effective adop- the southern Himalayan arc and the south-west Indian coast,
tion of machine learning (ML) algorithms for wireless sensor both of which are located in India [9]. David Petley’s dataset
networks (WSN) deployed for real-time monitoring and early [9] on landslide fatality provides deeper insights into the
warning of landslides. The WSN system performs real-time
monitoring of landslide triggering parameters such as rainfall, statistics and the trend of landslides. Petleys dataset has logged
soil moisture, pore-water-pressure (PWP), and movements. The 2620 non-seismically triggered landslides over a span of seven
challenges involved in landslide early warning and the data driven years, ranging from 2004 to 2010. Over the years, a sustained
context for overcoming these challenges are being presented. increasing linear trend of the number of landslide fatalities has
Linear models were applied on real-time data from the WSN been reported. Despite the advent of modern techniques for
system and the data from laboratory experiments to discover
the interrelationships existing between rain, soil, and moisture, landslide monitoring and early warning, providing sufficient
PWP and construct the feature-set required for forecasting warnings ahead of time and minimizing loss of lives and
slope stability conditions. Two algorithms Current-PWP and 24- property damage remains a challenge [10].
PWP based on the support vector regression (SVR) method is
implemented for efficacious nowcast and 24 h forecast of the slope
stability conditions for early warning notification of landslides. A. Existing Methods and the Limitations
Comparison of the results from the Current-PWP and 24-PWP
algorithms with the actual sensor measurements recorded by Landslide is a complex phenomenon affected by diverse
the WSN system revealed mean absolute error (MAE) values causative parameters including rainfall, seismic activity,
4.2k and 4.7 kPa, respectively. The results section elaborates 1) weathering, moisture dynamics, PWP dynamics, seepage and
on relevance of the Current-PWP algorithm towards ensuring drainage patterns and movements in the slope etc. Several
the reliability of early warning WSN systems, constrained by
data availability, 2) application of the 24-PWP algorithm to yield methods of remote sensing through satellite imagery for
an additional duration of 24 h for landslide preparedness, 3) landslide monitoring are in use today [11]–[13]. The main
the empirical appraisal of the regression model from laboratory advantage of such methods is that, they can be used for
experiments for estimation of soil moisture from PWP, and vice surveillance of large areas, with high spatial resolution [14]
versa, 4) the efficacy of multi-level decision making for the and 3D capabilities [15]. Remote sensing techniques provide
preclusion of false warnings, and 5) the timely early warnings
issued from the proffered WSN system. excellent results in vulnerability mapping [7], [16], landslide
risk mapping [7], [17] and, post disaster mapping [18],
Index Terms— Wireless sensor networks, machine learning, but limited for real-time monitoring, entailed by the long
forecasting, decision support system.
duration consumed by a satellite for revisiting the same
I. I NTRODUCTION place [19]. Seismic [20], Electromagnetic [21], Ground Pen-
etrating Radar [22], Electrical Resistivity Tomography [23],
W SN has a great lifesaving prudential for disasters
such as landslides [1]–[3] and application of ML in
early-warnings of natural or manmade calamities is one of
constitute few of the noninvasive, geophysical landslide mon-
itoring methods, employed to investigate large volumes of
subsurface regions. However these approaches being indirect,
the most impactful societal applications [4]–[8]. Incidents of
lead to a number of non-unique solutions, which limit their
catastrophic landslides have been reported from many regions
usage in reliable warnings [24]. Landslide monitoring methods
of the world including Asia, Europe and the Americas. Among
centered on geo-technical instruments like extensometers [25],
them the first two emerging global hotspots for landslides are
inclinometers [26], and piezometers [27] afford accurate mea-
Manuscript received June 1, 2019; accepted June 26, 2019. Date of surements, but these measurements are confined to smaller
publication July 12, 2019; date of current version October 4, 2019. This work areas where the instrument is installed. Instrument based
was partly funded by Ministry of Earth Sciences (MoES), Government of
India under the project titled “Advancing Integrated Wireless Sensor Networks methods cannot be used for a large scale deployment and
for Real-time Monitoring and Detection of Disasters” and partly funded by has its limitation for regional monitoring. These methods
Amrita University. The associate editor coordinating the review of this paper are quite cumbersome and expensive, involving heavy duty,
and approving it for publication was Prof. Huang Chen Lee. (Corresponding
author: T. Hemalatha.) unwieldy equipment and their effectiveness is limited to at
The authors are with the Amrita Center for Wireless Networks and most detecting the event [5], [28]–[30] and cannot really be
Applications (AmritaWNA), Amrita School of Engineering, Amrita Vishwa used for early warning.
Vidyapeetham, Kollam 690525, India (e-mail: hemalathat@am.amrita.edu;
maneesha@amrita.edu; venkat@amrita.edu). On the other hand, Early Warning System (EWS), based on
Digital Object Identifier 10.1109/JSEN.2019.2928358 rainfall threshold models from rain gauge are inexpensive and
1558-1748 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HEMALATHA et al.: EFFECTIVE AND ACCELERATED FOREWARNING OF LANDSLIDES USING WSN AND MACHINE LEARNING 9965

widely used models for landslide early warning [31], [32].


However, the warnings from these methods are vulnerable
to a large number of false positives and fail to provide a
reliable solution [33] either due to lack of sufficient historical
data or lack of fine grained rainfall data in both spatial and
temporal domain. Rainfall thresholds are highly general which
can provide regional warnings and if they have to be tailored
to be effective for a specific site they require a long number
of years of rainfall measurements and correlating them with
landslide events [32], [33]. Moreover, monitoring one vital
parameter related to landslide will not suffice to address the
need for early warning of landslide with certainty [34].
B. Our Approach Fig. 1. Landslide monitoring and early warning system deployment site
Our goal is to develop highly accurate algorithms that − Anthoniar Colony, Munnar, Kerala, India. Location of deep-earth probes
(DEPs) are indicated as L1, L2,.., rain gauge is indicated as RG, and
make use of sensor measurements. We embarked on such the location of gateway of the wireless sensor network (WSN) is also
an effort to design a multi sensor WSN based system that indicated [35].
captures multiple parameters such as PWP, soil moisture,
ground vibrations and movements in addition to rainfall for Deep Earth Probe (DEP), which incorporates these sensors
providing accurate site specific, near site-specific and large- at different depths under the ground. Deep bore holes were
area early warnings [35], [36]. In order to provide reasonably drilled in selected vulnerable locations (L1, L2,.. L6) as
dependable, reliable, and adequately advance early warnings depicted in Fig. 1. and DEPs accompanied with sensors are
a data driven approach was adopted. We have integrated deployed in these locations [2], [35]. More details regarding
SVR, linear and non-linear models on multi parameter data. the deployment of sensors and DEP’s can be found in [2] and
We have successfully devised the algorithm to (i) arrive at a [35]. Over 150 sensors at Munnar, and more than 200 sen-
24 hour ahead forecast of the slope stability conditions from sors are deployed at Chandmari. Each DEP is constituted
rainfall and PWP sensor measurements (ii) learn from the of multiple piezometer sensors, moisture sensors, tilt meters,
historical data and predict the real-time conditions in the slope. strain gauge, inclinometer and geophone at different depths.
The ML algorithms facilitated predictions of the real-time Next, the sensors in the DEP are then interfaced to a Data
conditions of the slope, which rendered the design of a resilient AcQuisition (DAQ) board and a sensor mote to form a WSN
system, during times of disaster-in-progress, when there is a system. Rain gauge is deployed at the location RG shown in
limitation to availability of reliable data. Another advantage of the Fig. 1. and it is connected to the WSN system.
incorporating ML to multi sensor data is that after learning the These sensors in the WSN system sense the vital parameters
slopes conditions for few years the learned knowledge can be like precipitation rate, soil moisture, PWP, ground vibration,
used as a virtual sensor, while the actual sensors can be reused slope movements respectively from heterogeneous soil layers.
for re-deployment at other sites. The early warnings, 24 hour Data from the WSN system are transmitted through hetero-
ahead forecast of slope stability conditions and predicting the geneous wireless network to the Data Management Centre
real-time sensor data are all validated by comparing it with in Amrita Vishwa Vidyapeetham [2]. Deployment site in
the real-time data from the WSN system. Munnar with the locations of DEP’s (L1, L2, ..L6) rain gauge
(RG) and gateway node is shown in Fig. 1. In the Data
II. OVERVIEW OF WSN BASED L ANDSLIDE M ONITORING Management Centre, the WSN data from the field is analyzed
AND E ARLY WARNING S YSTEM in real-time for the following (i) Rainfall Intensity-Duration
The landslide monitoring and EWS is deployed in two thresholds, (ii) Factor of Safety (FoS) of the slope, (iii) Tilt and
places (1) Munnar in Western Ghats, and (2) Chandmari in vibrations from movement sensors. FoS is a non-dimensional
North Eastern Himalayas. The average annual rainfall in these number that expresses the stability conditions of the slope at
regions are higher than 2500 mm. The monitoring terrain different locations, this value is calculated from slope angle,
in Munnar is approximately 7 acres of slope in Anthoniar soil properties and the PWP levels at that particular location
colony, (10.0892 N, 77.0597 E) with an elevation of 1532 m in the slope [42]. The FoS from Iverson model [42] is given
above sea level and is located in the landslide prone Western as
Ghats mountain region of Kerala in India. Idukki district, FoS = F f + Fw + Fc = 1 (1)
where Munnar is located is a landslide hotspot in the Western
Ghats and every year landslides are very common during mon- Here F f ,is a slope factor, Fw is a hydrogeological factor
soon season, [37]–[39]. The monitoring terrain in Chandmari and Fc is a soil cohesion factor and is given by
(27.3415 N, 88.6251 E) is approximately 150 acres of slope tanϕ
Ff =
with an elevation of 1650 m above sea level and is located tanα
in the landslide prone Sikkim Himalayas mountain region in −PW P(Z , t) γw tan
Fw =
India [40], [41]. γs Z si nα cosα
To characterize the changes materializing in the subsur- c
Fc = (2)
face using these sensors, we have designed and developed a γs Z si nα cosα
9966 IEEE SENSORS JOURNAL, VOL. 19, NO. 21, NOVEMBER 1, 2019

where ϕ is a soil friction angle, α is the slope angle, c is the


soil cohesion, γs is the depth averaged soil unit weight, γw
is the unit weight of ground water, PW P(Z , t)is the pressure
distribution with respect to depth Z over time t. In the above
equation the dynamically varying term is the pressure head
distribution PW P(Z , t), terms such as c, γs , ϕ are calculated
via soil test in laboratory, and terms α, γw are constants. When
PWP in the soil increases, it results in the decrease of effective
stress and reduction of strength in the soil. FoS equation
describes the failure in slopes which is characterized by the
ratio of downslope gravitational driving force to the resisting
stress mediated by PWP. Theoretically, both the forces balance
when the ratio is equivalent to 1. If the FoS value is greater
than 1; i.e FoS > 1, then the effective stress in the soil is
greater than the downward gravitational force and therefore
slope is stable at that location depth. When there is higher
PWP, the effective stress decreases and the FoS value becomes
lesser than 1; i.e FoS < 1 then the slope is susceptible to
Fig. 2. Temporal variation of moisture (at 1.2m depth deployed in the location
failure at that location depth. In our landslide early warning DEP-6) and PWP sensor (at 14m depth deployed in location DEP-6) with
system, FoS are determined at each of the vulnerable locations respect to rainfall for the years 2011 to 2016.
at different depths using the real-time pore pressure valued
collected from the piezometers deployed in those specific loca-
tion. The real-time system has the capability to continuously III. C HALLENGES OF F OREWARNING
collect these data and analyse the FoS variability in real-time Numerous mathematical models and relationships are devel-
with respect to the dynamically varying pore pressure values. oped between parameters for predicting landslides in real-
These FoS value act as the indictor for the imminent land- time [42]–[44]. However issuing early warnings with higher
slides. Aggregate FoS value from multiple location in the slope accuracy levels ahead of time is a challenging task. For the
will be used to determine the stability condition of the whole purpose of making decisions and providing early warnings
slope. ahead of time, we have adopted a data driven approach
A three level early warning system “EII” constituting of on the WSN data since 2009 and this has provided deep
“Early”, “Intermediate”, “Immediate” levels is designed to insights into the seasonal trends, spatial relationship, tem-
issue notifications of landslide warnings to the relevant govern- poral relationship, and interrelationship between parame-
ment agency and the general public. The multi-level warnings ters [45]. In this section we discuss about the seasonal
are derived as follows. trend, spatial-temporal variability and interrelationships of
• Early: Issued whenever the rainfall rate crosses its thresh- three vital parameters rain, moisture and PWP contributing
old. Site specific rainfall threshold for Munnar are derived to landslides. In this section we also discuss about the
from the historic rainfall data that caused landslide inci- challenges faced in providing early warnings from these
dents in Munnar. parameters.
• Intermediate: This warning is derived from the FoS
as shown in equation 1 computed for the slope from A. Partial knowledge of the Causative Relationship of
the forecasted and measured PWP. As mentioned ear- Parameters Leading to Landslides
lier, FoS of the slope is derived from PWP mea-
The spatial and temporal variation of soil moisture and PWP
surements in the slope and since slopes nearby the
data with respect to the rainfall for the years 2011 to 2016 is
deployment site possesses similar geology and receive
shown in Fig. 2. It can be seen from the graph that there
approximately the same amount of rainfall, the warning
is a seasonal trend associated with moisture and PWP with
based on FoS is applicable to the nearby slopes also.
respect to the rainfall rate. The peak values of PWP do not
Therefore the second level warning is a near site spe-
exactly coincide with the rainfall peaks. To attain a better
cific warning and provides early warning based on the
understanding of the interrelationship between these parame-
forecasted FoS.
ters, linear models were created for various antecedent rainfall
• Immediate: This warning is derived from the movement
conditions with moisture and PWP, which are elaborated in
sensors, when dynamic variability in stress is derived
Algorithm-1 in section V. The results of these models in terms
from the observations collected from these sensors at
of Pearson’s correlation coefficient is shown in Table-I. From
the deployment site. The movements observed are very
the table it can be seen that the correlation values for all
particular to the slope and therefore this is a site-specific
antecedent rainfall conditions with moisture and PWP are less
warning.
and the maximum correlation that can be obtained is 0.69. The
The challenges faced while providing early warnings is dis- correlation values in Table-I, delineate that rainfall measure-
cussed in the next section. ments and antecedent conditions are only partially correlated
HEMALATHA et al.: EFFECTIVE AND ACCELERATED FOREWARNING OF LANDSLIDES USING WSN AND MACHINE LEARNING 9967

TABLE I
A LGORITHM -1: R ESULTS OF L INEAR M ODEL B ETWEEN A NTECEDENT
R AIN Ar FOR t N UMBER OF D AYS VS M OISTURE ( f L − moi ),
A NTECEDENT R AIN Ar FOR t N UMBER
OF DAYS VS PWP ( f L − pwp)

Fig. 3. Spatial-Temporal variation of PWP sensors at various DEP locations


and depths for the year 2013 (January to December).

• PWP at 14m is at the location DEP-6 (middle region of


the slope) and the PWP value here varies from (0kPa to
45kPa)
• PWP at 8.03m is at the location DEP-5 (middle region of
the slope) and the PWP value here varies from (-90 kPa
to -40kPa)
• PWP at 6.15m is at the location DEP-2 (toe region of the
slope) and the PWP value here varies from (-30 kPa to
-20 kPa)
• PWP at 5.5m is at the location DEP-1(toe region of the
slope), but the PWP value here varies from (25kPa to
55kPa)
• PWP at 8.5m is at the location DEP-1(toe region of the
with moisture and PWP, which do not give a comprehensive slope), but the PWP value here varies from (-38kPa to
view of moisture and PWP dynamics. Furthermore, Fig. 2. 25kPa)
portrays that the moisture and PWP dynamics do not exhibit • PWP at 12.5m is at the location DEP-1(toe region of the
the same pattern every year and their variations are also slope), but the PWP value here varies from (-40 kPa to
affected by various geological and hydrological parameters. 10 kPa)
From the yearly seasonal pattern variations as shown in Fig. 2. From the foregoing notes it can be seen that the spatial
and from the partially correlating relations in Table-I, it can be variation of PWP is not uniform, at all DEP locations and
understood that there are some other parameters also which are depths. It can also be seen from the Fig. 3. that the PWP values
contributing to slope instability and it could be due to seepage, at different DEP locations and depths on a temporal scale
drainage, groundwater levels, weathering of rocks, manmade doesn’t follow the same trend, and the peaks of PWP values
slope modifications etc., which cannot be measured directly. are attained at different times. There is a spatio-temporal
Therefore the knowledge we have from the sensors are partial, variability of PWP values all over the slope and the there
and the sequence of impact of all the parameters leading to a is a difference in response time (rise of the PWP values)
disaster is unknown. In real-world most thresholds are derived for the same trigger, rainfall. Hence the current knowledge is
from this partial knowledge leading to uncertainities in early inadequate to estimate the expected variability in a particular
warning. parameter based on the variance in the other known parame-
ters. This spatio-temporal variability has an impact not only
on generating landslides but also on the expected time window
B. Spatial-Temporal Variability of Parameters at which the landslide may initiate. In a real-world scenario
both these are unknown.
The known and the unknown parameters in a slope con-
tributing to landslides vary spatially from one location to
C. Reducing False Warnings
another depending on the hydro geological properties over
time. Fig. 3. shows the spatial and temporal variability of PWP As discussed earlier, conventional EWSs are based on Rain-
measurements in the field for the year 2013. PWP observations fall threshold models. The shortcomings with these models
from Fig. 3 are summarized below ensue from the following (i) model inherently integrated with
9968 IEEE SENSORS JOURNAL, VOL. 19, NO. 21, NOVEMBER 1, 2019

uncertainties, as only one of the parameters, contributing


to slope instability has been reckoned with, (ii) enhanced
level of uncertainty due to the the unavailability of rainfall
observations, in the required spatial and temporal scales (iii)
lack of knowledge of landslide events and the accurate time
and duration of such incidents (iv) the lack of prior knowledge
of prospective landslide areas constrains the development of
rainfall models, specific to each site, leading to uncertainty in
the validity of the warnings; since they were primarily devel-
oped for global or regional scale scenarios. For the Munnar
deployment site, rainfall thresholds models such as [31], [46],
[47], were implemented using the historic rainfall data from
2004-2009. The derived rainfall intensity duration thresholds
delivered 25-40 false positive warnings, however the region Fig. 4. Amrita’s large scale landslide laboratory set-up.
has experienced less than ten landslides. If we had initiated
warning dissemination based on the rainfall threshold alone,
then there would have been many false positive warnings A. Learning Algorithms to Capture the Causative
leading to ambiguities in the veracity of the warnings. Relationship of Parameters Leading to Landslides
The problem of learning the interrelationship between the
vital parameters over time and to the landslides, is modelled
D. Reliability of the System During Disasters as a regression problem and we have used SVR to address
During torrential rainfall in monsoon season, the solar irra- the same. SVR is applied on both the historic and real-time
diance in Munnar is lesser than 2 kW h/m 2 /day. Sometimes data to (i) approximate the underlying unknowing mapping
due to heavy rainfall of more than 100 mm to 300mm in a day, of the input data from DEP to the output landslide or no
the power grid will also be completely shut down for varying landslide occurrences and (ii) approximate the underlying rela-
durations of time, which adversely affects and delays the data tion existing between different vital parameters, thus approx-
transfer. During unscheduled power outages,the reliability of imating the unknown relation in the mapping function. This
the WSN system to collect data for monitoring and early learned knowledge of the unknown relation from the mapping
warning is a formidable challenge. function, helps in predicting landslides with higher accuracy
for the current situations prevailing in the slope. In this
paper we have shown results for mapping the underlying
E. Inadequacy of Time for Fore Warnings unknown complex relationship existing between rainfall and
PWP in Algorithm-3, the Current-PWP algorithm. We have
In the course of early warnings, inadequacy of time for
also compared the interrelationship between rainfall and PWP
decision making, warning dissemination, relocating people and
captured using linear model in Algorithm-1, and the Current
their valuable assets to safer locations is a substantial practical
PWP model in Algorithm-3.
difficulty.

B. Deployment Strategies and Laboratory Experiments for


F. Cost of Deployment Capturing Spatio-Temporal Variability of Parameters
The instruments and sensors in the WSN system contributes In order to capture the spatial and temporal variability,
to the cost of the system and if this system has to be extended sensors in the field are deployed in different soil layers based
for deployment in large areas, it will add on to the cost. on the bore-hole logs characterizing them. The sensor in
The challenges mentioned above are prevalent in most of the each soil layer is representative of the subsurface in that
landslide monitoring and early warning applications. In next particular layer. In addition to that, laboratory experiments of
section, we give an overview of our approach to solve the creating landslides under controlled environmental conditions
above mentioned challenges. are conducted to simulate the variations happening in the vital
parameters leading to landslides in the field. Our research
center is facilitated with a large scale landslide laboratory
IV. O UR A PPROACH TO S OLUTION
setup as shown in Fig. 4. for studying and understanding
Landslide domain experts have addressed the above men- landslides by performing experiments. By controlled envi-
tioned challenges from the geological, simulation and labo- ronmental conditions, we mean to replicate a particular soil
ratory experimental contexts. In this paper, these challenges layer in the field and conduct experiments by simulating
have been addressed from a data driven context, which helps different rainfall conditions. By conducting an experiment for
overcome these challenges. The propounded solutions with creating landslides on a particular soil-layer, we can capture
respect to the learning, nowcasting and forecasting models as the temporal variation happening in the field. By conducting
well as slope stability conditions using the landslide triggering several such experiments and by replicating different soil
parameters - rainfall, moisture and PWP are described below. layers, we can collect the spatio-temporal variations of the field
HEMALATHA et al.: EFFECTIVE AND ACCELERATED FOREWARNING OF LANDSLIDES USING WSN AND MACHINE LEARNING 9969

without actual deployment in the field. ML techniques are also F. Cost Savings via Virtual Sensors and Reuse of Sensors
applied on this laboratory data to discover the relation between From a practical perspective, the cost of the deployment
different vital parameters and understand the spatio-temporal increases with an increase in the number of installed sensors.
variability of these parameters. The experimental results for Advantages of Current-PWP and 24-PWP models can reduce
the spatio-temporal variability of Munnars top soil layer and the cost of the deployment without compromising on the
deriving the interrelationship between parameters is shown in reliability of the system with reduced number of sensors. Our
the results section. ML models, Current-PWP model and the 24-PWP model we
have created uses rainfall as the input to learn PWP dynamics
and predicts PWP values using rainfall. After learning the
C. Multi-Level Decision Mechanism for Reducing False terrain for few years, the learned knowledge can be used as
Warnings a virtual sensor to approximately predict the slope stability
A multi-level decision mechanism was adopted to reduce the conditions from rainfall measurements or rainfall forecast
number of false alarms. This multi-level decision mechanism information. After collecting adequate data and learning the
makes decision at various stages from different sensor data. terrain for few years; relying on the learned knowledge, and
The current EWS are not tailored for making decisions at based on the domain expertise, we can decide to retain few
various stages. We have an effective combination of regional sensors and remove few sensors for reusing in other locations,
warnings from rainfall threshold, near-site specific warnings which brings down the deployment cost. In other words, after
from PWP response and FoS of the slope and site-specific adequate data has been collected from the sensor at one
warnings from movement sensors. During torrential rainfall, location, the actual sensor can be removed from that location
regional warnings helps in disseminating the likelihood for and repeat the process of collecting adequate data for other
landslides over a large area. Whereas near-site specific and locations.
site specific warnings provide accurate and reliable warnings
in highly populated areas where the deployment is performed. V. M ETHODOLOGY
The real-time multi-level warnings issued is discussed in The investigation reported in this paper utilized data
results section. acquired from a rain gauge, moisture sensor, and piezometer
sensor, for now-casting and forecasting of landslides, with
higher accuracy.
D. Achieving Reliability Through Learned Knowledge
A. Linear Models
As mentioned in previous sections, Algorithm-3, “Current
PWP” can be used for real-time nowcasting of PWP val- Rainfall or lack thereof is the prime contributing parameter
ues, from rainfall measurements. The major breakthrough for build-up or decline of moisture and PWP. As far as rainfall
discerned in the process of application of ML to the accquired is concerned, the antecedent rainfall has an impact on the
data was that even when the rainfall data and other sensor build-up along with the rainfall on the current day of the
data are unavailable, during adverse weather conditions, vital landslide event [36]. Furthermore, the amount of antecedent
parameters like moisture, and PWP can be predicted from rainfall conditions, that accounts for moisture and PWP build
rainfall forecast and we were successful in doing so. The rain- up at different soil layers and at different depths is unknown.
fall forecast data is provided by several agencies like Indian Algorithm-1 tries to find any direct linear relations existing
Meteorological Department (IMD), Skymet, etc. The results between Antecedent rainfall conditions Art with soil moisture
of the real-time prediction of vital parameters from rainfall and PWP at different location-depths. In Art , Ar represents
forecast information, (while actual rainfall data not available) antecedent rainfall for t number of days, here t varies from
during the monsoon period (June-September) in 2018 is shown 6 to 120 (number of days). We describe linear functions for
in Fig. 7. in the results section. pore water pressure versus antecedent rain, f L − pwp, moisture
versus antecedent rain, f L − moi , at various depths of preset
DEP locations. D E Pi Pk represents pore water pressure P at
depth k in the i t h DEP. D E Pi M j represents moisture M at
E. Forecasting methodologies to overcome inadequacy of
depth j in the i t h DEP. The output of the linear function
time:
in terms of Pearson’s correlation coefficient is checked for
In order to forewarn landslides ahead of time, an SVR correlation values greater than 50 %. If such a correlation
algorithm “24-PWP” shown in Algorithm-4 was devised, for exists then that antecedent rainfall condition are returned and
forecasts of PWP data from rainfall measurements, 24 hours used further in constructing training and testing data sets.
ahead of time. Output data from “24-PWP” algorithm can The results of Algorithm 1 for a moisture sensor at 1.2m
be used to calculate 24-FoS as shown in Algorithm-5, and depth and PWP at 14m depth deployed at the location DEP-
Fig. 9 which is the forecasted values of FoS 24 hours ahead 6 is shown in Table-I. From Table-I it can be seen that, with
of time. These 24-FoS values from 24-PWP algorithm afford the increase in the number of days of antecedent rainfall,
an extra time of 24 hours which is a boon for the researchers, the correlation with both moisture and PWP increases. The
government agencies and the general public to be prepared for maximum correlation, that can be obtained between Art and
landslides. Details of the same are shown in results section. moisture is 0.44, whereas the maximum correlation that can
9970 IEEE SENSORS JOURNAL, VOL. 19, NO. 21, NOVEMBER 1, 2019

be obtained between Art and PWP is 0.69. While comparing due to antecedent rainfall, short duration high intensity rainfall,
the correlation values between moisture and PWP, PWP is cor- extreme rainfall events or a combination of these. Therefore we
relating more with Art . Since PWP shows higher correlation need to consider both antecedent rainfall conditions and short
with Art , when compared to moisture, Art and PWP were term rainfall conditions for forecasting landslides. In a study
used, along with soil properties, for creating learning models. of extreme rainfall events, investigators in [49], highlighted
the increase in the frequency of 1-5 days of extreme rainfall
B. Current-PWP- Nowcasting of PWP maxima due to the anthropogenic warning in south India.
In recent days, many extreme rainfall events have been noticed,
From the literature, PWP response due to rainfall is con- all over south India. Therefore, both long term antecedent rain-
sidered to be more vital because, PWP is one of the main fall condition and short term rainfall need to be accounted for.
factors contributing to slope instability. Increase in PWP will The former contributes to PWP build-up leading to landslides,
lead to decrease in shear stress [48], which can lead to slip while the latter can create shallow landslides (if it is of high
surface formations. Moreover at the Munnar deployment site, intensity).
after moisture level saturation, PWP increases drastically in a →

Creating rainfall feature set vector fs : Short-term rainfall
span of two to three days during the monsoon period. It can Sr varying for t number of days is considered along with the
be seen from Fig. 2 and Fig. 3. that the PWP remains constant highly correlating antecedent rainfall condition Art∗ − pwp
at the beginning of the monsoon rains in June. The value of from Algorithm-1 for creating the rainfall feature set vector
PWP starts increasing after adequate amount of rainfall. When →

fs . Firstly, all unique combinations of short term rainfall,Srt
moisture values have saturated, PWP shows a drastic increase
is created, here t varies from 1 to 5 number of days, which
within a short duration of two to three days of elapsed time.
will result in a total of 5C1 + 5C2 + 5C3 + 5C4 + 5C5 = 31
Therefore, it is highly important to predict the PWP dynamics
combinations. The 31 combinations are [Sr1 ], [Sr2 ], [Sr3 ],
from other known parameters to estimate the slope stability
[Sr4 ], [Sr5 ], [Sr1 , Sr2 ], [Sr1 , Sr3 ], [Sr1 , Sr4 ], [Sr1 , Sr5 ],
conditions.
[Sr2 , Sr3 ], [Sr2 , Sr4 ], [Sr2 , Sr5 ], [Sr3 , Sr4 ], [Sr3 , Sr5 ],
[Sr4 , Sr5 ], [Sr1 , Sr2 , Sr3 ], [Sr1 , Sr2 , Sr4 ], [Sr1 , Sr2 , Sr5 ],
Algorithm 1 Antecedent Rain Art Relation With Moisture [Sr1 , Sr3 , Sr4 ], [Sr1 , Sr3 , Sr5 ], [Sr1 , Sr4 , Sr5 ],
amd PWP [Sr2 , Sr3 , Sr4 ], [Sr2 , Sr3 , Sr5 ], [Sr2 , Sr4 , Sr5 ],
Input: Art , D E Pi M j , D E Pi Pk [Sr3 , Sr4 , Sr5 ], [Sr1 , Sr2 , Sr3 , Sr4 ], [Sr1 , Sr2 , Sr3 , Sr5 ],
Output: Art∗ − pwp, Art∗ − moi [Sr1 , Sr2 , Sr4 , Sr5 ], [Sr1 , Sr3 , Sr4 , Sr5 ], [Sr2 , Sr3 , Sr4 , Sr5 ],
[Sr1 , Sr2 , Sr3 , Sr4 , Sr5 ].
for x = 1 to i do Secondly, we append the highly correlating antecedent
f L− pwp (Art , D E Px Pk ); Pk ∈ {P1 , P2 , , ...Pn } rainfall conditions Art∗ − pwp for PWP in each DEP location-
depth D E Pi Pk with all 31 unique combinations of short term
compute R 2 ( f L− pwp (Art , D E Px Pk )) rainfall Srt .
f L−moi (Art , D E Px M j ); M j ∈ {M1 , M2 , ...Mq }


compute R 2 ( f L−moi (Art , D E Px M j )) Algorithm 2 Creating Rainfall Feature Set Vector fs
end for Input: Srt , Art∗ − pwp for each D E Pi Pk


Output: Rainfall feature set vectors fs
if R 2 ( f L− pwp ) ≥ 0.5 then 1. Compute all combinations of Srt and store in Combset
return Art∗ − pwp for r = 1 to (n-1) do
end if
CombSet{} = nCr
if R 2 ( f L−Moi ) ≥ 0.5 then end for


return Art∗ − moi 2. Create rainfall feature set vector fs by appending Comb-
end if Set{} with Art∗ − pwp

for each Pk in D E Pi do
As mentioned before in section-III, the results from Table-I
for each Art∗ − pwp do
indicates that there is only a partial relation existing between
for i = 1 to length(CombSet) do
antecedent rainfall conditions with PWP dynamics. To dis- →

fs [i] = [CombSet{i}, Art∗ − pwp]
cover and model the unknown relation existing between rain-
end for
fall and PWP dynamics, we need to perform an assumption
end for
free learning and therefore, we have used SVR based learning
end for
model in Current-PWP algorithm to discover the underlying
complex relation. In this section we have discussed about the →

dataset construction and the Current-PWP algorithm. Creating training set vector fTr and testing set fTe : Apart
1) Dataset Construction: The known parameters to estimate from rainfall, the other known parameter contributing to PWP


the PWP dynamics and the resulting landslides are rainfall dynamics is the soil properties. A training set fTr is formed
conditions and soil properties. Landslides can be caused either from the known soil properties and the rainfall feature set
HEMALATHA et al.: EFFECTIVE AND ACCELERATED FOREWARNING OF LANDSLIDES USING WSN AND MACHINE LEARNING 9971



vector fs as the input and the corresponding PWP as the dataset from fTr∗ with the input variables at time stamp ‘T1’


output. Historic data is used for creating the fTr . There are and the output variable with time stamp ‘T1 + 24 hours’. The

− ultimate aim of training the support vector regression machine
31 training data sets in a single training set fTr corresponding
to the highly correlated antecedent rainfall model Art∗ − pwp in such a way is to forecast pore-water pressure that develops
for every DEP. Selecting the best training data set fTr∗ from the 24 hours later from the current rainfall and antecedent rainfall

→ conditions. C-SVR algorithm with radial basis kernel is used
31 data sets in fTr is an iterative process and it is accomplished
during the training phase in Current-PWP algorithm. Current in this section also for forecasting the 24-PWP. The algorithm
PWP model is formed from the best training data set fTr∗ for forecasting PWP values 24 hours ahead is detailed in
selected. The historic dataset since the year 2009 was used Algorithm-4, 24PWP.


for creating the training set vector fTr .
D. 24-PWP to 24-FoS; 24-FoS to Forewarning
The real-time streaming data from the WSN system is
prepared in the same format as fTr∗ and used for creating the Slope stability analysis is done for the Munnar slope and
testing data fTe . The Current-PWP model is tested using fTe , we have derived the Factor of Safety (FoS) equations for
and the output will be the predicted PWP value for the given various locations of the slope [35], [42]. In the FoS model,
fTe . Current-PWP algorithm is explained below. PWP is the dynamically changing parameter along with other
2) Current-PWP Algorithm: In the Current-PWP algorithm, constant parameters. The 24 predictive PWP can be used to

− derive FoS 24 hours ahead of time and is denoted as 24-FoS.
the best data set fTr∗ is selected from the training set vector fTr
for creating the Current-PWP model. The Current-PWP model This 24-FoS conveys about the stability conditions of the
is updated in real-time by accomplishing the learning in real- slope 24 hours ahead, which can be used for forewarning for
time which comprises of the historic data from past to the landslides 24 hours ahead.
recently streamed data. The learned knowledge is then used
to predict the PWP from the new rainfall data streaming in Algorithm 4 24-PWP
real-time. The actual PWP values from the WSN system and Training:
the predicted PWP are compared to assess the accuracy of Input: fTr∗ for each D E Pi Pk
prediction. SVR learning algorithm learns and adapts itself in Output: 24-PWP model for each D E Pi Pk
real-time by continuously adapting its weight as and when the
1. Create training data set from fTr∗ with input values at time
new data streams. C-SVR algorithm with radial basis kernel
stamp ‘T1’ and output values at time stamp ‘T1+24 hours’
is used for modelling the Current PWP.
2. Train the SVR with the training set and fine tune the
parameters
Algorithm 3 Current-PWP 3. Store the SVR model as 24-PWP model
Training:


Input: fs for each D E Pi Pk ; Testing:
Output: Current PWP model Input: Test data fT e

− Output: Forecasted PWP from 24-PWP model
1. Create training set vector fTr
1. Input fT e to the 24-PWP model


for i = 1 to length fTr do 2. Return Forecasted PWP from the 24-PWP model.
2. Train the SVR model and fine tune the parameters
3. Store the SVR model
end for Algorithm 5 24-PWP to 24-FoS; 24-FoS to Forewarning
4. Compare the accuracy and error for all the SVR models Input: 24-PWP, Soil properties, Constants in FoS equation
5. Select the model with high accuracy and less error as the Output: 24-FoS
Current PWP model
6. Return fTr∗ associated with Current-PWP model Calculate 24-FoS from the FoS equation
if FoS > 1 then
Testing: Slope is Stable
Input: Test data fT e
Output: Predicted PWP from Current PWP model end if
if FoS ≤ 1 then
1. Create testing data fT e in the same format as fTr∗ Slope is Unstable
2. Input fT e to the Current-PWP model end if
3. Return Predicted PWP from the Current-PWP model.
VI. R ESULTS
C. 24-PWP: Forecasting PWP 24 Hours Ahead A. Learning Algorithms to Capture the Causative
Support vector algorithm learns from the historic and real- Relationship of Parameters Leading to Landslides
time data, the data can be used in such a way to forecast future The highest correlation that could be obtained in a linear
PWP values. Forecasting is achieved by, creating a training model for a relation between rainfall and PWP is 0.69. In other
9972 IEEE SENSORS JOURNAL, VOL. 19, NO. 21, NOVEMBER 1, 2019

Fig. 5. Results of Current-PWP algorithm for the PWP values from


piezometer sensor at location DEP-1 and depth 5.5m.

words in a lower dimensional space, the relation between


rainfall and PWP can at the most be explained by only 70%.
Whereas while using the Current-PWP algorithm which uses
SVR, the mapping of rainfall to PWP happens in higher
dimension and the mapping function tries to capture from the
data, the underlying unknown relationship. By doing so the
SVR algorithm can better explain the complex relationship
existing between rainfall and PWP. The r-squared correlation
obtained using SVR algorithm is 0.95. Fig. 5. shows the
results of Current-PWP algorithm for the PWP values from
piezometer sensor at location DEP-1 at a depth of 5.5m. It can
be seen from the Fig.5., the period from June to August is the
monsoon month, and this is the period where the deployment
site receives higher amount of rainfall. Due to rainfall, the soil Fig. 6. Regression curve to derive PWP from moisture values & moisture
moisture increases and soil saturation happens. Once the soil from PWP values.
is saturated, the pressure inside the soil pores increases, which
leads to increase in PWP values with respect to the rainfall
rate. Depending upon the rainfall pattern during the months
of June, July, August, the PWP dynamics varies accordingly. top layer soil in terms of its properties like soil cohesion,
The Current-PWP algorithm captures the generalized trend of permittivity, permeability, etc. Multiple moisture sensors and
the PWP increase and decrease. There are few data point, piezometers are installed in the crown, middle and bottom
that is not accurately reflected in the Current-PWP values as regions of the test bed along with other sensors like geophones
compared to the actual values. This is expected because any and resistivity measuring electrodes. The experimentation was
ML based algorithms tries to learn a generalized trend and conducted continuously for 7 hours by providing artificial
cannot capture momentary anomalies. The mean absolute error rainfall using the rainfall simulator. The moisture, PWP data
obtained between the actual PWP and the PWP values from is acquired along with other sensor data. This data is used
Current-PWP algoritm is 4.2kPa with 95% accuracy. Current- to arrive at polynomial regression equations. The regression
PWP algorithm is used at situations when the real-time data equation to derive PWP from moisture values is PW P =
from the field is not available for various reasons. Since we −5E − 07x 4 + 0.0001x 3 + 0.2714x + 98.259 (here the symbol
have the prior knowledge of the error from the Current-PWP x refers to moisture) and the regression equation to derive
model, we know the uncertainty in the nowcasted results pro- moisture values from PWP values is 21.08y 4 − 8520.8y 3 +
duced from the Current-PWP model. Therefore it is acceptable 1E + 06y 2 − 9E + 07y + 2E + 09 (here the symbol y
to tolerate an error of 4.2 kPa with 95% accuracy. refers to PWP). The polynomial regression graphs are shown
in the Fig. 6. The regression equation shown in this paper
is very specific to the top soil layer in Munnar deployment
B. Laboratory Experiments for Capturing Spatial-Temporal site and it cannot be used for other soil layers in Munnar.
Variability of Parameters The regression equations obtained highly correlates with the
Laboratory experiments were conducted under controlled data points with an R-squared value greater than 0.95. These
environmental conditions in a large scale test bed as shown regression equations can be used to better approximate the
in Fig. 4. to simulate Munnar deployment site. Controlled value of PWP if moisture data and vice versa for Munanrs
environmental conditions include, artificial rain, prepared soil deployment sites top soil layer. By replicating different soil
to replicate the monitoring sites particular soil layer, wind, layers of Munnar deployment site in laboratory and studying
slope angle etc. Results of the experiment conducted to the variations of vital parameters will help in deriving relations
replicate Munnars top soil layer is shown in Fig. 6. in this for the whole deployment site. The regression equations thus
paper. Soil in the test bed is prepared to replicate Munnars obtained for different soil layers in the deployment site can
HEMALATHA et al.: EFFECTIVE AND ACCELERATED FOREWARNING OF LANDSLIDES USING WSN AND MACHINE LEARNING 9973

be used to approximate the value of a sensing parameter from


another available parameter for different soil layer in the site.

C. Reducing False Warnings


False warnings are reduced with the multi-level warnings in
our early warning system. Multilevel warning system stratifies
warnings into 3 likelihood levels. Earlier in the absence of
such a system, warning are delivered when rainfall crosses
threshold. Rainfall threshold corresponds to the lowest level in
landslide warning. Intermediate and immediate level warning
will be much rarer phenomenon compared to rainfall crossing Fig. 7. Current PWP during the monsoon 2018.
threshold. In 2013, Munnar experienced torrential rainfall in
the last week of July and first week of August. Our system was
operational 24/7 and the rainfall-intensity duration thresholds
were crossed on a daily basis, and for 3-day, 7-day and 15-
day durations. Therefore, we issued a first level warning,
which is a regional warning to the regions surrounding the
deployment site. As forecasted, a landslide did happen in
a nearby area, and many other landslides happened in the
Munnar region, which validated our first level warnings issued.
However at the same time, we didn’t not notice any significant
increase in PWP or the decrease in slope stability conditions,
and moreover we did not notice any movements from our Fig. 8. 24-PWP during the monsoon 2018.
movement sensor in our Munnar deployment site. Therefore
neither second level nor third level warnings were issued to
the inhabitants in the expected landslide runoff areas of our
deployment site. In the same way no landslides happened
in the Munnar deployment site, which validated our systems
capability of providing multi-level warnings.

D. Multilevel Reliability of the System During Adverse


Conditions
In Munnar, during the monsoon in 2018, power was inter-
mittent starting from July-2018 and the power grid was
Fig. 9. 24-FoS during the monsoon 2018.
completely shut from August-14-2018 to August-21-2018 for
a period of 7 days, which adversely affected the data transfer
from the field to the Data Management Center in the Uni- IMD’s rainfall forecast information instead of the actual rain-
versity. During adverse scenarios like this, we accomplish the fall data from the rain gauge in the field, for a piezometer
reliability of the system through two means sensor at location DEP-1 and at a depth 5.5m for the complete
1) Reliability Through Intelligent Algorithms: In spite of the monsoon period from June to September 2018. From Fig. 7.
power grid shut-off completely, our system was functioning a break in the actual PWP values can be noticed during the
until Aug-16-2018 with the minimal solar power available. period of June-21 to June23, August-16 to August-21 and from
This is achieved by coupling intelligent energy efficient algo- August-30 to September-9- 2018. It can be seen from Fig. 5.
rithms [50] to the nodes in the WSN system [50].The energy that the Current-PWP algorithm using the rainfall forecast
efficient algorithms are context aware and the sensor nodes information predicts the actual PWP values more accurately
are tailored for making local decisions to optimize the data with a mean absolute error of 4.2 kPa.
collection by using minimal energy thereby extending the life
time of the WSN system [50]. E. Inadequacy of Time for Forewarnings
2) Reliability Through Machine Learning: Since there was In order to forewarn landslides ahead of time, we devised
no power available in the Field Management Center, the data a SVR algorithm “24-PWP” Algorithm-4, which is used for
transfer got affected from August-16-2018 and no data was forecasting PWP data 24 hours ahead of time from rainfall
available until August-21-2018 in the central data management measurements. 24-PWP data can be used to calculate 24-FoS,
centre. During adverse scenarios like this, when the reliability which is the forecasted values of FoS derived from equation
of the system for early warnings is a challenge, Current-PWP 1 and 2. This 24-FoS values from 24-PWP algorithm gives
algorithm is used to nowcast the expected sensor values at extra time of 24 hours which is a boon for the researchers,
various location-depth from the IMD rainfall forecast data. government and public to be prepared for landslides. Fig. 8
Fig. 7. shows the results of Current-PWP algorithm from depicts the results of 24-PWP algorithm for a piezometer
9974 IEEE SENSORS JOURNAL, VOL. 19, NO. 21, NOVEMBER 1, 2019

sensor at location DEP-1 and at a depth of 5.5m, during the [2] M. V. Ramesh, “Design, development, and deployment of a wireless
monsoon season in July through, August-2018. Fig. 9. shows sensor network for detection of landslides,” Ad Hoc Netw., vol. 13,
pp. 2–18, Feb. 2014.
the corresponding FoS values forecasted from the 24-PWP [3] P. K. Mishra, S. K. Shukla, S. Dutta, S. K. Chaulya, and G. M. Prasad,
values for a piezometer sensor at location DEP-1 and at a “Detection of landslide using wireless sensor networks,” in Proc. Int.
depth of 5.5m, during monsoon season in July, August-2018. Conf. Organized, Aug. 2011.
[4] B. T. Pham, B. Pradhan, D. T. Bui, I. Prakash, and M. B. Dholakia,
It can be seen from Fig. 8. that the 24-PWP algorithm captures “A comparative study of different machine learning methods for land-
the generalized trend of the PWP dynamics and it doesnt slide susceptibility assessment: A case study of Uttarakhand area
capture the momentary aberrations, same as that of Current- (India),” Environ. Model. Softw., vol. 84, pp. 240–250, Oct. 2016.
PWP algorithm. From Fig. 9, it can be seen that the same [5] A. M. Fanos, B. Pradhan, S. Mansor, Z. M. Yusoff, and
A. F. B. Abdullah, “A hybrid model using machine learning methods
generalized trend of 24-PWP algorithm is reflected in 24-FoS. and GIS for potential rockfall source identification from airborne laser
24-PWP algorithm forecasts the PWP values 24 hours ahead scanning data,” Landslides, vol. 15, no. 9, pp. 1833–1850, 2018.
of time with a mean absolute error of 4.7kPa with 95% [6] X. Z. Li and J. M. Kong, “Application of GA–SVM method with
parameter optimization for landslide development prediction,” Natural
accuracy. Since this particular study was primarily undertaken Hazards Earth Syst. Sci., vol. 14, pp. 525–533, Mar. 2014.
for forecast of landslides, the authors believe that an error [7] H. R. Pourghasemi, A. G. Jirandeh, B. Pradhan, C. Xu, and
of 4.7 kPa with 95% accuracy are within the acceptable range C. Gokceoglu, “Landslide susceptibility mapping using support vector
machine and GIS at the Golestan Province, Iran,” J. Earth Syst. Sci.,
of tolerance limits. vol. 122, no. 2, pp. 349–369, 2013.
[8] H. Li, Q. Xu, Y. He, and J. Deng, “Prediction of landslide displacement
VII. D ISCUSSION AND F UTURE WORK with an ensemble-based extreme learning machine and copula models,”
With the steady advancements in sensor technology, com- Landslides, vol. 15, no. 10, pp. 2047–2059, 2018.
[9] D. Petley, “Global patterns of loss of life from landslides,” Geology,
munication and computing techniques, it is expected that vol. 40, no. 10, pp. 927–930, 2012.
disasters can be forecasted with higher accuracy than before. [10] M. J. Crozier and T. Glade, “Landslide hazard and risk: Issues, concepts
WSN and now recently IoT have facilitated scalablity, fault tol- and approach,” Landslide Hazard Risk, pp. 1–40, 2005.
erance, effective power consumption and ubiquitous computa- [11] J. Hervás, J. I. Barredo, P. L. Rosin, A. Pasuto, F. Mantovani, and
S. Silvano, “Monitoring landslides from optical remotely sensed
tional abilities in sensor devices and facilitate data availability imagery: The case history of Tessina landslide, Italy,” Geomorphology,
anywhere. These systems integrated with learning algorithms vol. 54, nos. 1–2, pp. 63–75, 2003.
and knowledge discovery of the data from these devices can [12] V. Singhroy, “Satellite remote sensing applications for landslide detec-
tion and monitoring,” in Landslides–Disaster Risk Reduction. Springer,
be of very great use in life saving disaster early warning 2009, pp. 143–158.
applications. In this paper, we proposed an efficient method [13] V. Tofani, S. Segoni, A. Agostini, F. Catani, and N. Casagli, “Use of
for nowcasting the real-time and forecasting the 24-hours remote sensing for landslide studies in Europe,” Natural Hazards Earth
Syst. Sci., vol. 13, no. 2, pp. 299–309, 2013.
ahead conditions of the slope. The key idea of the proposed
[14] A. Abellán, T. Oppikofer, M. Jaboyedoff, N. J. Rosser, M. Lim, and
method is to (i) render the system reliable through learning M. J. Lato, “Terrestrial laser scanning of rock slope instabilities,” Earth
algorithms at times like disaster when data availability is a Surf. Processes Landforms, vol. 39, no. 1, pp. 80–97, 2014.
constraint, (ii) using the learned knowledge as a virtual sensor [15] T. Oppikofer, M. Jaboyedoff, L. Blikra, M.-H. Derron, and R. Metzger,
“Characterization and monitoring of the Åknes rockslide using terres-
and reusing the sensor in other locations for monitoring. As a trial laser scanning,” Natural Hazards Earth Syst. Sci., vol. 9, no. 3,
future work, we would like to implement weightage based pp. 1003–1019, 2009.
learning approach, giving higher weightage to the real-time [16] C. J. Van Westen, E. Castellanos, and S. L. Kuriakose, “Spatial
data for landslide susceptibility, hazard, and vulnerability assessment:
conditions and also implement other learning methods such An overview,” Eng. Geol., vol. 102, nos. 3–4, pp. 112–131, 2008.
as deep learning and comparing those results. [17] A. Akgun, C. Kincal, and B. Pradhan, “Application of remote sensing
data and GIS for landslide risk assessment as an environmental threat to
ACKNOWLEDGMENT Izmir city (west Turkey),” Environ. Monit. Assessment, vol. 184, no. 9,
pp. 5453–5470, 2012.
The authors would like to express their immense gratitude [18] F. Tsai, J.-H. Hwang, L.-C. Chen, and T.-H. Lin, “Post-disaster assess-
to Sri. Mata Amritanandamayi Devi (AMMA), Chancellor, ment of landslides in southern Taiwan after 2009 Typhoon Morakot
Amrita Vishwa Vidyapeetham, who gave us the motivation using remote sensing and spatial analysis,” Natural Hazards Earth Syst.
Sci., vol. 10, no. 10, pp. 2179–2190, 2010.
and inspiration to pursue this research work. We undertook
[19] M. Pieraccini et al., “Landslide monitoring by ground-based radar
this work following our recognition as the “World Center of interferometry: A field test in Valdarno (Italy),” Int. J. Remote Sens.,
Excellence in landslide disaster risk reduction”, conferred to vol. 24, no. 6, pp. 1385–1391, 2003.
us by the IPL- “International Programme on Landslides” in [20] M. G. Raines et al., “The application of passive seismic techniques to
the detection of buried hollows,” Tech. Rep., 2015.
August 2017. The authors would also like to acknowledge [21] L. Pellerin, “Applications of electrical and electromagnetic methods for
the contributions of Prof. Balaji Hariharan for his valuable environmental and geotechnical investigations,” Surv. Geophys., vol. 23,
suggestions. We would also like to acknowledge the contribu- nos. 2–3, pp. 101–132, 2002.
[22] J. Busby, R. Cuss, M. Raines, and D. Beamish, “Application of ground
tions of Dr. M.R Kaimal, Dr. P.V. Usha kumari, Dr. Nirmala penetrating radar to geological investigations,” Tech. Rep., 2004.
Vasudevan, Ms. Divya P, Ms. Geethu TH, and the entire [23] J. E. Chambers et al., “Three-dimensional geophysical anatomy of
landslide team in our research center for their support in an active landslide in lias group mudrocks, Cleveland Basin, UK,”
various aspects. Geomorphology, vol. 125, no. 4, pp. 472–484, 2011.
[24] T. Hermans, D. Caterina, R. Martin, A. Kemna, and F. Nguyen,
R EFERENCES “Incorporation of prior information in the regularized inversion of ERT
data with CRTomo,” Tech. Rep., 2012.
[1] N. A. A. Aziz and K. A. Aziz, “Managing disaster with wireless sensor [25] G. Mentes, “Borehole wire extensometer for measurement of small
networks,” in Proc. 13th Int. Conf. Adv. Commun. Technol. (ICACT), displacements,” in Proc. 5th Int. Conf. Eng. Surveying Univ. Zagreb,
Feb. 2011, pp. 202–207. Fac. Geodesy Zagreb, Croatia, 2011, pp. 307–314.
HEMALATHA et al.: EFFECTIVE AND ACCELERATED FOREWARNING OF LANDSLIDES USING WSN AND MACHINE LEARNING 9975

[26] L. Simeoni and L. Mongiovì, “Inclinometer monitoring of the Castelrotto [49] S. Mukherjee, S. Aadhar, D. Stone, and V. Mishra, “Increase in extreme
landslide in Italy,” J. Geotech. Geoenviron. Eng., vol. 133, no. 6, precipitation events under anthropogenic warming in India,” Weather
pp. 653–666, 2007. Climate Extremes, vol. 20, pp. 45–53, Jun. 2018.
[27] A.-B. Huang, J.-T. Lee, Y.-T. Ho, Y.-F. Chiu, and S.-Y. Cheng, “ Stability [50] R. Prabha, M. V. Ramesh, V. P. Rangan, P. V. Ushakumari, and
monitoring of rainfall-induced deep landslides through pore pressure T. Hemalatha, “Energy efficient data acquisition techniques using context
profile measurements,” Soils Found., vol. 52, no. 4, pp. 737–747, 2012. aware sensing for landslide monitoring systems,” IEEE Sensors J.,
[28] F. Bardi et al., “Integration between ground based and satellite SAR data vol. 17, no. 18, pp. 6006–6018, Sep. 2017.
in landslide mapping: The San Fratello case study,” Geomorphology,
vol. 223, pp. 45–60, Oct. 2014.
[29] P. Savvaidis, “Existing landslide monitoring systems and techniques,”
From Stars Earth Culture, pp. 242–258, 2003.
[30] F. Karsli, A. Yalcin, M. Atasoy, O. Demir, S. Reis, and E. Ayhan, T. Hemalatha received the M.Tech. degree in
“Landslide assessment by using digital photogrammetric techniques,” in remote sensing and wireless sensor networks from
Proc. 20th ISPRS Congr. Commission VII Turkey, 2004, pp. 736–739. Amrita Vishwa Vidyapeetham University, where she
[31] N. Caine, “The rainfall intensity-duration control of shallow landslides is currently pursuing the Ph.D. degree in data ana-
and debris flows,” Geografiska Annaler Ser. A, Phys. Geography, vol. 62, lytics for wireless sensor network with the Amrita
nos. 1–2, pp. 23–27, 1980. Center for Wireless Networks and Applications.
[32] F. Guzzetti, S. Peruccacci, M. Rossi, and C. P. Stark, “The rain- She is also a Research Associate with the Amrita
fall intensity-duration control of shallow landslides and debris flows: Center for Wireless Networks and Applications,
An update,” Landslides, vol. 5, no. 1, pp. 3–17, 2008. Amrita Vishwa Vidyapeetham University, where she
[33] D. Lagomarsino et al., “Quantitative comparison between two different is involved in the research works for forecasting
methodologies to define rainfall thresholds for landslide forecasting,” imminent landslides and the decision support system
Natural Hazards Earth Syst. Sci., vol. 15, no. 10, pp. 2413–2423, 2015. work for monitoring and early warning of landslides in parts of Western Ghats
[34] L. Brocca, F. Ponziani, T. Moramarco, F. Melone, N. Berni, and and Himalayas. Her major focus is on arriving at thresholds for different
W. Wagner, “Improving landslide forecasting using ASCAT-derived soil sensors used in landslide forecasting, efficiently forecasting the possibility for
moisture data: A case study of the Torgiovannetto landslide in central imminent landslides, and risk analysis for landslides.
Italy,” Remote Sens., vol. 4, no. 5, pp. 1232–1244, 2012.
[35] M. V. Ramesh and N. Vasudevan, “The deployment of deep-earth sensor
probes for landslide detection,” Landslides, vol. 9, no. 4, pp. 457–474,
Dec. 2012. doi: 10.1007/s10346-011-0300-x.
[36] T. Hemalatha, M. V. Ramesh, and V. P. Rangan, “Adaptive learning Maneesha Vinodini Ramesh received the Ph.D.
techniques for landslide forecasting and the validation in a real world degree in computer science and engineering from
deployment,” in Proc. Workshop World Landslide Forum. Springer, 2017, Amrita Vishwa Vidyapeetham University in 2009.
pp. 439–447. She currently serves as the Director of the Amrita
[37] S. L. Kuriakose, G. Sankar, and C. Muraleedharan, “History of landslide Center for Wireless Networks and Applications,
susceptibility and a chorology of landslide-prone areas in the Western Amrita Vishwa Vidyapeetham University, where she
Ghats of Kerala, India,” Environ. Geol., vol. 57, no. 7, pp. 1553–1568, is currently a Professor of Computer Science and
2009. Engineering. She is also a Co-Principal Investiga-
[38] S. L. Kuriakose, V. G. Jetten, C. J. Van Westen, G. Sankar, and tor of the European Commission funded Wireless
L. P. H. Van Beek, “Pore water pressure as a trigger of shallow landslides Sensor Networks with Self Organization Capabilities
in the Western Ghats of Kerala, India: Some preliminary observations for Critical and Emergency Applications (WINSOC)
from an experimental catchment,” Phys. Geography, vol. 29, no. 4, Project, and a principal investigator of eight internationally recognized projects
pp. 374–386, 2008. funded by different organizations from all over the world. She has more than
[39] H. Vijith and G. Madhu, “Estimating potential landslide sites of an 100 papers, including several journals and best paper awards. She was a
upland sub-watershed in Western Ghat’s of Kerala (India) through recipient of the NABARD Award for Rural Innovation—Second Prize from the
frequency ratio and GIS,” Environ. Geol., vol. 55, no. 7, pp. 1397–1405, Honorable Finance Minister, Government of India, in 2012, for her research
2008. activities benefited to the rural community. She is an Editor of the Ad Hoc
Networks Elsevier. She has given invited talks at several eminent universities
[40] R. Bhasin et al., “Landslide hazards and mitigation measures at Gangtok,
all over the world.
Sikkim Himalaya,” Eng. Geol., vol. 64, no. 4, pp. 351–368, 2002.
[41] N. Vasudevan and K. Ramanathan, “Geological factors contributing
to landslides: Case studies of a few landslides in different regions of
India,” in Proc. IOP Conf. Ser. Earth Environ. Sci., vol. 30, no. 1, 2016,
Art. no. 012011.
Venkat P. Rangan founded and directed the Mul-
[42] R. M. Iverson, “Landslide triggering by rain infiltration,” Water Resour.
timedia Laboratory and the Internet and Wireless
Res., vol. 36, no. 7, pp. 1897–1910, Jan. 2000.
Networks (Wi-Fi) Research, University of California
[43] R. Greco, A. Guida, E. Damiano, and L. Olivares, “Soil water con-
at San Diego, where he served as a Professor of
tent and suction monitoring in model slopes for shallow flowslides
Computer Science and Engineering for 16 years. He
early warning applications,” Phys. Chem. Earth Parts A/B/C, vol. 35,
is currently the Vice Chancellor of Amrita Vishwa
nos. 3–5, pp. 127–136, 2010.
Vidyapeetham University. He has more than 85 pub-
[44] B. Thiebes et al., “Integration of a limit-equilibrium model into lications in international (mainly the IEEE and the
a landslide early warning system,” Landslides, vol. 11, no. 5, ACM) journals and conferences, and also holds
pp. 859–875, 2014. 22 U.S. Patents. He is a Fellow of the ACM (1998).
[45] M. V. Ramesh, D. Pullarkatt, T. H. Geethu, and P. V. Rangan, “Wireless He is the youngest to achieve this international
sensor networks for early warning of landslides: Experiences from a distinction. He was a recipient of the President of the India Gold Medal
decade long deployment,” in Proc. Workshop World Landslide Forum. in 1984, the NCR Research Innovation Award in 1991, and the NSF National
Springer, 2017, pp. 41–50. Young Investigator Award in 1993. In 2000, Internet World featured him
[46] P. Aleotti, “A warning system for rainfall-induced shallow failures,” Eng. on its cover page and named him as one of the top 25 Stars of Internet
Geol., vol. 73, nos. 3–4, pp. 247–265, 2004. Technologies. In 2012, Silicon India ranked him as one of the 50 Indians
[47] G. B. Crosta and P. Frattini, “Rainfall thresholds for triggering soil slips Who Redefined Entrepreneurship in the Last 65 Years of Independence. He is
and debris flow,” in Proc. 2nd EGS Plinius Conf. Medit. Storms, Siena, an internationally recognized pioneer of research in multimedia systems and
Italy, A. Mugnai, F. Guzzetti, and G. Roth, Eds. 2001, pp. 463–487. Internet E-Commerce. In 1993, he founded the first International Conference
[48] M. Rinaldi, N. Casagli, S. Dapporto, and A. Gargini, “Monitoring on Multimedia: ACM Multimedia 93, for which he was the Program Chair-
and modelling of pore water pressure changes and riverbank stability man. This is now the premier world-wide conference on multimedia. He also
during flow events,” Earth Surf. Processes Landforms, vol. 29, no. 2, founded the first International Journal on Multimedia: ACM/Springer-Verlag
pp. 237–254, 2004. Multimedia Systems, which is now the premier journal on multimedia.

You might also like