Content uploaded by Samiul Hasan
Author content
All content in this area was uploaded by Samiul Hasan on Mar 23, 2021
Content may be subject to copyright.
[This is the accepted version of the following article: Roy, KC, Hasan, S, Mozumder, P. A multilabel classification approach to
identify hurricane‐induced infrastructure disruptions using social media data. Comput Aided Civ Inf. 2020; 35: 1387– 1402,
which has been published in final form at https://doi.org/10.1111/mice.12573. This article may be used for non-commercial
purposes in accordance with the Wiley Self-Archiving Policy http://www.wileyauthors.com/self-archiving].
A Multi-label Classification Approach to Identify
Hurricane-induced Infrastructure Disruptions Using
Social Media Data
Kamol Chandra Roy, Samiul Hasan*
Department of Civil, Environmental, and Construction Engineering, University of Central Florida, Orlando,
Florida
&
Pallab Mozumder
Department of Earth & Environment and Department of Economics, Florida International University, Miami,
Florida.
Abstract: Rapid identification of infrastructure disruptions
during a disaster plays an important role in restoration and
recovery operations. Due to the limitations of using physical
sensing technologies, such as the requirement to cover a large
area in a short period of time, studies have investigated the
potential of social sensing for damage/disruption assessment
following a disaster. However, previous studies focused on
identifying whether a social media post is damage related or
not. Hence, advanced methods are needed to infer actual
infrastructure disruptions and their locations from such data.
In this paper, we present a multi-label classification approach
to identify the co-occurrence of multiple types of
infrastructure disruptions considering the sentiment towards
a disruption—whether a post is reporting an actual disruption
(negative), or a disruption in general (neutral), or not affected
by a disruption (positive). In addition, we propose a dynamic
mapping framework for visualizing infrastructure disruptions.
We use a geo-parsing method that extracts location from the
texts of a social media post. We test the proposed approach
using Twitter data collected during hurricanes Irma and
Michael. The proposed multi-label classification approach
performs better than a baseline method (using simple keyword
search and sentiment analysis). We also find that disruption
related tweets, based on specific keywords, do not necessarily
indicate an actual disruption. Many tweets represent general
conversations, concerns about a potential disruption, and
positive emotion for not being affected by any disruption. In
addition, a dynamic disruption map has potential in showing
county and point/coordinate level disruptions. Identifying
disruption types and their locations are vital for disaster
recovery, response, and relief actions. By inferring the co-
occurrence of multiple disruptions, the proposed approach
may help coordinate among infrastructure service providers
and disaster management organizations.
1 INTRODUCTION
Cities and communities all over the world largely depend
on critical infrastructure systems/services such as electrical
power, water distribution, communication services, and
transportation networks. The growing interconnectedness
and interdependency among these systems have changed the
organizational and operational factors and increased the
vulnerability in the face of unwanted disruptions. These
systems provide critical services to a large population, and
thus when disrupted they affect our quality of life, local and
regional economy, and the overall community well-being.
The need to quickly identify disaster-induced infrastructure
disruptions is growing because of the increasing number of
natural disasters such as hurricane Michael, Irma, Harvey,
and Florence and their enormous impacts to affected
communities.
For instance, hurricane Irma caused a substantial number
of power outages in addition to transportation,
communication, drinking water, and wastewater related
disruptions. More than six million customers faced power
outages during Irma. Storm related high winds and sustained
storm surges cost approximately 3,300 megawatts of power
generation (NERC, 2018). Around 27.4% of cell phone
towers in Florida were damaged due to hurricane Irma as
reported by the Federal Communications Commission (FCC)
(FCC, 2017). Irma caused flooding to several areas
throughout Florida, forcing health officials to issue unsafe
Roy et al.
2
drinking water and boiling water notices (“The Effect of
Hurricane Irma on Water Supply,” 2017; “Unsafe Drinking
Water After Hurricane Irma,” 2019). Moreover, dozens of
sewage systems were overflowed after the power went out,
which further exacerbated the drinking water condition
(“Unsafe Drinking Water After Hurricane Irma,” 2019).
To ensure efficient operation and maintenance, it is
important to gather real-time information about the
performance and integrity of engineering systems. This is
typically performed through a computational monitoring
process that involves observation of a system, analysis of the
obtained data, and prediction of future performance (Dyskin
et al., 2018). During a disaster, due to disruptions, the
performance of critical infrastructures degrades rapidly—
leading to cascading failures (Du, Cai, Sun, & Minsker, 2017;
Kadri, Birregah, & Châtelet, 2014). In such extreme events,
computational monitoring is required to assess the quickly
changing condition of infrastructure systems and warn about
an approaching failure or even a catastrophic event.
For effective disaster response and recovery operations,
coordinated actions are required from the responsible
organizations. Disruptions to infrastructure systems such as
electricity/power, cell phone, internet, water, waste water,
and other systems significantly affect the recovery time of a
community (Mitsova, Escaleras, Sapat, Esnard, & Lamadrid,
2019). Due to the interdependence among infrastructure
systems, multiple types of disruptions (e.g., power outages,
internet/cell phones, water service) are likely to co-occur
during a disaster. To ensure an expedited recovery of the
systems, rapid identification of the co-occurrence of
disruptions is necessary so that coordinated actions can be
taken by multiple agencies.
Although infrastructure performance data can be collected
through physical sensing technologies such as drones,
satellite, UAV etc. (Jongman, Wagemaker, Romero, & de
Perez, 2015; NERC, 2018), they might not be feasible due to
the rapidly evolving nature of a disaster spreading over a
large area (Fan, Mostafavi, Gupta, & Zhang, 2018). Social
media users have been used as sensors during disasters and
several studies have found its potential for understanding
situational awareness (Huang & Xiao, 2015; Kryvasheyeu Y,
Chen H, Moro E, Van Hentenryck P, 2015). Previous studies
investigated social media sensing for damage assessment
(Kryvasheyeu et al., 2016), recovery (Z. Li, Wang, Emrich,
& Guo, 2018), and inundation mapping (Jongman et al.,
2015). Studies have also proposed query based approaches to
identify topics related to critical infrastructure disruptions
(Fan & Mostafavi, 2019; Fan et al., 2018). However, these
studies have not considered the co-occurrences of the types
and extent of infrastructure disruptions.
During an unfolding disaster, people from the affected
regions share their opinions, views, concerns, and eye
witnessed events in social media platforms. Such user-
generated content can provide valuable information to extract
disruption-related information. However, during a disaster,
emergency managers face challenges to monitor the massive
volume of social media posts in real time (Oyeniyi, 2017).
Thus, to get actionable information, it is important to identify
whether a post indicates an actual disruption or simply
expresses user views or opinions about a disruption. Recent
studies have mainly focused on identifying whether a
particular social media post is damage related or not (Yuan &
Liu, 2018, 2019). However, since infrastructure systems are
more interconnected, co-occurrences of disruptions in
multiple infrastructures are more likely.
In this study, we develop a multi-label classification
approach to identify the co-occurrence and extent of multiple
types of infrastructure disruptions. We also present a
framework to create dynamic disruption maps and case
studies showing the developed approach based on Twitter
data collected during hurricanes Irma and Michael. This
study has the following contributions:
• We consider multiple types of infrastructure disruptions
(e.g., power, transportation, water, wastewater, and other
disruption) and their co-occurrences in a social media
post, instead of considering a simple binary classification
problem (i.e., whether a post is disruption related or not).
• To identify if a disruption related post reflects an actual
disruption, we associate sentiments with disruption
status—whether a post is reporting an actual disruption
(negative), or disruption in general (neutral), or not
affected by a disruption (positive).
• We propose a dynamic mapping framework for
visualizing infrastructure disruptions by adopting a geo-
parsing method that extracts location from tweet texts.
Instead of identifying disruption types and status in a single
label, we identify disruption types and disruption status
(through sentiment) separately. We adopt this approach since
the neutral and positive sentiment about a disruption may also
provide valuable information on the level of situational
awareness about disruptions during a disaster.
2 LITERATURE REVIEW
According to the Department of Homeland Security, there
are 16 critical infrastructure sectors (Homeland Security,
2019). Among these sectors, energy, communication,
transportation, water/wastewater systems are the most
vulnerable ones to a natural disaster. It is important to
identify, characterize, and model infrastructure disruptions
for a faster restoration and recovery operation (Fang &
Sansavini, 2019; Sriram, Ulak, Ozguven, & Arghandeh,
2019). Studies have focused on the recovery plans and
damages due to extreme weather events (Bryson, Millar,
Joseph, & Mobolurin, 2002; Fang & Sansavini, 2019;
Lambert & Patterson, 2002; Rosenzweig & Solecki, 2014;
Sörensen, Webster, & Roggman, 2002). Several studies have
proposed approaches to assess the reliability, resilience,
vulnerability and failure process of power, transportation, and
water supply networks individually (Barabási & Albert,
1999; Buldyrev, Parshani, Paul, Stanley, & Havlin, 2010;
Jenelius & Mattsson, 2012; Ouyang & Fang, 2017; Pietrucha-
Urbanik & Tchórzewska-Cieślak, 2018; Sumalee &
Kurauchi, 2006; Ulak, Kocatepe, Sriram, Ozguven, &
Arghandeh, 2018). However, these critical infrastructures are
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
3
inter-connected and inter-dependent (Alinizzi, Chen, Labi, &
Kandil, 2018; Homeland Security, 2019; Martani, Jin, Soga,
& Scholtes, 2016). Considering the increased connectedness
and interdependencies among infrastructure systems, studies
have proposed a holistic approach to assess the resilience to
disruptions (Hasan & Foliente, 2015; Lu et al., 2018; Pant,
Thacker, Hall, Alderson, & Barr, 2018; Rinaldi, Peerenboom,
& Kelly, 2001; Sriram et al., 2019). However, most of these
studies are based on synthetic data or post-event data. Thus,
they are not suitable for real-time decision-making.
Recently, real-time condition monitoring is becoming very
popular in manufacturing, maintenance, and usage of many
engineering systems (Dyskin et al., 2018) and civil
engineering infrastructures (Chang, Flatau, & Liu, 2003).
Computational models have been developed for estimating
the properties of constructional materials (Rafiei, Khushefati,
Demirboga, & Adeli, 2017), detecting damages to building
structures (Rafiei & Adeli, 2017b, 2018a), predicting
construction costs (Rafiei & Adeli, 2018b) etc. Another
potential approach for monitoring infrastructures is by
collecting real-time data using smartphones, leading to
citizen-centered and scalable monitoring systems in a disaster
context (Alavi & Buttlar, 2019).
During an ongoing disaster and post-disaster period, it is
important to collect disruption data to take necessary actions
as fast as possible. Due to the intensity and spread of a
disaster, physical sensing techniques such as satellite, UAV
(Unmanned Aerial Vehicle) etc. (Jongman et al., 2015;
NERC, 2018) are not suitable. For example, after hurricane
Irma, unmanned aerial drones, amphibious vehicles, airboats
are used to perform damage assessment on inaccessible
transmission and distribution lines (NERC, 2018). A crowd-
sourcing app that allows damage reporting might not be
useful because of fewer participants. On the other hand, the
ubiquitous use of social media on GPS enabled smartphone
device, allows us to collect large-scale user generated data
containing live and in situ events during a disaster
(Middleton, Middleton, & Modafferi, 2013). Studies have
already used social media data for crisis mapping (Birregah
et al., 2012; Gao, Barbier, & Goolsby, 2011; Middleton et al.,
2013). However, real-time crisis mapping requires location
information, but only around 1% to 4% of social media (e.g.,
Twitter) data posts are geo-tagged (Cheng, Caverlee, & Lee,
2010; C. Li & Sun, 2014; Middleton et al., 2013). Studies
have proposed several location-extraction methods from
content/textual data (Cheng et al., 2010; C. Li & Sun, 2014;
Middleton et al., 2013). In addition, the power of social media
to connect a large group of population has drawn significant
attention towards using social media platforms for disaster
management (Keim & Noji, 2010; Sadri, Hasan, & Ukkusuri,
2019; Tang, Zhang, Xu, & Vo, 2015). Studies have analyzed
social media data for understanding human mobility and
resilience during a disaster (Roy, Cebrian, & Hasan, 2019;
Wang & Taylor, 2014). Kryvasheyeu et al. proposed that
social media users can be considered as early warning sensors
in detecting and locating disasters (Kryvasheyeu Y, Chen H,
Moro E, Van Hentenryck P, 2015). Studies have also
explored social media data to understand evacuation behavior
(Fry & Binner, 2015; Martín, Li, & Cutter, 2017) and damage
assessment (Deng, Liu, Zhang, Deng, & Ma, 2016; Guan &
Chen, 2014; Kryvasheyeu et al., 2016; Yuan & Liu, 2018).
Damage assessment plays a vital role in resource allocation
and coordination in disaster response and recovery efforts.
Previous studies found that affected people provide damage
related situational updates in social media (Deng et al., 2016;
Guan & Chen, 2014; Kryvasheyeu et al., 2016; Yuan & Liu,
2018). However, these studies do not consider the types of
disruptions and are mainly suitable for post-disaster overall
damage assessment. Most of these studies adopted simpler
indicators of damage assessment such as frequency of
disaster related tweets (based on keywords such as ‘sandy’,
‘hurricane sandy’, ‘damage’). The limitation of using pre-
defined keywords is that a large number of such tweets/texts
may not contain any damage related information. Some
studies (Kotsiantis, Zaharakis, & Pintelas, 2007) adopted
supervised machine learning based classification approaches
to resolve this limitation. These studies (Cresci, Cimino,
Dell’Orletta, & Tesconi, 2015; Yuan & Liu, 2019) adopted
support vector machine, naïve Bayes, decision tree
classification algorithms to analyze damage related social
media posts. However, these studies considered damage
identification as a binary (damage related or not)
classification problem, which may include posts that are not
reporting an actual damage/disruption. In addition, deep
learning models were used for image and text data
(Mouzannar, Rizk, & Awad, 2018; D. T. Nguyen, Ofli,
Imran, & Mitra, 2017). Image data are limited,
computationally expensive, and cannot report disruptions in
functionality such as power outage, communication
disruptions etc.
The most relevant studies towards identifying an
infrastructure disruption using social media posts are
proposed by Fan et al. (Fan & Mostafavi, 2019; Fan et al.,
2018). The first study (Fan et al., 2018) has focused on
summarizing the overall topics during a disaster given some
predefined keywords, not suitable to identify disruptions
from real-time data. In the second study (Fan & Mostafavi,
2019), the authors have developed a graph-based method to
identify situational information related to infrastructure
disruptions by detecting time slices based on a threshold
number of tweets. They compute content similarity within the
detected time slices to get credible information. Some
limitations of this approach include: it depends on keyword
based filtering, which can miss out important information if
appropriate keywords are not chosen; it requires the whole
dataset as an input, which is not suitable for a real-time
prediction; it considers the content posted only on the burst
timeframe that might miss out some actual disruption related
posts. Moreover, this study does not consider that a single
post may have information about multiple types of
disruptions and cannot distinguish if a particular post is
reporting an actual disruption or not.
In summary, to the best of our knowledge, currently no
study exists to identify the co-occurrence of multiple types of
infrastructure disruptions using social media data. For this
Roy et al.
4
task, a multi-label classification approach (Sorower, 2010)
identifying multiple labels from a single input, can be useful.
In this study, we use a multi-label text classification
approach to identify multiple disruption types and their status
using social media data. To develop our multi-label
disruption classification approach, we use eight well-known
models on text classification. We present two case studies to
identify disruptions using Twitter data from hurricanes Irma
and Michael. Finally, we visualize the spatio-temporal
dynamics of infrastructure disruptions in a map of the
affected regions.
3 DATA PREPARATION
In this study, we use Twitter data collected during
hurricanes Irma and Michael for creating a dynamic
disruption map of critical infrastructure disruptions. We use
two different methods (Twitter streaming API and rest API)
for data collection. A brief description of the data is provided
in Table 1.
Using the streaming API, we collected about 1.81 million
tweets posted by 248,763 users between September 5, 2017
and September 14, 2017 during hurricane Irma. We collected
the tweets using a bounding box covering Florida, Georgia,
and South Carolina. To collect data for more time span and
to fill some missing values contained in the steaming API
data, we used Twitter’s rest API to gather user-specific
historical data. Twitter’s rest API allows collecting the most
recent 3,200 tweets of a given user. We collected user-
specific data for 19,000 users, who were active for at least
three days within the streaming data collection period.
Similarly, we collected data for hurricane Michael using a
bounding box covering Florida, Georgia, South Carolina, and
North Carolina, containing 3.53 million tweets posted by 1.29
million users covering from October 8, 2018 to October 18,
2018.
To create an annotated disruption dataset, we manually
labeled 1,127 tweets from hurricane Irma and 338 tweets (for
testing purpose only) from hurricane Michael. The tweets
were labeled by 5 human annotators. To ensure that we
retrieve the right labels of the disruption types and
sentiments, we only considered the labels when all 5
annotators agreed on it. Each tweet can have one or more
labels out of the ten possible labels including: not hurricane
related, power/electricity disruption, communication
disruption, road/transportation disruption, drinking water
disruption, wastewater related disruption, other disruption,
positive, negative, and neutral. The first label indicates
whether a tweet is hurricane related or not. The next five
labels indicate five types of infrastructure disruptions. The
label, other disruption, indicates a disruption that does not
fall into the five types of infrastructure disruptions considered
here. The last three labels indicate the possible sentiment
towards a disruption. We give below three examples of
disruption related tweets:
• This tweet -“Update I'm the only community in my area
with power I feel really lucky right now but I hope
everyone else is safe”- mentions about power/electricity
disruption but in a positive way. We would label such a
tweet as (power/electricity disruption, positive).
• This tweet- “we are in Clermont on Lake Minnehaha. We
have no cable or power & cell service is spotty. When
will be the worst here”- mentions about both
power/electricity and communication disruptions. We
would label this tweet as (power/electricity disruption,
communication disruption, negative).
• This tweet -“im trying to eat and watch as much netflix
as i can just incase my power go out”- mentions about
power/electricity disruption but does not indicate an
actual disruption. We would label it as (power/electricity
disruption, neutral).
Figure 1 shows the frequencies and co-occurrences of the
labels in the annotated dataset. It shows that the annotated
data contain many “not hurricane related” tweets. Among
the tweets related to different types of disruptions,
power/electricity related disruptions have the highest
frequency. Among the sentiment related labels, negative
sentiment has the highest frequency. On the other hand,
power/electricity disruption and negative sentiment are the
most frequently co-occurred labels in the annotated dataset.
Table 1 Data Description
Hurricane
Name
Regions
(USA)
No. of
Tweets
No. of
Users
Irma
(Streaming
API)
FL, GA,
SC
1,810,000
248,763
Irma
(Rest API)
2,478,383
16,399
Michael
(Streaming
API)
FL, GA,
SC, NC
3,534,524
1,289,204
Figure 1 Distribution of Label Frequency and Label Co-
occurrence Frequency
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
5
4 METHODOLOGICAL APPROACH
The methodological approach adopted in this study has
three main parts. The first part takes tweet texts as input and
identifies disruptions and the sentiment towards the
disruption. The second part extracts the geo-location from the
tweet’s metadata or text. The third part visualizes the
disruptions geographically in a dynamic map of disruptions.
Figure 2 shows the steps and information flow among those
steps. Each part of the framework is described below:
4.1 Disruption Identification
The objective of this step is to identify infrastructure
disruptions and sentiments from a given text input, where
more than one disruption type might be present. We use a
supervised multi-label classification approach. The input
texts collected from Twitter posts contain many noises, which
may degrade classification performance. Therefore, we
process the data before feeding it into the model. The
sequential steps are shown in Figure 2 (left side).
Data Pre-processing
In this step, we discard the unnecessary tweets and remove
noise from a tweet. Since a retweet (starting with RT in the
texts) does not provide any new information in the dynamics
of disruption, we discard retweets from the data to avoid false
spike in the disruption count. To clean the tweet texts, we
remove the stop words (e.g., ‘a’, ‘an’ and ‘the’), short URLs,
emoticons, user mentions (@Twitter user name),
punctuations, and special characters (\@/#$ etc.). Finally, we
tokenize (splitting texts into words) the texts and apply
lemmatization (converting the words into noun, verb etc.) and
stemming (converting words into root form) to the tokens.
Data Processing
In this step, we process the data for training models and
predicting disruptions. In machine learning, training of a
model refers to providing it with training data, which contains
both inputs and correct answers, so that the algorithm can find
the pattern to map the input features to the target/output
features. We convert the preprocessed tokens as TF-IDF
(Term Frequency-Inverse Document Frequency), which
measures the importance of a word in a document of a corpus
(collection of documents). The details on TF-IDF can be
found in this study (Ramos & others, 2003). The TF-IDF of a
term/word ( is calculated as follows:
⁃
(1)
where,
We create the TF-IDF using both unigram and bigram of
words. We remove the features that appear in less than 2
documents. To remove the effect of total word counts in a
document, we apply normalization (sum of the squared
value of TF-IDF =1 for a document). To prevent data leakage,
we calculate the TF-IDF considering the tweets available in
the training dataset. The output of the model may contain
multiple disruptions; thus we convert the annotated labels
into multi-label formats. We represent the multi-label output
as a binary/one hot encoded matrix indicating the presence of
disruption type and the sentiment label. In our study, we have
10 possible labels, so, each converted label is represented as
binary matrix where the value 1 represents the
presence and the value 0 represents the absence of a particular
label.
Model Selection
The objective of this step is to find the best model that
maps an input tweet text to the binary matrix representing one
or more types of infrastructure disruptions and sentiment. In
our study, we choose a multi-label classification approach for
identifying disruptions and sentiments. This approach
generalizes the multiclass classification, where a single
input/tweet can be assigned to multiple types of disruptions.
Let be the set of labels containing disruption types
and sentiment, where, . In our case, .
The objective of our disruption identification model, is that:
given the input tweet, the model has to predict the
disruption types and sentiments,
(2)
We apply three methods that allow using the multiclass
classification models for a multi-label classification task. The
first method transforms a multi-label classification into
multiple binary classification problems. This method is also
known as binary relevance (BR) (Sorower, 2010) that trains
one binary classifier for each label independently. The
equation for a binary classifier, for a label can be
expressed as below:
(3)
The BR method transforms the training data into
datasets. The dataset for label contains all the original
dataset labeled as if the original example contains ,
otherwise, as . For an unseen sample, the combined
model predicts all labels using the respective classifier. One
of the disadvantages of the BR method is that it does not
consider the correlation between labels.
The second method transforms the multi-label
classification problem into a multi-class classification
problem. This method is known as label powerset (LP) that
considers each subset of as a single label. Let, be the
powerset of , which contains all possible subset of . LP
method considers each element of as a single label.
Now, in training LP learns one single label classifier ,
where:
(4)
The LP method has advantages over the BR method,
because it takes the label correlations into account. However,
it requires high computation time if the size of is very
Roy et al.
6
big and majority of the subsets have very few members. Also,
the LP method tends to overfit (performs well on training data
but performs poorly on test data), when the number of labeled
samples of the generated subsets is low.
As the third method, we apply an ensemble technique,
known as Random k-Labelsets (RAKEL) adopted from the
study (Tsoumakas & Vlahavas, 2007). This method
constructs an ensemble of LP classifiers, where each LP
classifier is trained on a small random subset of labels.
Instead of using , it creates k-labelset , where
. If the set of all distinct -labelset is , then
.
Given a user specified integer value for and , where,
and , the RAKEL algorithm iteratively
constructs an ensemble of numbers of LP classifiers.
However, for and , RAKEL method becomes
a binary classifier ensemble of BR method. On the other
hand, for , becomes , and consequently, RAKEL
method becomes a single label classifier of the LP method.
Given a meaningful parameter of ( , at each
iteration, without replacement it randomly
selects a k-labelset, from and learns an LP classifier, .
Where,
)
(5)
For a given input, the label prediction is accomplished by
a voting scheme from the ensemble combination. The
RAKEL method solves the overfitting problem of the LP
method but loses some correlations as it considers a random
subset of the labels (LP method considers all possible
Figure 2 Methodological framework: disruption identification module (left); geo-parsing module (right);
and visualization module (middle)
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
7
subsets). The full description of the RAKEL method can be
found in this study (Tsoumakas & Vlahavas, 2007).
In multi-label classification, a prediction cannot be
assigned as a hard right or wrong value, because a prediction
containing a subset of the actual classes should be considered
better than a prediction that contains none of them. Thus,
traditional performance metrics (e.g., precision, recall) are
not suitable for evaluating our disruption identification
model. We choose the best model based on three generally
used performance metrics in multi-label classification: subset
accuracy, micro F1 score, and hamming loss. Here, subset
accuracy and hamming loss are example-based metrics and
micro F1 measure is a label-based metric. For each test
sample, an example-based metric computes the difference
between true and predicted class labels and then calculate the
average over all test samples. Whereas, a label-based metric
first computes the performance for each class label, and then
calculates the average over all class labels. Assuming as the
set of true class labels, as the predicted set of labels, as
the set of labels, as the subset of with label , the
subset of with label , the number of samples, the
equations of these metrics are given below:
(6)
(7)
(8)
We further check the predictive performance of the model
computing a confusion matrix for each label (representing
disruption types and sentiment). Table 2 shows the
components of a confusion matrix. The rows represent the
actual labels and the columns represent the predicted labels
where positive means the existence of a particular label and
negative means the absence of a particular label. For a
particular sample, if the actual label is negative, a negative
prediction by the model is assigned as true negative and a
positive prediction is assigned as false positive. Similarly, if
the actual label is positive, a positive prediction is assigned
as true positive and a negative prediction is assigned as false
negative.
4.2 Disruption Location Extraction
The objective of this step is to extract the location of the
disruptions that are identified by the previous step. Geo-
tagged tweets provide location information either as a point
type (exact latitude-longitude) or as a polygon type
(bounding box). We use this location to indicate the location
of a disruption either at a point resolution or a city/county
resolution. However, geo-tagged tweets are only a few
percentages (1% to 4%) of the total number of tweets. To
address this limitation, we implement a location extraction
method from tweet texts. This approach has several steps
within it. Given a tweet text, the first step is to label each
word (e.g., person’s name, location, organization etc.), which
is known as Named Entity Recognition (NER). We
implement our NER model using the Natural Language
Toolkit (NLTK), developed by (Bird, Klein, & Loper, 2009).
The second step is to extract the location entity, words that
are tagged as location, from the labeled words. In the third
step, we match the extracted location with the county/city
names of the affected regions. Finally, if the extracted
locations are matched, we collect the coordinates using the
geo-coding API provided by Google Maps. The process of
location extraction is shown in Figure 2.
4.3 Dynamic Disruption Mapping
This part of the methodology enables the visualization of
the locations of disruptions with disruption types in a
dynamic way. We visualize the exact disruption location,
only if the location has the exact co-ordinate (location type:
point or latitude-longitude). We choose a time interval to
count the number of disruptions within a geographical
boundary (e.g., county) and then visualize the disruption
intensity as a geographical heat map. We did not consider
disruption severity in this study. But severity can be assumed
to be correlated with the frequency of disruption related
tweets from a given area; the higher the frequency of
disruption related tweets the higher will be the severity level
of disruptions. Hence, a dynamic disruption map can provide
insights about the severity of infrastructure disruptions of an
Table 2 Confusion matrix
Predicted Label
Negative (0)
Positive (1)
Actual Label
Negative (0)
True Negative
(TN)
False Positive
(FP)
Positive (1)
False Negative
(FN)
True Positive
(TP)
Roy et al.
8
area based on the frequency of a specific or all disruption
related posts generated from that area.
5 RESULTS
Using Twitter data from real-world hurricanes, we present
our results to identify infrastructure disruptions and visualize
those disruptions in a dynamic map. To identify disruptions
types and sentiment from text data, we use Binary Relevance,
Label Powerset, and ensemble based multi-label
classification approaches. We compare the performance of
these approaches using eight existing models namely:
Multinomial Naïve Bayes (MNB), Logistic Regression (LR),
K-Nearest Neighborhood (KNN), Support Vector Machine
(SVC), Random Forest (RF), Decision Tree (DT), Multilayer
Perceptron (MLP), and Deep Neural Network (DNN)
methods. The details of these well-known methods can be
found in these studies (Binkhonain & Zhao, 2019; Khan,
Baharudin, Lee, & Khan, 2010). We convert the annotated
tweet text as TF-IDF and annotated label as binary matrix
(multi-label format) by following the steps described in the
data processing section. We use the TF-IDF as input and the
binary matrix as output. For each model, we use 70% (788
tweets) of the annotated samples as training and the rest 30%
(339 tweets) as test samples. We further validate our best
model over 338 tweets from hurricane Michael to test model
performance on the data from an unseen hurricane (i.e., for
hurricane data which were never used for training the model).
We implement all the models in a personal computer using
Python programming language and model parameters are
selected using a grid search approach (Pedregosa et al., 2011).
Moreover, we implement a baseline method that uses
keyword matching and sentiment analysis to identify
disruptions and sentiment characteristics, respectively.
Currently no benchmark method exists that can identify the
co-occurrence of multiple types of disruptions from social
media posts. Since a keyword based approach has been used
in similar studies (Fan & Mostafavi, 2019; Yuan & Liu,
2018), we choose to use this as a baseline method. The
keywords used are listed in Table 3.
For sentiment identification, we use a pre-trained model
adopted from this study (Hutto & Gilbert, 2014); this model
has been trained on social media texts. We consider this
combined (keyword matching and sentiment identification)
approach as a baseline method to evaluate if the trained
models perform better than this baseline method. Table 4
presents the performance of each model on hurricane Irma
test dataset with respect to the selected performance metrics:
subset accuracy, micro F1 measure, and hamming loss.
From the results, we can see that Logistic Regression
classifier (LP method) has the best subset accuracy and micro
F1 scores and Support Vector classifier (RAKEL method) has
the best hamming loss score. The models (LR, KNN, SVC,
MLP, and Deep DNN) perform better than the baseline
method in all approaches (BR, LP, and Ensemble) (see Table
4). Among the three multi-label approaches, LP has the best
performance; RAKEL is second; and BR method is the last
in terms of the considered performance metrics. The reasons
for this result are the following: (i) BR method considers the
labels as mutually exclusive or the correlation between the
disruptions is ignored; (ii) LP method considers the
correlations between the labels/disruptions by considering all
label combination; and (iii) RAKEL method falls between the
BR and LP methods with respect to label correlations as it
considers a random small subset of labels.
To select the best model, we further check the confusion
matrix and choose Logistic Regression (LP method)
classifier. Figure 3 shows the confusion matrix for the LR
(LP) on the test samples from hurricane Irma. The selected
best model (LR-LP) shows 74.93% increase (0.351 to 0.614)
in subset accuracy, 30.73% increase (0.550 to 0.719) in micro
F1 measure, and 44.65% decrease (0.159 to 0.088) in
hamming loss compared to the baseline method.
We also check the performance of our best model (LR-LP)
for disruption and sentiment identification separately. We
validate for hurricanes Irma and Michael, using 339 test data
from hurricane Irma and 338 test data from hurricane
Michael. Table 5 shows the performance on disruption
identification.
Table 3 Keywords for Identifying Disruption
Related Tweets
Disruption Types
Keywords
Power/Electricity
Disruption
power, electricity, outage,
(power, outage), (without,
power)
Communication
Disruption
internet, wi-fi, cell, (no,
internet), (no, network)
Transportation
Disruption
road, roads, traffic,
transportation, turnpike, i-4,
i-95, jam, closed, (traffic,
signal), (road, closed)
Drinking Water
Disruption
drinkingwater,
drinking_water,
bottledwater, bottled_water,
(drinking, water), (bottled,
water)
Wastewater
Related Disruption
wastewater, waste_water,
drainage, drainagewater,
(waste, water), (drainage,
water)
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
9
Except hamming loss for hurricane Michael, our model
performed better for both hurricanes with respect to accuracy,
micro F1, and hamming loss. The baseline method performed
better in hurricane Michael test set than the Irma test data set.
On the other hand, LR-LP model performed better in Irma
data than the Michael data since the model is trained on Irma
dataset.
Table 6 shows the performance of LR-LP model against
the baseline sentiment model (adopted from (Hutto & Gilbert,
2014)). The (LR-LP) model performed better than the
baseline for both Irma and Michael datasets. The baseline
method also performed better for Michael data than Irma data
for sentiment classification. In summary, our developed
model (LR-LP) performed better than the baseline for both
hurricanes Irma (hurricane data used to train the model) and
Michael (unseen hurricane data representing a future
hurricane).
To understand the features that help to correctly identify a
disruption, we analyze the training samples that our model
correctly predicted (i.e., true positive samples in Table 2). For
each disruption type, we rank the words based on their
average TF-IDF score. A higher score represents more
importance of a word for a disruption type. Figure 4 shows
Figure 3 Confusion Matrix (In each panel, the x axis
represents the predicted label and the y axis represents the
actual label in the test set of hurricane Irma. For a
particular label, the value 1 means the presence of this
label whereas 0 means the absence of the label. The value
within a cell represents the number of times a predicted
label matched or mismatched with the actual label)
Table 4 Model Performance Values (Accuracy, Micro F1-measure, Hamming-loss) (A higher score of subset accuracy or micro
F1 measure indicates better performance and a lower score of hamming loss indicates better performance)
Model Name
Binary Relevance (BR)
Label Power set (LP)
Ensemble (RAKEL)
Baseline
(keyword search + sentiment)
0.351, 0.55, 0.159
Multinomial Naïve Bayes (MNB)
0.218, 0.519, 0.145
0.472, 0.615, 0.14
0.268, 0.527, 0.151
Logistic Regression (LR)
0.463, 0.709, 0.090
0.614, 0.719, 0.092
0.525, 0.702, 0.094
K-nearest Neighborhood (KNN)
0.490, 0.613, 0.130
0.525, 0.612, 0.125
0.510, 0.598, 0.126
Support Vector Classifier (SVC)
0.472, 0.707, 0.089
0.608, 0.699, 0.096
0.519, 0.709, 0.088
Random Forest (RF)
0.124, 0.471, 0.170
0.54, 0.635, 0.116
0.357, 0.588, 0.109
Decision Tree (DT)
0.292, 0.628, 0.129
0.522, 0.634, 0.119
0.366, 0.617, 0.124
Multilayer Perceptron (MLP)
0.440, 0.662, 0.099
0.540, 0.615, 0.119
0.507, 0.635, 0.11
Deep Neural Network (DNN)
0.466, 0.342, 0.138
0.569, 0.684, 0.103
-
Table 5: Performance Comparison of disruption
identification
Baseline
Model (LR-LP)
Hurricane
Accuracy, Micro F1-measure, Hamming-loss
Irma
0.351, 0.55, 0.159
0.614, 0.719, 0.092
Michael
0.476, 0.656, 0.115
0.515, 0.658, 0.119
Table 6: Performance comparison of sentiment model
Baseline
Model (LR-LP)
Hurricane
Accuracy, Micro F1-measure, Hamming-
loss
Irma
0.383, 0.368 0.311
0.673, 0.596, 0.165
Michael
0.571, 0.501, 0.226
0.609, 0.656, 0.175
Roy et al.
10
the TF-IDF scores of the top ten words of each disruption type
(shown as horizontal bars) and the TF-IDF scores of the same
words calculated over all disruption types in the training set
(shown as color intensity). We can see that overall words
such as ‘power’, ‘water’, ‘wifi’, ‘internet’, ‘traffic’,
‘drainage’ etc. have higher TF-IDF scores (see the color
intensity of the corresponding bars in Figure 4). It means that
these words are highly important in the overall classification
performance. On the other hand, ‘power’, ‘cell’, ‘stop’,
‘water’, ‘drainage’, ‘close’ are the highest ranked words for
power/electricity disruption, communication disruption,
road/transportation disruption, drinking water disruption,
waste water related disruption, and other disruption,
respectively. Some words (e.g., ‘power’, ‘water’, ‘cell’) are
present in multiple disruption types, indicating that these
words would help identify the co-occurrence of multiple
disruption types. For example, the presence of ‘cell’ and
‘signal’ in the top 3 words of power/electricity and
communication disruptions indicates the co-occurrence of
these two types of disruptions. Regarding sentiment features,
the word ‘power’ is common in all the three sentiments. The
differences among the words present in these three sentiment
classes are: (i) the negative (actual disruption) contains the
words that are mostly present in the disruption types, (ii) the
positive sentiment contains slang words such as ‘hell yeah’,
‘yeah’, ‘ac loll’, (iii) neutral sentiment contains situation and
forecast related words such as ‘update’, ‘best update’,
‘situation’, ‘chance wont’, ‘good chance’ (see Figure 4).
6 CASE STUDIES: HURRICANES IRMA AND
MICHAEL
In this section, we present two case studies of our proposed
approach, one for hurricane Irma and another for hurricane
Michael. Our best model (LR-LP) predicts the disruption
types and status over the input data described in Table 1. As
shown in Figure 2, for a geotagged tweet, we obtain the
disruption location from the tweet geo-location information.
Otherwise, we extract the location from the tweet texts using
the geocoding module. We match the extracted location with
the city/county of a state and then obtain the coordinate using
Google Maps API.
Finally, we plot the disruption types and status in a
disruption map. To understand the hurricane context, we also
Figure 4 Important features for different disruption types. The X axis shows the mean TF-IDF score
(calculated over individual disruption type) and the Y axis shows the words/features. The color of the bar
indicates the mean TF-IDF score (calculated over all disruption types). The calculated scores and important
features are based on the training dataset.
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
11
present the hurricane track and wind speed data collected
from the National Hurricane Center (NOAA, 2019). Two
snapshots of the power/electricity disruption map from each
hurricane are shown in Figure 5 (5a.1 and 5a.2 for hurricane
Irma, 5b.1 and 5b.2 for hurricane Michael). We use a 3-hour
time-interval for aggregating the tweets to create the county-
level disruption heat map. The inset plot shows the locations
of power/electricity disruptions. We show the location of
hurricane center (shown as a circle at the beginning of the
hurricane track line), wind speed (through the color of the
circle), and disruption related tweets (geographic heat map)
which will be updated dynamically as we receive data from
Twitter stream. Figure 5a.1 shows a snapshot of Irma at
around 7 PM on Sept. 10, 2017. It shows that majority of the
power/electricity related posts were generated from Miami-
Dade, Broward, Palm-Beach counties when Irma’s center
was near Collier county with a wind speed of around 120
mph. However, not all the posts are about the actual power
outage incident (disruptions are represented by black circles
in the inset plot of Figure 5a.1), and a substantial number of
these posts are expressing concerns about power outage or
expressing that they still have power. The second snapshot
(Figure 5a.2) shows that when the center of Irma was near
Tampa, most of the disruption related tweets were posted
from Orlando, Tampa, and Miami-Dade counties. A dynamic
disruption map of Michael shows similar results. On October
a.1
a.2
(a) Hurricane Irma
b.1
b.2
(b) Hurricane Michael
Figure 5 Dynamic Disruption Map for Power/Electricity Disruption: (a) Hurricane Irma, (b) Hurricane Michael.
Roy et al.
12
10, 2018 around 6 PM (Figure 5b.1), when Michael was
about to make its landfall near Tallahassee, most of the
power/electricity disruption related tweets were coming from
Tallahassee area. Figure 5b.2 shows the second snapshot of
Michael around midnight of October 12, 2018 when the
center of Michael was over North Carolina. It shows that
most of power/electricity related disruptions were coming
from Wake, Johnston, Durham and Orange counties of North
Carolina.
Finally, we visualize the co-occurrence of multiple
disruption types in an interactive map. Figure 6 shows a
snapshot of the co-occurrence map for hurricane Irma (Figure
6a) and Michael (Figure 6b). We plot this map using only the
actual disruption samples (negative sentiment) aggregated
over a 1-hour interval. This interactive map allows to explore
the disruptions type separately as well as a combination of
them. The co-occurrence heat map shows a relative intensity
of the disruptions based on the co-occurrences of all the
disruption types. For Irma, mostly co-occurred disruptions
are power, communication, and transportation disruptions.
On the other hand, for hurricane Michael (see Figure 6b) the
most co-occurred disruptions are power and transportation
disruptions.
In summary, we find that during hurricanes Irma and
Michael affected people posted infrastructure related tweets.
Those posts may represent actual infrastructure disruptions. A
multi-label classification approach (a logistic regression
model adopted over a label powerset) has been developed to
predict both the disruption types and disruption status from
such data. After locating the disruptions using a geocoding
approach, a map can visualize the disruptions spatially and
temporally. The training time of the model is about 7 sec, and
it takes about 1 sec to process, predict, and visualize the data
collected over one hour. Thus, this approach can be easily
applied in a real-time setting.
7 LIMITATIONS AND FUTURE RESEARCH
DIRECTIONS
Our study has some limitations. For instance, the annotated
dataset is small in comparison to the entire dataset. More
annotated samples are likely to increase the accuracy of the
model. Although the co-occurrence of multiple disruptions is
considered, the approach cannot infer if a disruption is caused
by another disruption. Incorporating causality as an input to
the model may improve its performance. Another limitation
of our approach is that we have checked the accuracy of the
method based on human-annotated tweets, which may not
represent the total number of disruptions observed in the
ground. To check the extent to which the reported disruptions
match actual ones, ground truth data on disruptions occurring
in different infrastructure systems are required. These
datasets, often collected by infrastructure service providers
including private companies and public agencies, may
contain sensitive information. Collecting ground truth data on
infrastructure disruptions from a variety of sources covering
multiple states will be a very challenging task. Further studies
are needed to verify what percentage of actual disruptions is
reported in social media and to what extent these disruptions
can be identified using the method developed in this study. In
addition, our data cover hurricanes only. Future studies can
transfer and validate our approach across other disasters such
as wildfire, earthquake, snowfall, and thunderstorms.
In this study, we assume that a post with a negative
sentiment is associated with an actual disruption, and a post
with a neutral or positive sentiment is associated with no
disruption. However, there could be a post with a positive
(a) Hurricane Irma
(Time 2017-09-11 05:00:00)
(b) Hurricane Michael
(Time 2018-10-12 06:00:00)
Figure 6 Disruption Co-occurrence Map (a) Hurricane Irma (b) Hurricane Michael
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
13
sentiment, but associated with an actual disruption. These
tweets are likely to be a small portion of the entire dataset. In
our annotated dataset, we did not find such tweets. Future
studies, based on natural language processing, can develop
more advanced methods to capture the situations where even
a positive tweet could be associated with a disruption.
When the communication network is disrupted, affected
people may not have access to social media platforms. In such
situations, our model cannot detect disruptions. In the geo-
parsing method, we use exact matching process between the
extracted location and county/city of the affected regions.
Since our approach finds city/county names only, it cannot
extract location if street or any finer level location is
mentioned in the text. In future studies, text-based location
matching can be developed with finer resolution (e.g., street
name), which may help in locating disruptions with more
specific location information.
For training our models, we adopt a batch learning
approach which requires retraining the model to incorporate
new data from the data stream. Future studies can explore an
incremental learning approach (T. T. Nguyen et al., 2019;
Read, Bifet, Holmes, & Pfahringer, 2012) to dynamically
train models on newly available data from the ongoing/future
disasters (NOAA National Centers for Environmental
Information (NCEI) U.S., 2018). Such an incremental
learning approach is likely to increase the accuracy of the
model as it utilizes data from an ongoing disaster.
To achieve a better classification accuracy, more complex
classification methods such as probabilistic neural networks
(Ahmadlou & Adeli, 2010), dynamic neural networks (Rafiei
& Adeli, 2017a), and hierarchy-based models (Cerri,
Basgalupp, Barros, & de Carvalho, 2019; Wehrmann, Cerri,
& Barros, 2018) can be considered. A probabilistic neural
network is a fast, efficient, and flexible model to add/remove
new training data and hence may be more suitable for real-
time disruption prediction for an unseen disaster. Since
textual data have a large feature space, a dynamic neural
network might be useful in finding an optimal number of
features to achieve better performance. Moreover, hierarchy-
based models might be more suitable when there exists more
hierarchy in the disruption types, especially considering
disruptions from multiple disasters (hurricane, wildfire,
snowstorm etc.). A hierarchy-based model can have classes
for disaster type, disruption type, and disruption status. A
hierarchical relationship can be created from disaster type to
disruption type to disruption status (e.g., if a post is not
disaster related it has no disruption type and disruption
status).
8 CONCLUSIONS
This paper presents an approach to identify infrastructure
disruptions and a dynamic disruption mapping framework
using social media data. While previous research focused
mainly on identifying hurricane or damage related social
media posts, we consider five types (power/electricity,
communication, drinking water, and wastewater) of
infrastructure disruptions, their co-occurrence, and their
status (whether a post is reporting an actual disruption,
disruption in general, or not affected by a disruption). The
result shows that our multi-label classification approach
(logistic regression adopted in a label powerset approach)
performs better than a baseline method (based on keyword
search and sentiment analysis). Moreover, we present a
method, to visualize disruptions in a dynamic map.
Identifying disruption types and disruption locations is vital
for disaster recovery, response and relief operations. The
developed approach of identifying the co-occurrence of
multiple disruptions may help coordinate among
infrastructure service providers and disaster management
organizations.
ACKNOWLEDGMENT
The authors are grateful to the U.S. National Science
Foundation for the grants CMMI-1832578, CMMI-1832693,
and CMMI-1917019 to support the research presented in this
paper. However, the authors are solely responsible for the
findings presented in this paper.
REFERENCES
Ahmadlou, M., & Adeli, H. (2010). Enhanced probabilistic
neural network with local decision circles: A robust
classifier. Integrated Computer-Aided Engineering,
17(3), 197–210.
Alavi, A. H., & Buttlar, W. G. (2019). An overview of
smartphone technology for citizen-centered, real-time
and scalable civil infrastructure monitoring. Future
Generation Computer Systems, 93, 651–672.
Alinizzi, M., Chen, S., Labi, S., & Kandil, A. (2018). A
Methodology to Account for One-Way Infrastructure
Interdependency in Preservation Activity Scheduling.
Computer-Aided Civil and Infrastructure Engineering,
33(11), 905–925.
Barabási, A.-L., & Albert, R. (1999). Emergence of scaling
in random networks. Science, 286(5439), 509–512.
Binkhonain, M., & Zhao, L. (2019). A review of machine
learning algorithms for identification and classification
of non-functional requirements. Expert Systems with
Applications.
Bird, S., Klein, E., & Loper, E. (2009). Natural language
processing with Python: analyzing text with the natural
language toolkit. “ O’Reilly Media, Inc.”
Birregah, B., Top, T., Perez, C., Châtelet, E., Matta, N.,
Lemercier, M., & Snoussi, H. (2012). Multi-layer crisis
mapping: a social media-based approach. In 2012 IEEE
21st International Workshop on Enabling
Roy et al.
14
Technologies: Infrastructure for Collaborative
Enterprises (pp. 379–384).
Bryson, K.-M. N., Millar, H., Joseph, A., & Mobolurin, A.
(2002). Using formal MS/OR modeling to support
disaster recovery planning. European Journal of
Operational Research, 141(3), 679–688.
Buldyrev, S. V, Parshani, R., Paul, G., Stanley, H. E., &
Havlin, S. (2010). Catastrophic cascade of failures in
interdependent networks. Nature, 464(7291), 1025.
Cerri, R., Basgalupp, M. P., Barros, R. C., & de Carvalho, A.
C. (2019). Inducing Hierarchical Multi-label
Classification rules with Genetic Algorithms. Applied
Soft Computing, 77, 584–604.
Chang, P. C., Flatau, A., & Liu, S. C. (2003). Health
monitoring of civil infrastructure. Structural Health
Monitoring, 2(3), 257–267.
Cheng, Z., Caverlee, J., & Lee, K. (2010). You are where
you tweet: a content-based approach to geo-locating
twitter users. In Proceedings of the 19th ACM
international conference on Information and
knowledge management (pp. 759–768).
Cresci, S., Cimino, A., Dell’Orletta, F., & Tesconi, M.
(2015). Crisis mapping during natural disasters via text
analysis of social media messages. In International
Conference on Web Information Systems Engineering
(pp. 250–258).
Deng, Q., Liu, Y., Zhang, H., Deng, X., & Ma, Y. (2016). A
new crowdsourcing model to assess disaster using
microblog data in typhoon Haiyan. Natural Hazards,
84(2), 1241–1256.
Du, E., Cai, X., Sun, Z., & Minsker, B. (2017). Exploring the
role of social media and individual behaviors in flood
evacuation processes: An agent-based modeling
approach. Water Resources Research, 53(11), 9164–
9180.
Dyskin, A. V, Basarir, H., Doherty, J., Elchalakani, M.,
Joldes, G. R., Karrech, A., … others. (2018).
Computational monitoring in real time: review of
methods and applications. Geomechanics and
Geophysics for Geo-Energy and Geo-Resources, 4(3),
235–271.
Fan, C., & Mostafavi, A. (2019). A graph-based method for
social sensing of infrastructure disruptions in disasters.
Computer-Aided Civil and Infrastructure Engineering.
Fan, C., Mostafavi, A., Gupta, A., & Zhang, C. (2018). A
system analytics framework for detecting
infrastructure-related topics in disasters using social
sensing. In Workshop of the European Group for
Intelligent Computing in Engineering (pp. 74–91).
Fang, Y.-P., & Sansavini, G. (2019). Optimum post-
disruption restoration under uncertainty for enhancing
critical infrastructure resilience. Reliability
Engineering & System Safety, 185, 1–11.
FCC. (2017). Communications Status Report for Areas
Impacted by Hurricane Irma.
Fry, J., & Binner, J. M. (2015). Elementary modelling and
behavioural analysis for emergency evacuations using
social media. European Journal of Operational
Research, 249(3), 1014–1023. Retrieved from
http://www.scopus.com/inward/record.url?eid=2-s2.0-
84930690948&partnerID=tZOtx3y1
Gao, H., Barbier, G., & Goolsby, R. (2011). Harnessing the
crowdsourcing power of social media for disaster
relief. IEEE Intelligent Systems, 26(3), 10–14.
Guan, X., & Chen, C. (2014). Using social media data to
understand and assess disasters. Natural Hazards,
74(2), 837–850.
Hasan, S., & Foliente, G. (2015). Modeling infrastructure
system interdependencies and socioeconomic impacts
of failure in extreme events: emerging R&D
challenges. Natural Hazards, 78(3), 2143–2168.
Homeland Security. (2019). Critical Infrastructure Sectors.
Retrieved January 8, 2019, from
https://www.dhs.gov/cisa/critical-infrastructure-sectors
Huang, Q., & Xiao, Y. (2015). Geographic Situational
Awareness: Mining Tweets for Disaster Preparedness,
Emergency Response, Impact, and Recovery. ISPRS
International Journal of Geo-Information, 4(3), 1549–
1568. Retrieved from http://www.mdpi.com/2220-
9964/4/3/1549/htm
Hutto, C. J., & Gilbert, E. (2014). Vader: A parsimonious
rule-based model for sentiment analysis of social
media text. In Eighth international AAAI conference on
weblogs and social media.
Jenelius, E., & Mattsson, L.-G. (2012). Road network
vulnerability analysis of area-covering disruptions: A
grid-based approach with case study. Transportation
Research Part A: Policy and Practice, 46(5), 746–760.
Jongman, B., Wagemaker, J., Romero, B., & de Perez, E.
(2015). Early flood detection for rapid humanitarian
response: harnessing near real-time satellite and
Twitter signals. ISPRS International Journal of Geo-
Information, 4(4), 2246–2266.
Kadri, F., Birregah, B., & Châtelet, E. (2014). The impact of
natural disasters on critical infrastructures: A domino
effect-based study. Journal of Homeland Security and
Emergency Management, 11(2), 217–241.
Keim, M. E., & Noji, E. (2010). Emergent use of social
media : A new age of opportunity for disaster
resilience. American Journal of Disaster Medicine,
A Multi-label Classification Approach to Identify Hurricane-induced Infrastructure Disruptions Using Social Media Data
15
6(1), 47–54.
Khan, A., Baharudin, B., Lee, L. H., & Khan, K. (2010). A
review of machine learning algorithms for text-
documents classification. Journal of Advances in
Information Technology, 1(1), 4–20.
Kotsiantis, S. B., Zaharakis, I., & Pintelas, P. (2007).
Supervised machine learning: A review of
classification techniques. Emerging Artificial
Intelligence Applications in Computer Engineering,
160, 3–24.
Kryvasheyeu Y, Chen H, Moro E, Van Hentenryck P, C. M.
(2015). Performance of Social Network Sensors
During Hurricane Sandy. PLoS One 10.2 (2015):
E0117288., 10(2). Retrieved from
http://arxiv.org/abs/1402.2482
Kryvasheyeu, Y., Chen, H., Obradovich, N., Moro, E.,
Hentenryck, P. Van, Fowler, J., & Cebrian, M. (2016).
Rapid assessment of disaster damage using social
media activity. Science Advances 2.3 (2016):
E1500779.
Lambert, J. H., & Patterson, C. E. (2002). Prioritization of
schedule dependencies in hurricane recovery of
transportation agency. Journal of Infrastructure
Systems, 8(3), 103–111.
Li, C., & Sun, A. (2014). Fine-grained location extraction
from tweets with temporal awareness. In Proceedings
of the 37th international ACM SIGIR conference on
Research & development in information retrieval (pp.
43–52).
Li, Z., Wang, C., Emrich, C. T., & Guo, D. (2018). A novel
approach to leveraging social media for rapid flood
mapping: a case study of the 2015 South Carolina
floods. Cartography and Geographic Information
Science, 45(2), 97–110.
Lu, L., Wang, X., Ouyang, Y., Roningen, J., Myers, N., &
Calfas, G. (2018). Vulnerability of interdependent
urban infrastructure networks: Equilibrium after failure
propagation and cascading impacts. Computer-Aided
Civil and Infrastructure Engineering, 33(4), 300–315.
Martani, C., Jin, Y., Soga, K., & Scholtes, S. (2016). Design
with uncertainty: the role of future options for
infrastructure integration. Computer-Aided Civil and
Infrastructure Engineering, 31(10), 733–748.
Martín, Y., Li, Z., & Cutter, S. L. (2017). Leveraging Twitter
to gauge evacuation compliance: Spatiotemporal
analysis of Hurricane Matthew. PLoS ONE, 12(7), 1–
22. https://doi.org/10.1371/journal.pone.0181701
Middleton, S. E., Middleton, L., & Modafferi, S. (2013).
Real-time crisis mapping of natural disasters using
social media. IEEE Intelligent Systems, 29(2), 9–17.
Mitsova, D., Escaleras, M., Sapat, A., Esnard, A.-M., &
Lamadrid, A. J. (2019). The Effects of Infrastructure
Service Disruptions and Socio-Economic Vulnerability
on Hurricane Recovery. Sustainability, 11(2), 516.
Mouzannar, H., Rizk, Y., & Awad, M. (2018). Damage
Identification in Social Media Posts using Multimodal
Deep Learning. In ISCRAM.
NERC. (2018). Hurricane Irma Event Analysis Report.
Www.Nerc.Com.
Nguyen, D. T., Ofli, F., Imran, M., & Mitra, P. (2017).
Damage assessment from social media imagery data
during disasters. In Proceedings of the 2017
IEEE/ACM International Conference on Advances in
Social Networks Analysis and Mining 2017 (pp. 569–
576).
Nguyen, T. T., Dang, M. T., Luong, A. V., Liew, A. W.-C.,
Liang, T., & McCall, J. (2019). Multi-Label
Classification via Incremental Clustering on Evolving
Data Stream. Pattern Recognition.
NOAA. (2019). National Hurricane Center Data Archive.
Retrieved from https://www.nhc.noaa.gov/data/
NOAA National Centers for Environmental Information
(NCEI) U.S. (2018). Billion-Dollar Weather and
Climate Disasters. Retrieved January 15, 2018, from
https://www.ncdc.noaa.gov/billions/
Ouyang, M., & Fang, Y. (2017). A mathematical framework
to optimize critical infrastructure resilience against
intentional attacks. Computer-Aided Civil and
Infrastructure Engineering, 32(11), 909–929.
Oyeniyi, D. (2017). How Hurricane Harvey Changed Social
Media Disaster Relief. TexasMonthly. Retrieved from
https://www.texasmonthly.com/the-daily-post/how-
social-media-managers-responded-to-hurricane-harvey/
Pant, R., Thacker, S., Hall, J. W., Alderson, D., & Barr, S.
(2018). Critical infrastructure impact assessment due to
flood exposure. Journal of Flood Risk Management,
11(1), 22–33.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,
Thirion, B., Grisel, O., … others. (2011). Scikit-learn:
Machine learning in Python. Journal of Machine
Learning Research, 12(Oct), 2825–2830.
Pietrucha-Urbanik, K., & Tchórzewska-Cieślak, B. (2018).
Approaches to failure risk analysis of the water
distribution network with regard to the safety of
consumers. Water, 10(11), 1679.
Rafiei, M. H., & Adeli, H. (2017a). A new neural dynamic
classification algorithm. IEEE Transactions on Neural
Networks and Learning Systems, 28(12), 3074–3083.
Rafiei, M. H., & Adeli, H. (2017b). A novel machine
Roy et al.
16
learning-based algorithm to detect damage in high-rise
building structures. The Structural Design of Tall and
Special Buildings, 26(18), e1400.
Rafiei, M. H., & Adeli, H. (2018a). A novel unsupervised
deep learning model for global and local health
condition assessment of structures. Engineering
Structures, 156, 598–607.
Rafiei, M. H., & Adeli, H. (2018b). Novel machine-learning
model for estimating construction costs considering
economic variables and indexes. Journal of
Construction Engineering and Management, 144(12),
4018106.
Rafiei, M. H., Khushefati, W. H., Demirboga, R., & Adeli,
H. (2017). Supervised Deep Restricted Boltzmann
Machine for Estimation of Concrete. ACI Materials
Journal, 114(2).
Ramos, J., & others. (2003). Using tf-idf to determine word
relevance in document queries. In Proceedings of the
first instructional conference on machine learning
(Vol. 242, pp. 133–142).
Read, J., Bifet, A., Holmes, G., & Pfahringer, B. (2012).
Scalable and efficient multi-label classification for
evolving data streams. Machine Learning, 88(1–2),
243–272.
Rinaldi, S. M., Peerenboom, J. P., & Kelly, T. K. (2001).
Identifying, understanding, and analyzing critical
infrastructure interdependencies. IEEE Control
Systems Magazine, 21(6), 11–25.
Rosenzweig, C., & Solecki, W. (2014). Hurricane Sandy and
adaptation pathways in New York: Lessons from a
first-responder city. Global Environmental Change, 28,
395–408.
Roy, K. C., Cebrian, M., & Hasan, S. (2019). Quantifying
human mobility resilience to extreme events using geo-
located social media data. EPJ Data Science, 8(1), 18.
Sadri, A. M., Hasan, S., & Ukkusuri, S. V. (2019). Joint
Inference of User Community and Interest Patterns in
Social Interaction Networks. Social Network Analysis
and Mining, 9, 11. Retrieved from
http://arxiv.org/abs/1704.01706
Sörensen, S., Webster, J. D., & Roggman, L. A. (2002).
Adult attachment and preparing to provide care for
older relatives. Attachment & Human Development,
4(1), 84–106.
Sorower, M. S. (2010). A literature survey on algorithms for
multi-label learning. Oregon State University,
Corvallis, 18, 1–25.
Sriram, L. M. K., Ulak, M. B., Ozguven, E. E., &
Arghandeh, R. (2019). Multi-Network Vulnerability
Causal Model for Infrastructure Co-Resilience. IEEE
Access, 7, 35344–35358.
Sumalee, A., & Kurauchi, F. (2006). Network capacity
reliability analysis considering traffic regulation after a
major disaster. Networks and Spatial Economics, 6(3–
4), 205–219.
Tang, Z., Zhang, L., Xu, F., & Vo, H. (2015). Examining the
role of social media in California’s drought risk
management in 2014. Natural Hazards, 79(1), 171–
193. Retrieved from
http://link.springer.com/10.1007/s11069-015-1835-2
The Effect of Hurricane Irma on Water Supply. (2017).
KRAUSZ.COM.
Tsoumakas, G., & Vlahavas, I. (2007). Random k-labelsets:
An ensemble method for multilabel classification. In
European conference on machine learning (pp. 406–
417).
Ulak, M. B., Kocatepe, A., Sriram, L. M. K., Ozguven, E. E.,
& Arghandeh, R. (2018). Assessment of the hurricane-
induced power outages from a demographic,
socioeconomic, and transportation perspective. Natural
Hazards, 92(3), 1489–1508.
Unsafe Drinking Water After Hurricane Irma. (2019). Wayde
King Water Filtration.
Wang, Q., & Taylor, J. E. (2014). Quantifying human
mobility perturbation and resilience in hurricane sandy.
PLoS ONE, 9(11), 1–5.
Wehrmann, J., Cerri, R., & Barros, R. (2018). Hierarchical
multi-label classification networks. In International
Conference on Machine Learning (pp. 5225–5234).
Yuan, F., & Liu, R. (2018). Feasibility study of using
crowdsourcing to identify critical affected areas for
rapid damage assessment: Hurricane Matthew case
study. International Journal of Disaster Risk
Reduction, 28, 758–767.
Yuan, F., & Liu, R. (2019). Identifying Damage-Related
Social Media Data during Hurricane Matthew: A
Machine Learning Approach. Computing in Civil
Engineering 2019.
A preview of this full-text is provided by Wiley.
Content available from Computer-Aided Civil and Infrastructure Engineering
This content is subject to copyright. Terms and conditions apply.
