Dynamics of Crisis Communications in Social Media: Spatio-temporal and Text-based Comparative Analyses of Twitter Data from Hurricanes Irma and Michael

Abstract

Social media platforms play critical roles in information dissemination, communication and coordination during different phases of natural disasters as it is crucial to know the type of crisis information being disseminated and user concerns. Large-scale Twitter data from hurricanes Irma (Sept. 2017) and Michael (Oct. 2018) are used here to understand the topic dynamics over time by applying the Dynamic Topic Model, followed by a comparative analyses of the differences in such dynamics for these two hurricane scenarios. We performed a spatio-temporal analyses of user activities with reference to the hurricane center location and wind speed. The findings of spatio-temporal analyses show that differences in hurricane path and the affected regions influence user participation and social media activity. Besides, topic dynamics reveals that situational awareness, disruptions, relief action are among the patterns common for both hurricanes; unlike topics such as hurricane evacuation and political situation that are scenario dependent.

Full text

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
Dynamics of Crisis Communications in
Social Media: Spatio-temporal and
Text-based Comparative Analyses of
Twitter Data from Hurricanes Irma and
Michael
Kamol Roy
Ph.D. Candidate
Department of CECE
University of Central Florida
Email: roy.kamol@knights.ucf.edu
MD Ashraf Ahmed
Ph.D. Student
Department of CEE
Florida International University
Email. mpave002@fiu.edu
Samiul Hasan
Assistant Professor
Department of CECE
University of Central Florida
Email: samiul.hasan@ucf.edu
Arif Mohaimin Sadri, Ph.D.
Assistant Professor
Department of MDCM
Florida International University
Email. asadri@fiu.edu
ABSTRACT
Social media platforms play critical roles in information dissemination, communication and co-ordination during
different phases of natural disasters as it is crucial to know the type of crisis information being disseminated and
user concerns. Large-scale Twitter data from hurricanes Irma (Sept. 2017) and Michael (Oct. 2018) are used here
to understand the topic dynamics over time by applying the Dynamic Topic Model, followed by a comparative
analyses of the differences in such dynamics for these two hurricane scenarios. We performed a spatio-temporal
analyses of user activities with reference to the hurricane center location and wind speed. The findings of spatio-
temporal analyses show that differences in hurricane path and the affected regions influence user participation
and social media activity. Besides, topic dynamics reveals that situational awareness, disruptions, relief action are
among the patterns common for both hurricanes; unlike topics such as hurricane evacuation and political situation
that are scenario dependent.
Keywords
Social Media, Dynamic Topic Modeling, Irma, Michael, Disaster Management.
INTRODUCTION
Natural disasters have severe damaging impacts on society and economy. To take necessary actions during a
disaster, affected population need to be well informed about the threats of a disaster and emergency management
agencies need to understand community concerns and needs in different phases of disaster. Besides traditional
media, social media plays critical role in disseminating weather and situational awareness information (Imran,
Castillo, Diaz, & Meier, 2013; Oyeniyi, 2017; Sadri, Hasan, Ukkusuri, & Cebrian, 2017b; Xiao, Huang, & Wu,
2015). Social media platforms such as Facebook, Twitter, and Instagram etc. are uniquely different from
traditional media because traditional media is mainly designed for one way communication whereas social media
allows two way communication (Brown, 2015). However, it is critical to harness the rich information from these
information sources because of the velocity, variety, and volume of the big data (Brown, 2015; Stieglitz,
Mirbabaie, Ross, & Neuberger, 2018). In addition, with ubiquitous adoption of platforms—such as 2.38 billion
812

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
monthly active users on Facebook (Noyes, 2019) and 321 million on Twitter (Shaban, 2019)—online social media
has become a very important data source of user-generated content during a disaster. Many studies have shown
the potential of social media data in disaster management mainly in crisis communication (Lachlan, Spence, Lin,
Najarian, & Del Greco, 2016; Sadri, Hasan, Ukkusuri, & Cebrian, 2017a; Sadri, Ukkusuri, et al., 2017), human
mobility analysis (Beiró, Panisson, Tizzoni, & Cattuto, 2016; Pan, Zheng, Wilkie, & Shahabi, 2013; Rashidi,
Abbasi, Maghrebi, Hasan, & Waller, 2017; Roy, Cebrian, & Hasan, 2019; Q. Wang & Taylor, 2016), damage
assessment (Kryvasheyeu et al., 2016) and event detection (Dong, Mavroeidis, Calabrese, & Frossard, 2015;
Kryvasheyeu & Chen, 2015) among others.
Collecting ongoing topics in pre-disaster, during disaster and post-disaster period is vital in making necessary
decisions and take systematic actions for preparation, response and recovery logistics. Social media such as
Twitter has been one of the mostly used crowd-sourcing data sources for this purpose(Kryvasheyeu & Chen, 2015;
Martín, Li, & Cutter, 2017; Z. Wang, Ye, & Tsou, 2016). Machine learning approaches such as topic modelling,
e.g. using Latent Dirichlet Allocation (LDA) model (Blei, Ng, & Jordan, 2003), can identify topics from a
collection of documents. Thus, this approach has been adopted by recent studies for identifying disaster related
topics (Patel, 2018; Resch, Usländer, & Havas, 2018; Sadri, Hasan, et al., 2017b). Topic modeling can help
disaster social media analysis in many ways, for example, Resch, Usländer, and Havas(Resch et al., 2018) used
spatiotemporal analysis and topic modeling for earthquake footprint and damage assessment, thus adding one
more information layer to disaster management. Yang et al. has extended the LDA model to understand location
specific topic and topic sentiment for hurricane Sandy(Yang et al., 2017). However, most of the topics that have
been explored in the previous studies are static in nature, and thus cannot answer how a particular topic evolve
over time during different phases of disaster.
Moreover, a comparative study on how a topic evolves for two different disasters is still unexplored in the existing
literature. Each hurricane is different in terms of intensity, speed and hurricane path, and thus the topics triggered
by these hurricanes can be different except some commonalities. The findings of a comparative study will provide
an empirical basis to monitor general topics (based on similar topics in the comparison) and context-dependent
topics during such future hurricanes. In this paper, we present a dynamic extension of topic modeling approach,
Dynamic Topic Model (DTM) (Blei & Lafferty, 2006), to understand the topic dynamics using Twitter data
collected during hurricanes Irma and Michael. Our study focuses on three main research questions:
1. How does the spatio-temporal distribution of social media activity vary in response to the hurricane
center location and wind speed and does such spatio-temporal distribution depend on hurricane path
and affected regions?
2. How do the topics, discussed in social media, evolve over time during different phases of disaster?
3. What are similarities and dissimilarities in the topic dynamic in two different disaster contexts?
LITERATURE REVIEW
Social media has become one of the most feasible platforms for communication in many recent natural disasters
such as hurricane Florence, Michael, and wildfire at California. The affected people have used social media in
many ways such as —for sharing ground realities from the affected regions, taking informed decision and seeking
supports through peer networks (Cecile Wendling, Jack Radisch, 2013; Subba & Bui, 2017). Fast dissemination
of information about the ground realities of disaster-affected areas through social media helped emergency
management agencies to prepare, mobilize, and coordinate their resources for rescue and relief efforts (Slamet et
al., 2018). In disaster affected areas, communication among people is also critical as it provides essential
information regarding community needs and concerns about major infrastructures (different evacuation routes
and transportation facilities) during warning, response and recovery phases (Sadri, Hasan, et al., 2017b). Social
media data (twitter data) associated with the tornado in Moore and Oklahoma also provided similar information
related to public behavior as well as peoples’ critical needs (Ukkusuri et al. 2014; Sadri et al. 2018a).
Topic modelling, especially latent Dirichlet allocation (LDA), is used in disaster related information retrieval
from social media data. For example, Sadri et al. (Sadri, Hasan, et al., 2017a) have used static LDA model to
identify disaster related topics using Twitter data. This study has trained an LDA model over the whole dataset
irrespective of the time slice, thus cannot capture the evolution of topic over time. This study (Patel, 2018) has
used the static LDA model trained on both for the whole dataset (entire period) and by date (24 hours). A static
LDA for each time slice (e.g. 6 hours or 12 hours or 24 hours) will not consider the topics evolved from the
previous time slice. Thus, each time slice may have completely different topic distribution disregarding the
relation with the topics from previous time slice. Resch et al(Resch et al., 2018) has proposed a use case of static
LDA model to detect earthquake damage. Thus, this study also cannot capture the dynamics of the word
813

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
distribution within a topic. Yang et al. (Yang et al., 2017) proposed an approach to incorporate sentiment with the
topic given some keywords as priori knowledge and did not focus on the temporal dynamics of the topics.
In our study, we have used a Dynamic Topic Model (DTM) (Blei & Lafferty, 2006) to understand the topic
dynamics using Twitter data collected during hurricanes Irma and Michael.
DATA PROCESSING AND DESCRIPTIONS
In this study, we have used Twitter data collected through its API during hurricanes Irma and Michael. Hurricane
Irma made its landfall at the Florida Keys on Sept. 10, 2017 as a category 4 hurricane and Michael made the
landfall on Oct. 10, 2018 at Florida Panhandle as a category 5 hurricane. To identify the evolving topics only
from the affected regions, we filtered the data for the Florida users. For hurricane Irma, we collect 1.81 million
tweets posted by 248,763 users. To collect data for more timestamp we collect user specific data (up to 3200
tweets) using rest API for 16399 Florida users, identified from the streaming data. The combined data contains
~2.48 million tweets from 16,399 users. Besides for Michael, ~12 million raw tweets from ~3.8 million users
were collected by using relevant keywords and geolocation boundary. From that, ~200 thousand users were fr om
the four affected states (Florida, North Carolina, Georgia and Virginia) of Michael and we found ~1.5 million
tweets from them. For our study, we selected 10 days of data, 5 days before landfall and 5 days after landfall, for
each hurricane. During this 10 days of timestamp, ~ 640 thousand tweets were separated among which ~200
thousand tweets were found as repeated, hence resulting in ~450 thousand unique tweets from the Michael
affected states. In social media like Twitter, the users publish content about many different topics, and a lot of
these contents are not relevant to our study. Thus, appropriate filtering is necessary to get meaningful insights
from the data (Roy, Hasan, Sadri, & Cebrian, 2020). To get the data relevant to a particular disaster, we further
filter the datasets by keywords such as ‘hurricane’ and the name of the hurricane (‘irma’ or ‘michael’). For
dynamic topic models, we use only tweet texts which contain a lot of noisy information that can deteriorate the
quality of the topics. Thus, it is necessary to remove the noise from the tweet text before fitting the data into the
model. To be consistent with the tweet language, we take only the English tweets for our analysis. We remove the
emoticons from the texts. Although emoticons can represent sentiments and emotions, it is not standardized and
hard to interpret. In the tweet texts, users can mention other users starting with @ symbol. Although it might
represent some communication pattern, a user name does not give any meaningful interpretation within a topic
and thus, we remove all the user mentions from the text. We also remove URLs from the tweets because they do
not provide any semantically interpretable information. Further, we remove the stop words, punctuation and
special characters (\@/#$ etc.) from the text. Finally, we convert all the tokens/words as lowercase before creating
the corpus for fitting the topic model. After applying relevance filtering, the data size is as following-
 Irma data contains 32,021 tweets with vocabulary size 11,872
 Michael data contains 78,479 tweets with vocabulary size 23,837
MODELING APPROACH
In this study, we have used Dynamic Topic Model to understand the topic dynamics during a disaster. Dynamic
Topic Model (DTM) is introduced by Blei et al. (Blei & Lafferty, 2006) as an enhancement to the previous Latent
Dirichlet Allocation (LDA) (Blei et al., 2003). Compare to the static LDA model, DTM includes a notion of time
that describes the evolution of word-topic distributions over time. In a static topic model, such as LDA, if 𝛽: be
𝐾 topics, each topic is considered as a distribution over a fixed vocabulary and each document is considered as a
mixture of 𝑘 topics. In LDA, the documents are drawn exchangeably from the same set of 𝑘 topics, whereas in
DTM the documents are grouped by time slice (e.g. 6 hours in our study) and the documents of a time slice 𝑡
come from a set of topics that evolved from the previous time slice. Figure 1 shows the plate notation of DTM,
wher e 𝛽, denote the 𝑉 (assuming 𝑉 terms/vocabularies) dimensional vectors of natural parameter for topic 𝑘 in
time slice 𝑡. The component of 𝛽 is represented by 𝑉 dimensional multinomial, 𝜋 where, the 𝑖 component is
𝛽=log (
). The state space model evolves over time by chaining Gaussian distribution in a dynamic model and
mapping the emitted value to simplex.
𝛽|𝛽~𝑁(𝛽, 𝜎𝐼) (1)
Logistic normal with mean 𝛼 is used to represent document specific topic proportion, and the sequence between
models are captured by a dynamic model shown in equation (2).
814

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
𝛼|𝛼~𝑁(𝛼, 𝜎 𝐼) (2)
Therefore, the DTM sequentially ties a collection of topic models by chaining the topics and topic proportion
distribution.
More in depth description of DTM and posterior inference can be found here (Blei & Lafferty, 2006). In our study,
we choose 10 topics to understand the topic dynamics. We divide the total timestamps (10 days) into 40 six-hourly
slices. We implement our dynamic topic models using genism packages (Rehurek & Sojka, 2010).
RESULTS
This section presents the analysis of Twitter data (tweet) in response to two hurricane scenarios. We use Twitter
data from two recent hurricanes: hurricane Irma and Michael. At first, we present a spatio-temporal analysis of
tweet frequencies by showing the county wise tweet density with respect to the location of hurricane center (eye)
and its wind speed. Next, we present the topic evolution over time using a dynamic topic model. We train our
models in a personal computer having a core i5 processor and 12 GB of RAM. The training times are around 30
minutes and 45 minutes for Irma and Michael data, respectively. First, we show the results for hurricane Irma and
hurricane Michael separately, and then we compare between these two hurricane scenarios:
Hurricane Irma
Spatio-temporal Analysis of Twitter Activity
We focus our analysis on the tweets that come from different counties (total 67) in Florida. Hurricane Irma hit
Cudjoe
Figure 1. Plate Notation of Dynamic Topic Model with Variational Approximation (adopted from (Blei &
Lafferty, 2006))
815

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
Key, 20 miles north of Key West, and then Naples on September 10, 2017 (see Figure 2). Although Miami-Dade
County was not hit directly by Irma, but still it received life-threatening conditions. Figure 2 shows the county
wise heat-map representing Twitter activity on six different time-stamps indicating pre-landfall, landfall and post
landfall periods. From 3 AM, Sept. 9, 2017 to 1 AM, Sept. 10, 2017, when Irma was affecting Cuba, people from
Miami-Dade, Broward, Orange, and Hillsborough County, were highly active on Twitter (see Figure 2a and 2b).
The wind speed at these two periods were 180 mph (category 5) and 135 mph (category 4), respectively. On
(a) Pre-landfall (b) Pre-landfall
(c) Landfall (d) Landfall
(e) Post Landfall (f) Post Landfall
Figure 2. Spatio-temporal County-wise Tweet Density along with Hurricane Path and Wind Speed for Hurriane
Irma
816

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
September 10 (at around 5 AM), while Irma makes landfall at Florida key as a category 4 (wind speed around 130
mph) hurricane, twitter activity decreases compared to a pre-landfall period (see Figure 2c). Except in orange
County, Twitter activity decreases further when Irma approaches towards Naples County and made landfall as a
category 3 hurricane (wind speed around 115 mph). Mandatory evacuation was declared prior to hurricane landfall
causing one of the largest evacuation in Florida (Wong, Shaheen, & Walker, 2018); population displacement
trigged by massive evacuation causes this significant drop in twitter activities during the landfall period. On Sept.
11 and onwards (see Figure 2e), while Irma passes through Tampa as a category 1 hurricane, Twitter activity is
the lowest possibly due to power outage and communication disruptions (NERC, 2018). On September 12, while
Irma almost diminishes after passing Florida (see Figure 2f), Twitter activity starts increasing possibly because
evacuated people returned their home and power/electricity was restored in some areas (Wong et al., 2018).
Topic Evolution Over Time
Figure 3 shows the result of dynamic topic model. Before interpreting the topics, we need to understand that we
consider only the tweets coming from the affected regions (Florida). From the spatio-temporal analysis (see Figure
2), we see that Miami and Tampa regions have the highest share in the spatial distribution. The topic distributions
of the dynamic topic model also represent this spatial distribution. From Figure 3, we can see that the word
‘florida’ or ‘floridians’ have very high probability in topic #3, #4, #6, #7 and #8. Topic #1 contains the word
‘tampa’. On the other hand, topic #5 and #7 contain both ‘florida’ and ‘miami’. Now, looking at the contexts of
the tweets we can see that:
 Topic #3 and Topic #7 describe the situational awareness and updates about hurricane Irma. The words
– ‘storm’, ‘mph’,’category’,’news’, ‘florida’- have very high probability at the beginning (Sept. 5 to
Sept. 6). This is because the news channels and media were forecasting about the hurricane strength
(wind speed) by providing update about the upcoming hurricane that might potentially hit the Florida. In
the next two days (Sept. 7 to Sept. 8), words like ‘watch’ and ‘advisory’ have high probability. It is
because people were recommended to watch about hurricane advisory as the landfall time (Sept. 10) was
coming closer. Coming next on the landfall day (Sept. 10), words like ‘key’, ‘move’ have high
probability, representing the landfall location of ‘florida key’. After the landfall on Florida key, on Sept.
13 words like ‘break’, ‘jose’ represent about the hurricane damage and another hurricane Jose that was
formed in the Atlantic Ocean as a category 4 hurricane.
 Topic #8 and Topic #9 describe about hurricane evacuation and return activity. On Sept. The words –
‘cause’, ‘south’, ‘florida’,’days’, ‘house’ have high probability in the beginning (Sept. 5 to Sept. 6). This
is because the south Florida was under the mandatory evacuation zone. Between Sept. 6 to Sept. 7, the
words- ‘come’, ‘bring’,’center’, ‘want’- have high probability indicating necessary preparation for
staying home or evacuation. Coming next on Sept. 8 to Sept. 10 (landfall), the words– ‘go’,’evacuate’,
‘coast’- have high probability indicating evacuation and going to a safer place from the high risk coastal
area. After the landfall, on Sept. 14 and afterward, the words- ‘back’,’home’ have high probability
indicating coming back to home
 Topic #4 indicate the disruptions caused by the hurricane. Words like ‘school’, ‘close’ represent that the
schools started closing 5 days before landfall (Sept. 5). The words ‘flood’ on Sept. 12 indicate the
flooding due to hurricane. The increasing probability of the word, ‘power’ indicates that power outage
disruption became higher after the landfall (Sept. 10) and onwards.
 Topic #6 is about the sentiments about the ongoing events during hurricane. The topic words – ‘pray’,
‘amp’ (ain’t my problem),’US’, ‘Cuba’, ‘please’, ‘safe’, ‘thank’- represent public sentiment during
different period of hurricane Irma.
 Topic #1 and topic #5 represent local topic on Tampa and Miami respectively. The word ‘miami’ have
high probability before Irma made landfall to Florida key, and the word ‘tampa’ had high probability
while hurricane was approaching towards Tampa. The probability of the words like ‘relief’ and ‘help’
became high on Sept. 14 (post landfall) in topic #1.
 Topic #0 and topic #2 represents the general topics about day-to-day things and concerns (words such as
‘worry’, ‘need’, ‘snack’).
817

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
Figure 3. Change in Word probability over Time (hurricane Irma)
Hurricane Michael
Spatio-temporal Analysis of Twitter Activites
Similar to Irma, we plot county wise tweet density with respect to the hurricane center. Figure 4(a) shows that the
Hurricane Michael is approaching to Florida on Oct. 9 (it was generated on Oct. 7), 2018 at a wind speed of 100
mph and not that much people were tweeting about Michael at that time. But, Figure 4(b) depicts that people of
south Florida and some counties of North Carolina started talking about Michael over twitter and the hurricane
was approacing very qucikly with 130 mph wind speed at the early morning of Oct. 10, 2018. Then, the hurricane
made landfall with 160 mph wind speed (Figure 4c) on Florida Panhandle during afternoon of the same day and
the tweet density increased in a significant number from Florida and North Carolina state. On the following day
(Oct. 11, 2018, morning) of landfall, from Figure 4(d), its obviuos that the hurricane became weak as the wind
speed went down to 40 mph while passing over Georgia and people became less active in twitter as the tweet
818

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
density decreased. But, from 4(e), at the evening of Oct. 11,2018, Michael streangthens as the wind speed went
up to 60 mph and was traveling over South Carolina towards North Carolina and again the tweet density increased
from Florida and North Carolina.
So, it seems like most of the tweets were generated from these two states as the south and central Florida was not
affected by the Hurricane and North Carolina was not expected to hit by Michael strongly, hence people were
spreading information over twitter instead of evacuating or preparing for hurricane. But, we did not observe
significant tweets from Georgia and South Carolina as it can be assumed that most of the people evacuated from
these states. Lastly, Virginia people also did not tweet a lot because this state was not expected to be hit by
Michael, hence it seems people of Virginia was not concerned about this Hurricane in twitter although the state
was affected.
(a) Pre-landfall (b) Pre-landfall
(c) Landfall (d) Post Landfall
(e) Post Landfall
Figure 4. Spatio-temporal County-wise Tweet Density along with Hurricane Path and Wind Speed for Hurricane
Michael
819

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
Topic Evolution Over Time
From Figure 5, a heat-map of dynamic topics found, we can observe the word probability of the most frequent
topics over 10 day’s timestamp, which are discussed in following-
 In topic #0, people discussed more about state emergency while the hurricane generated and predicted
that Florida could be hit tomorrow by Michael on the previous day (Oct. 09,2018) of the landfall.
 Topic #2 is related to hurricane Florence, a previous hurricane which made its landfall one month before
of Michael. The discussion was focused to help the victims of hurricane Florence by fundraising
program.
 Topic #3 indicates that people were concerned about witnessing hurricane for another time and the
moving of people or evacuation just before the landfall as Michael had a very small warning phase.
 In topic #2, ‘relief’ word co-appeared with ‘florence’ before Michael formation; which indicates people
were concerned to help hurricane Florence victims just before Michael occurrence. But, for topic #3,
‘relief’ is found after Michael landfall with ‘democrat’ which means people expected help from
government after the landfall.
 From topic #4, ‘tropical’ and ‘storm’ words found together just before two days of landfall mentioning
that Michael was not a major hurricane (but, it hit as a category 5 hurricane). It reflects the very short
noticed characteristics of Michael.
 Topic #5 shows that people were searching for information about expected affected area (Tallahassee)
before landfall. Besides, after the landfall they were also interested to know about the busted condition
of affected area (Panama City).
 Topic #8 is also about getting live updates of affected areas (Florida Panhandle).
 Topic #9 indicates that people were discussing about the forecast of Michael before the landfall and
about the wind speed of the hurricane just after the landfall. In the same topic, after landfall people were
concerned about the social connections to recover from the disaster and the power outage issue.
 Topic #3, topic #4 and topic #7 showed some political words like ‘democrat’, ‘trump’,’avenatti’
(renowned attorney) and ‘presidential’ before and after the landfall. The reason behind this is the US
midterm election was held within few weeks of hurricane Michael landfall.
 Topic #0, topic #1, topic #5 and topic #8 contain ‘Florida’, ‘panhandle’, ‘tallahassee’; topic #4 and topic
7# contain ‘north carolina’, ‘nc’ which depicts the more affected locations of Michael.
 Situational awareness and live updates words like ‘forecast’, ‘wind’, ‘mph’ and ‘emergency’ are found
in topic #0 and #9 for hurricane Michael.
 Topic #1 and #6 reflects the general topics about daily life issues and concerns by words like ‘advisory’,
‘defrosting’, ’weekend’, played etc.
820

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
Figure 5. Change in word probability over time within a topic (hurricane Michael)
Comparison Between Two Hurricane Scenarios
Hurricane Irma (2017) and Michael (2018) are among the strongest hurricanes in the US history. Although both
of these two hurricanes made landfall in Florida, Irma as category 4 and Michael as category 5, the hurricane
paths and the affected regions are different. Hurricane Irma mostly affected Florida; on the other hand, Michael
affected only a part of the Florida, North Carolina, Virginia, and Georgia. Therefore, these two hurricanes have
two different contexts and scenarios. The objective of this section is to identify the similarities and dissimilarities
in the spatio-temporal activity distribution and topic evolution for these two hurricane scenarios. From the above
findings between hurricane Irma and Michael, we notice the following similarities and dissimilarities:
821

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
Similarities
 During both hurricanes, Miami-Dade, Broward and Orange counties were highly active in Twitter though
these counties were not affected by hurricane Michael. Previous experience of multiple hurricanes might
be the reason of their increased activity.
 Both hurricanes represent some topics that reflect the affected locations with high density of Twitter
activity found in spatial distributions. For example: for hurricane Irma topic #1 contains the word
‘tampa’; topic #5 and #7 contain ‘miami’. For hurricane Michael, topic #0, topic #1, topic #5 and topic
#8 contain ‘Florida’, ‘panhandle’, ‘tallahassee’; topic #4 and topic 7# contain ‘north carolina’, ‘nc’.
 There are some topics which overlap, such as situational awareness and live updates. Topic #3 and topic
#7 from hurricane Irma and topic #0 and #9 from hurricane Michael reflect similar issues.
 In both cases, power outage related topics became active after the landfall, Topic #4 for hurricane Irma
and topic #9 for hurricane Michael.
 Likewise, hurricane relief related topics became active after the landfall, topic #1 for Irma, topic #3 for
Michael.
Dissimilarities
 Spatio-temporal density map shows higher density along the locations of the hurricane Irma’s path. For
hurricane Michael, we found less density at the counties in Georgia and North Carolina, although these
locations were along the path of Michael. Hurricane Irma’s path was along the coastline where these
areas have experienced many hurricanes/storms. On the other hand, Michael’s path was along the land-
area, where some of the regions were not affected by previous hurricanes. This might be the possible
reason for lower participants in social media activities from these areas.
 Topic related to hurricane evacuation became highly active at the evacuation period of hurricane Irma.
However, we do not find much attention related to evacuation in the topic dynamics of hurricane Michael.
Hurricane Irma prompted massive evacuation in Florida, whereas for Michael, evacuation period was
short noticed. This might be the possible reason for less attention to evacuation topic for Michael.
 In both hurricane scenarios, the topic dynamics show a topic related to another approaching or dissipating
hurricane. However, Irma’s topic dynamic the word, ‘Jose’ became active after Irma’s landfall. On the
other hand, in Michael’s topic dynamic, the word ‘Florence’ became active at the beginning of the topic
dynamic.
 In hurricane Michael’s topic dynamic some topics contain political words as the US midterm election,
2018 was held within the next few weeks of the landfall (the word ‘democrat’ in topic #3, the wor d
‘trump’ in topic #4). However, Irma’s topic dynamics do not contain any political words, probably due
to the fact that it occurred in a non-election year.
CONCLUSIONS
Learning from experiences has been a major focus by disaster scholars and we need enough empirical evidence
that will facilitate people-oriented future hazard mitigation scenarios. Online social networks are ever-growing
and large scale social media data that are publicly available can provide such opportunities in terms of effective
crisis communication and logistics of disaster response and recovery. Recent studies have used data mining
techniques such as topic models to understand the social media user need and sentiment during disasters. However,
these studies cannot answer how such communication patterns in disaster prone communities evolve over time
during different phases of a major disaster from warning to recovery, and if there exist common trends in the topic
dynamics for different hurricane scenarios. This study presents an analysis of time dependent communication
patterns during two different hurricane scenarios. To the best of authors’ knowledge, our study is the first in the
literature presenting an analysis of topic dynamics considering two distinct major hurricane scenarios.
In this study, we use large scale Twitter data that we collected during two recent hurricanes: hurricane Irma (2017)
and hurricane Michael (2018). At first, we present a spatio-temporal analysis of the Twitter activity along different
counties along the hurricane path, and then we use Dynamic Topic Model to understand the topic dynamics for
each hurricane using the tweet texts from the affected regions. This was followed by a comparative analysis of
the commonalities and differences in such dynamics between two hurricane scenarios. Topics related to situational
awareness, live updates and disruptions are common across both hurricane scenarios. On the other hand,
difference in hurricane path and characteristics of the affected regions lead to dissimilarities in the topic dynamics.
Although our work can reveal the dynamics of topics, it has some limitations. In the dynamic topic models, we
set the number of topics beforehand due to associated computation challenges and the time slices we used are
discrete in nature. However, it may be possible that topic evolutions are continuous in nature, some topics may
822

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
have dissipated in the process, and new topics emerged. However, in our current approach, the number of topics
are fixed and hence we cannot model this behavior with our current approach. The computational and memory
requirement increase quickly with the increase in time slice thus can pose a problem in determining appropriate
resolution of time granularity. Future studies can consider the continuous time topic dynamics considering a
flexible number of topics over the time during a disaster. Approaches like Brown ian motion based continuous
time DTM (C. Wang, Blei, & Heckerman, 2012) or non-Markov assumption based continuous topic model (X.
Wang & McCallum, 2006) can be can be adopted in the future studies.
REFERENCES
Beiró, M. G., Panisson, A., Tizzoni, M., & Cattuto, C. (2016). Predicting human mobility through the
assimilation of social media traces into mobility models. EPJ Data Science, 5(1).
Blei, D. M., & Lafferty, J. D. (2006). Dynamic topic models. In Proceedings of the 23rd international
conference on Machine learning (pp. 113–120).
Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent dirichlet allocation. Journal of Machine Learning
Research, 3(Jan), 993–1022.
Brown, J. (2015). Is Social Media the Key to Effective Communication During Campus Emergencies?
Government Technology.
Cecile Wendling, Jack Radisch, S. J. (2013). The Use of Social Media in Risk and Crisis Communication.
Dong, X., Mavroeidis, D., Calabrese, F., & Frossard, P. (2015). Multiscale event detection in social media. Data
Mining and Knowledge Discovery, 29(5), 1374–1405.
Imran, M., Castillo, C., Diaz, F., & Meier, P. (2013). Practical Extraction of Disaster-Relevant Information from
Social Media. In Proceedings of the 22nd International Conference on World Wide Web (pp. 2012–1024).
Kryvasheyeu, Y., & Chen, H. (2015). Performance of Social Network Sensors During Hurricane Sandy. PLoS
One, 10(2), e0117288. 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.
Lachlan, K. A., Spence, P. R., Lin, X., Najarian, K., & Del Greco, M. (2016). Social media and crisis
management: CERC, search strategies, and Twitter content. Computers in Human Behavior, 54, 647–652.
Retrieved from http://dx.doi.org/10.1016/j.chb.2015.05.027
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
NERC. (2018). Hurricane Irma Event Analysis Report. Www.Nerc.Com.
Noyes, D. (2019). The Top 20 Valuable Facebook Statistics – Updated July 2019. Zephoria. Retrieved from
https://zephoria.com/top-15-valuable-facebook-statistics/
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/
Pan, B., Zheng, Y., Wilkie, D., & Shahabi, C. (2013). Crowd sensing of traffic anomalies based on human
mobility and social media. In Proceedings of the 21st ACM SIGSPATIAL International Conference on
Advances in Geographic Information Systems (pp. 344–353).
Patel, V. H. (2018). Topic modeling using latent dirichlet allocation on disaster tweets.
Rashidi, T. H., Abbasi, A., Maghrebi, M., Hasan, S., & Waller, T. S. (2017). Exploring the capacity of social
media data for modelling travel behaviour: Opportunities and challenges. Transportation Research Part
C: Emerging Technologies, 75, 197–211. Retrieved from
http://linkinghub.elsevier.com/retrieve/pii/S0968090X16302625
Rehurek, R., & Sojka, P. (2010). Software Framework for Topic Modelling with Large Corpora. In Proceedings
of the LREC 2010 Workshop on New Challenges for NLP Frameworks (pp. 45–50). Valletta, Malta:
ELRA.
Resch, B., Usländer, F., & Havas, C. (2018). Combining machine-learning topic models and spatiotemporal
analysis of social media data for disaster footprint and damage assessment. Cartography and Geographic
Information Science, 45(4), 362–376.
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.
Roy, K. C., Hasan, S., Sadri, A. M., & Cebrian, M. (2020). Understanding the efficiency of social media based
crisis communication during hurricane Sandy. International Journal of Information Management, 102060.
Sadri, A. M., Hasan, S., Ukkusuri, S. V., & Cebrian, M. (2017a). Understanding Information Spreading in
Social Media during Hurricane Sandy: User Activity and Network Properties. ArXiv Preprint
823

Roy et al. Dynamics of Social Media Communications in Major Hurricanes
WiP Paper –Social Media for Disaster Response and Resilience
Proceedings of the 17th ISCRAM Conference – Blacksburg, VA, USA May 2020
Amanda Lee Hughes, Fiona McNeill and Christopher Zobel, eds.
ArXiv:1706.03019. Retrieved from http://arxiv.org/abs/1706.03019
Sadri, A. M., Hasan, S., Ukkusuri, S. V., & Suarez Lopez, J. E. (2018). Analysis of social interaction network
properties and growth on Twitter. Social Network Analysis and Mining, 8(1), 1–13.
https://doi.org/10.1007/s13278-018-0533-y
Sadri, A. M., Hasan, S., Ukkusuri, S. V, & Cebrian, M. (2017b). Crisis Communication Patterns in Social
Media during Hurricane Sandy. Transportation Research Record, 0361198118773896.
Sadri, A. M., Ukkusuri, S. V., Lee, S., Clawson, R., Aldrich, D., Nelson, M. S., … Kelly, D. (2017). The role of
social capital, personal networks, and emergency responders in post-disaster recovery and resilience: a
study of rural communities in Indiana. Natural Hazards, 1–30. https://doi.org/10.1007/s11069-017-3103-0
Shaban, H. (2019). Twitter reveals its daily active user numbers for the first time. The Washington Post.
Retrieved from https://www.washingtonpost.com/technology/2019/02/07/twitter-reveals-its-daily-active-
user-numbers-first-time/?noredirect=on&utm_term=.1aded2c5a7f2
Slamet, C., Rahman, A., Sutedi, A., Darmalaksana, W., Ramdhani, M. A., & Maylawati, D. S. (2018). Social
Media-Based Identifier for Natural Disaster. IOP Conference Series: Materials Science and Engineering,
288(1). https://doi.org/10.1088/1757-899X/288/1/012039
Stieglitz, S., Mirbabaie, M., Ross, B., & Neuberger, C. (2018). Social media analytics--Challenges in topic
discovery, data collection, and data preparation. International Journal of Information Management, 39,
156–168.
Subba, R., & Bui, T. (2017). Online Convergence Behavior, Social Media Communications and Crisis
Response: An Empirical Study of the 2015 Nepal Earthquake Police Twitter Project. Proceedings of the
50th Hawaii International Conference on System Sciences (2017), 284–293.
https://doi.org/10.24251/hicss.2017.034
Ukkusuri, S., Zhan, X., Sadri, A., & Ye, Q. (2014). Use of Social Media Data to Explore Crisis Informatics.
Transportation Research Record: Journal of the Transportation Research Board, 2459, 110–118.
Wang, C., Blei, D., & Heckerman, D. (2012). Continuous time dynamic topic models. ArXiv Preprint
ArXiv:1206.3298.
Wang, Q., & Taylor, J. E. (2016). Patterns and limitations of urban human mobility resilience under the
influence of multiple types of natural disaster. PLoS ONE, 11(1), 1–14.
Wang, X., & McCallum, A. (2006). Topics over time: a non-Markov continuous-time model of topical trends.
In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data
mining (pp. 424–433).
Wang, Z., Ye, X., & Tsou, M. H. (2016). Spatial, temporal, and content analysis of Twitter for wildfire hazards.
Natural Hazards, 83(1), 523–540.
Wong, S., Shaheen, S., & Walker, J. (2018). Understanding Evacuee Behavior: A Case Study of Hurricane
Irma.
Xiao, Y., Huang, Q., & Wu, K. (2015). Understanding social media data for disaster management. Natural
Hazards, 79(3), 1663–1679. https://doi.org/10.1007/s11069-015-1918-0
Yang, M., Mei, J., Ji, H., Zhao, Z., Chen, X., & others. (2017). Identifying and Tracking Sentiments and Topics
from Social Media Texts during Natural Disasters. In Proceedings of the 2017 Conference on Empirical
Methods in Natural Language Processing (pp. 527–533).
824