The document describes a methodology for analyzing the social impact of severe weather events using Twitter data. The researchers analyzed tweets during a 2011 heat wave in Italy to see how social media usage correlated spatially and temporally with weather patterns. Key findings included that conversation topics on high heat days centered around heat more than usual, and that geographic names in tweets matched heat-impacted areas. The open-source R code allows reproducing the analysis to understand community resilience to weather emergencies via social media insights.
ClimateScope: a Google Earth storytelling of climate change
Social Media and Severe Weather: Tracking a Heat Wave on Twitter
1. Social Media and
severe weather events: an heat
wave footprint on Twitter
Disaster 2.0 - MasterClass
Bruxelles 17/01/2013
Valentina Grasso - grasso@lamma.rete.toscana.it
Alfonso Crisci - a.crisci@ibimet.cnr.it
2. The 5 C of Social Media
• contents (UGC)
• conversation
• connection
• collaboration
•A community behaviour
- big lens on human
- Extract useful information from Big Data
3. SM e weather services
Plenty of weather content on SM and mobile APP
- weather is a common conversation topic
- personalization of weather forecast
- local dimension
- weather is a special case of "emergency" issue
4. Weather as emergency issue
main features
•FREQUENT: vs to other emergencies #fires
•FAMILIAR: people deal with weather daily #earthquake
•PREDICTABLE: important for warnings #chemical
•LOCATED: specific spatial and temporal
dimension
#nuclear
#disaster
#health
#terrorism
who
when
where
5. Building resilient
communities
Weather as an operational context where community may
increase "resilience" attitude.
In emergency "behaviours" modulate "impacts" on society.
If I'm aware and prepared I act responsibly.
US tornado warning:
people get used to "weather
warnings" and they learnt to be
proactive in protection.
6. Changing climate - changing
awareness
In Italy and Europe in the last 10 years climate change
made us more exposed to extreme weather events -
"preparedness"
Tornado hits: US - Italy 1999-2009
Geographical spreading and
magnitude of events
are important
for awareness
7. Working on Italian Twitter-
sphere
Weather event: early heat wave on 5-7 April 2011
Research objectives
• investigate time/space
coherence between the
event extension and its
social footprint on Twitter
• semantic analysis of
Twitter stream on/off
peaks days
8. Heat wave as a good case
Emergency as consequence of "behaviour"
Communication is key: "how to act"
9. Severe weather definition
Heat wave: it's a period with persistent T°
above the seasonal mean. Local definition
depends by regional climatic context.
Severe weather
refers to any dangerous
meteorological phenomena
with the potential to cause
damage, serious social
disruption, or loss of
human life.[WMO]
Types of severe weather
phenomena vary,
depending on the latitude,
altitude, topography, and
atmospheric conditions.
Ref:
http://en.wikipedia.org/wiki/Severe_weather
10. Target and Products
Consorzio LaMMA - CNR Ibimet developed a methodology and a set of
products to quantitative evaluate the social impact of weather related
events.
Products: Stakeholders:
•DNKT metric •forecasters
•association of the time vector •institutional stakeholders
(DNKT) and a time coupled •EM communities
gridded data stack
•media agents
•spatial associative map
•semantic analysis Twitter Target
stream:
Detect areas where it's worth
- clustering
focusing attention, also for
- word clouds communication purpose.
11. Data used
Heat wave period considered (7-13 April 2011)
Social
- Using Twitter API key-tagged (CALDO-AFA-SETE) 6069
tweets collected through geosearch
service for italian area.
- Retweets and replies included (full volume stream)
Climate & Weather (7-10 April 2011)
- Urban daily maximum T°
- Daily gridded data (lon 5-20 W lat 35-50)
WRF-ARW model T°max daily data (box 9km)
12. Twitter metric
DNKT - "daily number of key-tagged tweets"
*
*
*
DNKT shows time coherence with daily profiles of areal averaged temperature
*Critical days identified as numerical neighbour of peaks (7-8-9-April):
social "heaty days"
13. Geographic associative maps
Semantic based social stream in
1D * time space (DNKT)
Weather informative
layers in 2D time* space
Linear
Association
Statistically Geographic
based Associative
Verifier Map
by pixel (2D space)
14. Impacted areas
This is not a Twitter map
It's a weather map at
X-rays:
Twitter stream
is used as a
"contrast medium"
to visualize impacted
areas.
15. A question of shape
weather phenomena and
peak social/communication streams
as "analogue" time delayed
information waves
start decline
time
17. Semantic analysis
- Corpus creation
DNKT classification by heat-wave peak days:
heat days ( 7-8-9 April) no-heat days (6-10-11 April).
- Terms Word Clouds (min wd frequency>30)
heat days vs no-heat days
Clustering associated terms
Term frequency ranking comparison
- Hashtag Word Clouds heat days
heat days vs no-heat days
R Stat 15.2 Packages used:
tm (Feinerer and Hornik, 2012) & wordcloud (Fellows , 2012)
23. Semantic: some results
On peak days:
- widening of lexical base during "heat critical days" - heat as a
conversation topic
- ranking of terms (i.e.:adjectives as "troppo"!) is useful to detect change in
communication during climatic stress
- geographic names appears in terms and hashtags wordsets ( "#milano" !).
This fits with recent researches on "social media
contribution to situational awareness during
24. conclusions
- Methodology for a social "x-rays" of
a weather event: Twitter stream as
a "contrast medium" to understand
the social impact of severe weather
events
- Methodology social geosensing is able
to map severe weather impacts and
overcome the weakening in
geolocation of social messages and
eliminate the bias due to "social
fakes".
Weather as a key emergency context where it's worth working on
community resilience - also with the help of social insightful contents.
25. Reproducible R code
Github Master class socialsensing Code & Data
https://github.com/alfcrisci/socialgeosensing.git
Wiki Recipes in
https://github.com/alfcrisci/socialgeosensing/wiki
weather channel app is the fifth most downloaded app of all time - http://www.huffingtonpost.com/2012/03/05/most-popular-free-apps-iphone-ipad_n_1321852.html
An associative map could be done by the level of significance for the linear association calculated between a generic time vector (DKTN) and a time coupled gridded data stack (Weather time consecutive layers). The significant impacted areas emerges geographically where the social unidimensional signal is coherent and linearly correlated with the correspondent time vector of pixel/box values inside the grid. Tree significance levels are generally considered in function to the p.value of null significance statistical test (NO:p >0.1 Weak :p >0.05 Strong : p<0.05)
Method 's strength is due to the validity of assumption that exist a linear association between weather process and " verbosity " concerning words semantically linked to weather sensations felt by people. This is the basis of geographical matching between social streams and weather data. Method 's gain is also the overcoming of the bias due to "false tweet" and population density lackness.