With the surge in digital information systems, there is data deluge from various sources that can be analyzed and integrated to produce relevant, reliable and actionable information, preserving low-level details and surfacing essential information for better decision making.
We employ multi-modal data (i.e., unstructured text, gazetteers, and imagery) for an aggregate level analysis and location-centric demand/request matching in the context of disaster relief. After classifying the 'Need' expressed in a tweet (the WHAT), we leverage OpenStreetMap to geolocate that 'Need' on a computationally accessible map of the local terrain (the WHERE) populated with location features such as hospitals and housing. Further, our novel use of flood mapping based on satellite images of the affected area supports the elimination of candidate resources that are not accessible by road transportation.
The resulting map-based visualization serves two levels of users. A community level user (first-responders) can visualize aggregated summary of a selected geographical area and an individual level user can identify current needs and available resources in their geographic proximity. Additionally, our pluggable modularized pipeline (DisasterRecord) is extensive and addition functionalities can be layered on top of the map. The integration of disaster-related tweets, imagery and pre-existing knowledge-base resources (gazetteers) reduce decision-making latency and enhance resiliency by assisting decision-makers and first responders involved with relief effort coordination.
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
Multi-Scale and Multi-Modal Streaming Data Aggregation and Processing for Decision Support during Natural Disasters
1. Multi-Scale and Multi-Modal Streaming Data
Aggregation and Processing for Decision
Support during Natural Disasters
rebrand.ly/HazardSEE
S
1
DisasterRecord: Disaster Response and Relief Coordination Pipeline
Shruti Kar
M.S. Thesis Defense
COMMITTEE MEMBERS:
Dr. Krishnaprasad Thirunarayan (ADVISOR)
Dr. Amit Sheth
Dr. Valerie L. Shalin
MENTOR:
Hussein S. Al-Olimat
Fall 2018
3. Problem Definition
3
Given a disaster event with a known spatial context, can we extract the
knowledge from multimodal data including targeted crowd-sourced data,
Twitter stream, Satellite imagery and pre-disaster data and intertwine them
with locations on the map to provide decision-support for real-time situational
awareness?
Part of this work was published in ACM SIGSPATIAL ARIC 2018 Workshop.
Shruti Kar, Hussein S. Al-Olimat, Krishnaprasad Thirunarayan, Valerie Shalin, Amit Sheth, and Srinivasan Parthasarathy. D-record: Disaster
response and relief coordination pipeline. In Proceedings of the ACM SIGSPATIAL International Workshop on Advances in Resilient and
Intelligent Cities (ARIC 2018). Association for Computing Machinery, 2018.
4. Location-centric Ontology
Use-Case
4
Eg: “I am stuck at ‘XYZ’.. Help me!”
Need Location OSM Locations
Need Matching
Need Classification Location Affordability
Need type? Location type?
Flood Mapping
6. Data Sources
6
Data
Streaming and Crowd-
sourced data
Satellite imageryPre-disaster data
Excel sheets from volunteersTwitter
Training set Twitris collection
CrisisLexT26
CrisisNLP
Chennai Flood in 2015
Houston Flood in 2016
(http://rebrand.ly/reliefspreadsheet)
8. Data Sources - Streaming data
8
● Twitter Data
LOCATIONS NEEDS
Extract Locations names from Text
LNEx
Extract Needs type from Text
9. Disaster Text Classification - Related Literature
9
▪ Olteanu et al., 2015 ▪ CREES
▪ Cameran et al., 2012 - Binary classifier for Infrastructure damage
10. Location Name Extraction
10
● User profile location vs location
name mentioned?
● Location Name mentioned in
unstructured text streams
● Intertwine with geo
coordinates
LOCATIONS
(github.com/halolimat/LNEx)
11. Need Classification
11
Our Need Class CrisisLexT26,CrisisNLP Classes
Shelter/Food/Supplies Need
Donation_needs_or_offers_or_volunteering_services,
displaced_people_and_evacuations
Medical/Rescue Help Need
Missing_trapped_or_found_people, deaths_reports,
injured_or_dead_people, affected_people
Twitter
Training set Twitris Collection
CrisisLexT26
CrisisNLP
Chennai Flood in 2015
Houston Flood in 2016
12. Need Classification
12
● Preprocessing - Stemming, case folding and removal of noise.
● Semantics - Feature Engineering
■ TF-IDF vectors
■ Gensim’s word2vec embeddings
● Class Imbalance - SVM-SMOTE synthetically oversampled the minority class
● Gradient Boosting Classifier
Feature set
Without SVM-SMOTE With SVM-SMOTE
RF SVM GB RF SVM GB
tf-idf 0.50 0.57 0.58 0.55 0.61 0.63
word2vec 0.53 0.59 0.60 0.58 0.66 0.68
tf-idf + word2vec 0.60 0.64 0.71 0.67 0.72 0.77
14. Data Sources - Crowd Sourced data
14
● Excel sheets from volunteers
(http://rebrand.ly/reliefspreadsheet)
15. Data Sources - Crowd Sourced data
15
● Excel sheets from volunteers
(http://rebrand.ly/reliefspreadsheet)
...
16. Data Sources - Pre-Disaster
16
Houston
● OpenStreetMap Location features
17. Location-centric Ontological Modeling
17
▪ Matching Need and availability
▪ Event Classification
Event Ontology &
Situational Data
Seakers / Providers
Need Classification
Matching Need
and Availability
18. Location-centric Ontological Modeling
19
▪ Matching Need and availability - Relationships with Location features
▪ Event Classification - Concepts with Lexicon based features
Event Ontology &
Situational Data
Seakers / Providers
Need Classification
Matching Need
and Availability
Concepts with Lexicon
features
Location features
19. Location-centric Ontological Modeling
20
▪ Matching Need and availability - Relationships with Location features
▪ Event Classification - Concepts with Lexicon based features
Event Ontology &
Situational Data
Eg: “I Broke my leg..
Help me!” Kindred Hospital
Medical Need Medical Help Availability
From Lexicon features From Location features
20. Disaster-centric Ontologies - Related Literature
22
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
▪ CrisisNLP
▪ CrisisLex
▪ CREES
▪ MOAC - the Management of Crisis
▪ HXL - Humanitarian eXchange Language
▪ SOCC - SOCS crisis ontology
▪ EDXL-RESCUER
▪ SMEM - Social Media and Emergency Management
▪ Wang et al., 2014
21. Disaster-centric Ontologies - Related Literature
23
CrisisNLP, CrisisLex, CREES - Lexicons for classification
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
22. Disaster-centric Ontologies - Related Literature
24
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
MOAC - the Management of Crisis Vocabulary
Answers - WHO? WHAT? WHERE?
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
HXL - Humanitarian eXchange Language
Framework for Interoperability
23. Disaster-centric Ontologies - Related Literature
25
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
SOCC - SOCS crisis ontology
Extended HXL and MOAC
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
EDXL-RESCUER
Data exchange with legacy systems
24. Disaster-centric Ontologies - Related Literature
26
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
SMEM - Social Media and Emergency
Management
Combines crisis domain with social
media
Competency Questions
▪ Coverage - depth and breadth
▪ Spatial details
▪ Temporal details
▪ Thematic details
▪ Response details
Wang et al., 2014- Lacks hierarchical
relationships between concepts
25. Location-centric Ontological Modeling
27
▪ Uses information from the past disaster and risk reports from ACAPS.
▪ Distinction between response and relief phase of disaster
▪ Concepts
□ Needs - Topic modeling
□ Shelter/Food/Supplies
□ Medical/Rescue Help
□ Availability - OpenStreetMap features
▪ Relevance scores of each term to the Need class - used while Text Classification
Relationship : Available As
26. Location-centric Ontological Modeling
28
▪ Topic Modeling of Concepts of Interest:
Shelter/Food/Supplies
relief
donate
food
fund
assist
shelter
evacuate
volunteer
money
blood
contribute
Medical/Rescue Help
death
injure
missing
collapse
trapped
rubble
police
found
search
killed
blast
▪ Collected, cleaned and processed data.
▪ Latent Dirichlet Allocation (LDA) - A probabilistic
topic model
▪ p(topic t | document d) x p(word w | topic t)
27. Location-centric Ontological Modeling
29
▪ Topic Modeling of Concepts of Interest:
Completely Justified Quiet Justified To be reviewed Remove
rubble Collapse help deadli
police Rescue baby control
army Team old find
relief Caution water emergency
power hospital response
devast blast benefit
aid clean force
evacuation loss factory
35. Need Matching - Related Literature
37
▪ Purohit et al., 2014 - uses hand-crafted rules to match “Seekers” and “Suppliers”.
▪ Currion et al., 2012- Uses structures input from users and utilizes moderators in the
loop to match.
▪ Palomares et al.,2015- quantifies the degree of impact of a disaster and prioritizes the
matching accordingly.
▪ Murali et al., 2016- provides online and offline solution for requesting, providing and
coordinating resources.
▪ Limitations:
□ Faulty assumption that all routes are available for matching during a disaster.
□ Matching problem being solved for a different level of granularity
36. Need Matching
38
Seakers / Providers
Need Classification
+
Need
Concepts
Location
Affordability
+
Need Locations Available Locations
Flood Mapping
Need Matching
37. Data Sources - Satellite Imagery
39
● Flood mapping
Liang et al. "Human-Guided Flood Mapping: From Experts
to the Crowd." WWW 2018.
38. Flood Mapping
40
▪ Liang et al., 2018 - interpolates satellite imagery on a map in order to determine flooded
geo-coordinates
□ Eliminate flooded OSM Locations
□ Prune away flooded/closed routes
41. Data Visualization - Related Literature
43
▪ Wang et al., 2014 - News report
▪ Cameron et al., 2012 - Infrastructure
needs
42. Data Visualization - Related Literature
44
▪ MacEachren et al., 2011 - places geo-tagged tweets
on map, Heatmap of tweet frequency
(spatially), Ranked, sorted relevant tweets
▪ Neis et al., 2010 - User’s structured
Input.
▪ Junior et al., 2015 - showed effectiveness of
layered visualization systems.
45. Data Visualization
47
▪ Community-Level of Analysis
□ Location-specific textual data
□ Categorizing tweets to
need types
□ Aggregating tweets with
respect to location vicinity
46. Data Visualization
48
▪ Community-Level of Analysis
□ Location-specific image features
□ Filtering for Flooded images
□ Detecting objects of
interest
□ Location-specific available help
□ Location-specific thematic profile
47. Data Visualization
49
▪ Individual-Level of Analysis
□ Flooded areas around
their vicinity
□ The possible available
help around their location
matching their needs
□ Route guidance of non-
flooded routes for relief
workers and individuals
seeking help.
48. Results and Evaluation
50
Datasets Chennai Flood 2015 Houston Flood 2016
Original number of tweets 169,838 415,057
Locations extracted with LNEx 85,564 (23,401 needs) 241,684 (60,421 needs)
OSM-featured Location 1395 2,826
49. Conclusion and Future Work
● Our multi-scale and multi-modal streaming data aggregation and processing system supports
individual and aggregated level analysis for better-informed decision support during natural
disasters.
● A domain-specific location-centric event ontology is crucial for situation awareness and
disaster response.
● DisasterRecord is a modularized pipeline enabling multi-modal data as input demonstrates
Need-offer matching.
● Finer-grained classifier can be designed to do flexible and specific matching.
● Other background knowledge such as Weather data, storm surge model can help in
preparedness in addition to response.
● Online and offline infrastructure support can be provided.
● Studying Geo-tagged Locations in sync with Location Name mentions in the text to infer trust.
● Weighing the edges of road network with flood mapping to reduce ETAs while solving matching
problem.
51
50. References
● Muhammad Imran, Prasenjit Mitra, and Carlos Castillo. Twitter as a lifeline: Human-annotated twitter corpora for nlp of crisis-related messages. arXiv preprint
arXiv:1605.05894, 2016.
● SOCS:Social Media Enhanced Organizational Sensemaking in Emergency Response. http://knoesis.org/projects/socs. Accessed: 2018-02-22.
● Hussein S. Al-Olimat, Krishnaprasad Thirunarayan, Valerie Shalin, and Amit Sheth. Location name extraction from targeted text streams using gazetteer-based
statistical language models. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1986–1997. Association for Computational
Linguistics, 2018.
● Rebeca Barros, Pedro Kislansky, Láis Salvador, Reinaldo Almeida, Matthias Breyer, and Laia Gasparin Pedraza. Edxl-rescuer ontology: Conceptual model for semantic
integration. In ISCRAM, 2015.
● Shreyansh P Bhatt, Hemant Purohit, Andrew Hampton, Valerie Shalin, Amit Sheth, and John Flach. Assisting coordination during crisis: a domain ontology based
approach to infer resource needs from tweets. In Proceedings of the 2014 ACM conference on Web science, pages 297–298. ACM, 2014.
● Paul Currion, Chamindra de Silva, and Bartel Van de Walle. Open source software for disaster management. Communications of the ACM, 50(3):61–65, 2007.
● Zhi-Hua Hu. A container multimodal transportation scheduling approach based on immune affinity model for emergency relief. Expert Systems with Applications,
38(3):2632–2639, 2011.
● Paulo Sim̃ oes J́unior, Renato Novais, aninha Vieira, Laia G Pedraza, Manoel Mendonc ̧a, and Karina Villela. Visualization mechanisms for cro wdsourcing information
in emergency coordination. In Proceedings of the 14th Brazilian Symposium on Human Factors in Computing Systems, page 35. ACM, 2015.
● Carsten Keßler and Chad Hendrix. The humanitarian exchange language: coordinating disaster response with semantic web technologies. Semantic Web, 6(1):5–
21,2015.
● Jiongqian Liang, Peter Jacobs, and Srinivasan Parthasarathy. Human-guided flood mapping: From experts to the crowd. In WWW 2018, pages 291–298, 2018.
● Alan M MacEachren, Anuj Jaiswal, Anthony C Robinson, Scott Pezanowski, Alexander Savelyev, Prasenjit Mitra, Xiao Zhang, and Justine Blanford. Senseplace2:
Geotwitter analytics support for situational awareness. In Visual analytics science and technology (VAST), 2011 IEEE conference on, pages 181–190. IEEE, 2011.
● Smriti Murali, V Krishnapriya, and Aadhiya Thomas. Crowdsourcing for disaster relief: A multi-platform model. In Distributed Computing, VLSI, Electrical Circuits and
Robotics (DISCOVER), IEEE, pages 264–268. IEEE, 2016.
● Pascal Neis, Peter Singler, and Alexander Zipf. Collaborative mapping and emergency routing for disaster logistics–case studies from the haiti earthquake and the UN
Portal for Afrika. na, 2010.
● Alexandra Olteanu, Sarah Vieweg, and Carlos Castillo. What to expect when the unexpected happens: Social media communications across crises. In Proceedings of
the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing, pages 994–1009. ACM, 2015.
● Iv́an Palomares, Leo Galway, Martin Haran, Martijn Neef, Conor Woods, and Hui Wang. A knowledge management and need-capacity matching approach
for community-based disaster management and recovery. In ISK, pages 389–396. IEEE, 2015.
● Hemant Purohit, Andrew Hampton, Shreyansh Bhatt, Valerie L Shalin, Amit P Sheth, and John M Flach. Identifying seekers and suppliers in social media communities
to support crisis coordination. CSCW, 23(4-6):513–545, 2014.
● Hemant Purohit, Nikhita Vedula, Krishnaprasad Thirunarayan, and Srinivasan Parthasarathy. Transportation uncertainty in matching help seekers and
suppliers during disasters. In First SIGIR Workshop on Intelligent Transporation Informatics. ACM, 2018.
52
51. Resources
53
● Paper - http://knoesis.org/node/2915
● Video - https://goo.gl/wNp3S2
● DisasterRecord wiki page - http://wiki.knoesis.org/index.php/DisasterRecord
● Github repo - https://github.com/shrutikar/DisasterRecord
● Ontology link - https://goo.gl/k344EH
53. Acknowledgement
55
Team Members of DisasterRecord :
○ Michael Partin, MSc Student of Computer Engineering
○ Dipesh Kadariya, MSc Student of Computer Science
○ Hussein Al-Olimat, PhD Student of Computer Science
Jeremy Brunn for helping with UI design
Alan Smith for collaboration to make the tool real-time.
TO THE ENTIRE KNOESIS FAMILY
Our collaborators Jiongqian (Albert) Liang, Jiayong Liang (Jay), Desheng Liu, and Nikhita Vedula
from Ohio State University for providing flood mapping data.
This research was supported by the NSF award EAR-1520870 “Hazards SEES:Social and Physical
Sensing Enabled Decision Support for Disaster Management and Response”. All views are those of
researchers and do not necessarily reflect the views of the sponsor.
Recent past has seen number of natural disasters which resulted in massive destruction, loss in economy, and had long-lasting effects on the daily lives of affected people.
During each of these natural disasters have smaller sub-events that result from this bigger event and needs timely attention. Like power outage, or blocked roads.
Timely, accurate decisions and response to these events can minimize the cost of rehabilitation.
For such critical decisions, we need to know
WHAT has happened
WHEN it has happened as most of these events are time critical
WHERE it has happened with the spatial context
Timely extraction of this information , integration and organization for a real-time response is challenging.
With this we move on to our problem definition
Twitter data -Major source of information
1000s of people post everyday their daily activities, and numerous tweets are recorded during major natural disasters.
Surveys have proved that among such big data, there might be biased, false or unverified information. But there is still a significant number of relevant and meaningful information that can be gleaned out.
We crawled significant amount of data for major disasters.
The text streams were filtered using hashtags and disaster event related keywords.
Let us look at a tweet that was posted during Chennai 2015 rain.
It reads ….
The tweet comes with meta data like … user profile information- name, location, tweet time, external multimedia
As we said earlier, we want to know what what is happening where is it happening and when.
The WHEN can be answered with the tweet’s timestamp,
The need classification gives us the WHAT and
the location gives us the WHERE
In the disaster text classification, there have been work to classify the kind of disaster say is it hydrological disaster/ a climatological disaster.
The other works concentrate on finer-grained classes - like affected people, injured, missing, donations, volunteers etc
And One another work that does binary classification for infrastructure damage. Just this one has a visualization on a map and provides clustering based on location. Every other disaster text classification do not have any actionable details.
Furthermore, Cameran et al visualized the geo-tagged tweets on map.
About 4% of data are geo-tagged. If we end up relying on just geo-tagged tweets, we would miss out on a lot of critical information.
Our tool LNEx extracts location names from text data and attaches each of these location to their lat-long information on the map. The text data is then classified into one of the need classes.
We care more about the location that is talked about rather than the location it came from.
We visualize Location from a context of a need which can locate the need on a map
we repurposed their fine grained data to form coarse-grained classes like
(such as URLs, non-ASCII characters, mentions, punctuations, dataset-specific stopwords, and hashtags)
focuses on generating new minority class instances near borderlines and this boundary line is defined with SVM so as to help establish boundary between classes.
Produces prediction model in the form of ensemble method. It relies its decision over numerous weak learners to make strong and intelligent model.
The knowledge extracted from these crowdsourced excel sheets provide situational awareness and various kinds information available at each location.
pre-disaster data provides us with available locs inside a bounding box -- which affords the matching
As pre-disaster data we use Open Street Map
OSM is a Gazetteer of location names with meta data.
It includes all the existing locations on the earth, including street names, highways, towns, cities, state and countries.
Meta data would include : geo coordinates, key-value pair explaining the type of the location, the parent it belongs to.
Captured using a boundingbox of the disaster event. These available help locations are confirmed or pruned using satellite images.
Tweets expresses the needs (shelter and rescue) and open street map locations contains the types of locations that are inside a bounding box. Therefore, we needed a way to match a need with a particular location, and that’s why we built the ontology to express the affordability of locations for us to support the matching functionality.
To improve our need classifier, we included in the ontology concepts which represents a given need and used these concepts to improve the classifier later.
For example, “I am stuck at xyz”, xyz is the location of the need, stuck means I need rescue, and our matching functionality should show the nearby locations which can afford the response. For example, a hospital or an emergency response team.
Merely Lexicons
No explicit relationships between concepts
Crisis type
Information type
Used as labels for classification
Approx 11-16 classes as a whole
MOAC
http://observedchange.com/moac/ns/
Answers - Who? What? Where?
Sources : Traditional humanitarian agencies (shelter cluster), Volunteer and technical committees (Ushahidi platform), Disaster affected communities
HXL
http://hxlstandard.org/
Relies on Semantic Web technology and Linked Open Data.
Reuses generic ontologies but not disaster and event ontologies.
Framework for interoperability
UNOCHA's Humanitarian eXchange Language (HXL) which relies on Semantic Web technology and Linked Open Data lacks coverage. It merely reuses generic ontologies (FOAF [#], DCMI [#] and Open Geospatial Consortium [#]), but ignores existing disaster and event ontologies.
pen Geospatial Consortium [#]), but ignores existing disaster and event ontologies.
SOCS
http://knoesis.org/projects/socs
Extended HXL and MOAC
Not available online
No concepts for relief coordination
EDXL-RESCUER
http://ceur-ws.org/Vol-1442/paper_19.pdf
Data exchange with legacy systems
Coarse grained concepts
Answers Where? What? Who? When? Severity, Urgency
Not compatible with social media data
SMEM
http://www.iadisportal.org/ijwi/papers/2016141203.pdf
Combines crisis domain with social media
Not open-source
Coarse-grained concepts
Wang et al., 2014
https://www.sciencedirect.com/science/article/pii/S0198971514001252
Good coverage of concepts
No hierarchical relationships between concepts
So we resorted to building our own ontology but have references to the existing classes with rdfs: references.
We started with labelled class data.
(case folding, lemmatizing, stemming and removing ”noisy”
lexical elements such as URLs, non-ASCII characters, mentions, punctuations, dataset-
specific stopwords, numbers, and hashtag)
In LDA, each document may be viewed as a mixture of various topics and each of the words in the topic belongs to atleast one topic.
M - total number of docs
N-number of words in a doc
Alpha,beta parameters- dirichlet priors
Alpha - per doc topic distribution -- (higher-- likely to contain mixture of most topics, low -- few topics)
Beta- per topic word distribution -- (high -- per topic will contain most of the words , low -- few words)
Theta - topic distribution for document M
Z - each topic
W - word
But there was a critical problem we noticed.
The words as output comes with probability relevance to the topic. And so we wanted to use these as weights to rank the words in a text stream.
But the range of probabilities of words from the two classes were different and thus were incomparable. Therefore, we resorted to normalize these to a different but comman range. And finally used these weights to multiply with the frequency of their occurrences.
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
Hospital - key : value pair would be - building:hospital
Temporal details - There are few events that expire after sometime, - like blocked road - Our ontology does not cover the evolution of events dynamically. This is a strong future work that can be looked at.
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
One of the most important source of data.
Satellite imagery is used for flood mapping of the affected area
A tool developed by our team that takes in a satellite image and interpolate it with the actual map of the region and finds the geo-cordinates that are flooded with any precision value that you require.
The output of the flood mapping tool is a bitmap representing geo-coordinates with boolean values representing flooded or not. These flooded geo-coordinates are used in two ways: to eliminate flooded OpenStreetMap locations and to prune away flooded/closed/inaccessible routes.
D-record caches all of this information in Elasticsearch. We create the map layers on top of MapBox .
Each layer of the map represents a theme of data providing Visual hierarchy is the separation of the map layers into planes of information.
In the aggregation level of analysis, DisasterRecord analyzes, categorizes, and geolocates location names extracted from non-georeferenced tweets. It visualizes aggregated information in a selected geographical bounding box. DisasterRecord eliminates the low-level clutter to surface essential information about disasters to responders. It also abstracts and categorizes data to provide community-level analysis and insights. Upon choosing a location of interest, the key aggregated information includes:
Detecting objects of interest to assess severity and needs, including objects like people (to allow for learning about the disaster impact), animals (for monitoring livestock), and vehicles (to assess traffic maneuverability).
Location-specific available help:
Categorization of OpenStreetMap features to types and their counts.
Location-specific thematic profile:
Highlighting and summarizing the most discussed concepts in a region.
In the individual level analysis, DisasterRecord preserves low-level details to provide individuals with the situational awareness and the kind of available help for their needs. These include:
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
A Problem Well-stated is Half-solved
Charles Kettering, the famed inventor and head of research for GM
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
With the surge in digital information, we have seen numerous options of getting information from the web. Imagine you have a tool that automatically gets all of this data and tells you, during a disaster, what has happened, where it has happened, when it has happened. And assists you with an action plan.
Tool in the video is a full scale working tool- collects, processes and presents the real-time data.
Motivation and goals of this problem were to -
Reduce decision making latency and enhance resiliency
Utilize Multimodal data and multiscale data to confirm real-time events
Provide access to information to the users as well as the responders is as much important as access to food, water and shelter. And other needs during a disaster
Information is most perishable - time critical. - updates dynamically changes with time
Response and Relief Coordination
Let us begin with a short introductory video.
One of the most important source of data.
Satellite imagery is used for flood mapping of the affected area
The satellite images are the output from our flood mapping method which helped prune out routes that were unavailable during the matching process for the location seeking help.
A tool developed by our team that takes in a satellite image and interpolate it with the actual map of the region and finds the geo-cordinates that are flooded with any precision value that you require.
This is the team and we hope to contribute to the humanitarian world and build a resilient city.
Digital information, mobile communication technologies, and access to high volumes of all kinds of data like, digital text, SMS, tweets, satellites etc, have made a remarkable progress in the last decade. Especially, in the crisis domain. This in turn led to the birth of Big Crisis data. Since then several applications have come up that transformed the way we respond to a crisis event. Mapping of such information are done by volunteers called “digital humanitarians” who manually point the locations of these events on a map. One such application is Ushahidi which lets you map information received from various platforms to be mapped on one map, and give the first responders a real-time situation of the affected area for better relief and rescue operations. Other applications are made to analyze the situation by plotting a heatmap depicting the emotions or the magnitude of the event.
The state of the art techniques for disaster response lack in spatial and thematic details, and are limited with respect to usability and dissemination of recommended response for relief coordination. Therefore, we curated a location-centric disaster ontology to assist in matching relevant needs with available help. This ontology was created using information from past disasters and risk reports capturing distinction between response and relief disaster phases.
The class “concepts” has two concepts - Needs and Availability.
The other Needs concept was formed from topic modeling on various labeled data for the two classes Shelter/Food/Supplies and Medical/Rescue help. Each of the terms in these concepts were associated with a relevance probability to the respective need class which are used While Text Classification.
The concept - Availability comes from the location features of OpenStreetMap and are available as help for respective Need classes.