Integrating Social Media into a Pan-European Flood Awareness System: A Multilingual Approach Valerio Lorini European Commission, Joint Research ISCRAM19 Centre (JRC), Ispra, Italy Universitat Pompeu Fabra, Barcelona, Spain valerio.lorini@ec.europa.eu V.Lorini valerio.lorini@ec.europa.eu, C.Castillo chato@acm.org F.Dottori francesco.dottori@ec.europa.eu Milan Kalas milan.kalas@kajoservices.com D.Nappo domenico.nappo@gmail.com P.Salamon peter.salamon@ec.europa.eu
https://ec-jrc.github.io/lisflood/ Open Source Hydrological model @valeriolorini https://bitbucket.org/lorinivalerio valerio.lorini@ec.europa.eu lorinivalerio@gmail.com
This presentation Collector Copernicus case study JRC Aggregator Annotator future dev EFAS GLOFAS Geotagger Context SMFR Deployment valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Copernicus JRC EFAS GLOFAS Context valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Weather driven disasters are on the rise… Weather driven disasters are on the rise…
Total losses for natural disaster 45 % 225 Billion USD hydro events
Paris climate agreement: 185 countries have committed to limit the increase of average temperature to 1.5°C 5.700 # of death Dottori et al, Nature Climate Change, 2019 >10.000 1.5°C >12.000 2°C >20.000 3°C
Copernicus Emergency Management Services Complementary to national efforts Providing European wide information to the EU’s Emergency Response and Coordination Centre (ERCC) Knowledge exchange on emergency management for disaster risk at European level Focus on Europe but available globally
Copernicus Emergency Management Services
EFAS – European Flood Awareness System Provide complementary, added value flood early warning & monitoring products to improve the preparedness and emergency response of relevant stakeholders different forecasting & monitoring products (probabilistic, multi-ensemble, medium-range flood forecasts, flash flood indicators, radar nowcasting, etc.) impact forecasting (possible consequences of predicted events, e.g. flood extent, population affected)
GloFAS – Global Flood Awareness System Provide complementary, added value flood early warning & monitoring products to improve the preparedness and emergency response of relevant stakeholders different forecasting & monitoring products (probabilistic, multi-ensemble, medium-range flood forecasts, flash flood indicators, radar nowcasting, etc.) impact forecasting (possible consequences of predicted events, e.g. flood extent, population affected)
Preparedness Response Recovery Isn’t it perfect??? What can go wrong??? valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System Photo by Inge Wallumrød from Pexels
2016: let’s Iterative check social keywords media refinement Encouraging results Let’s use SM “The Seine river is rising. 2pm in Paris, Pont Neuf . More flooding coming!” [{ 'country_conf': 0.96474487 , 'country_predicted': 'FRA', 'geo': {'admin1': 'Île-de-France', 'country_code3': 'FRA', 'feature_class': 'A', 'feature_code': 'ADM2', 'geonameid': '2968815', 'lat': '48.8534', 'lon': '2.3486', 'place_name': 'Paris'}, 'spans': [{'end': 5, 'start': 0}], 'word': 'Paris'}] On a side note…among us…by the way… If only ALL the tweets were like these… valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Photo by Inge Wallumrød from Pexels Preparedness Response Recovery Provide forecast verification Improve situation awareness Connecting flood early warning systems with social media information valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Photo by Bhavesh Jain from Pexels social media analysis (passive, general-purpose user contributions) Social and mainstream media monitoring can provide early information and data on hazardous events at large scale crowdsourcing (active, targeted contributions requested by emergency responders) There were not yet approaches able to provide seamless and reliable integration of this information with existing forecasting, monitoring and mapping tools It is difficult to process It is difficult to provide data in a time frame multilingual coverage appropriate for coherent with CEMS emergency domains management.
Collector Annotator Aggregator Geotagger SMFR valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Main technical challenges Twitter's API restrictions : limit data collection Multiple Lack of languages : explicit multiplies geographical data coordinates : annotation requires geo- requirements coding Language ambiguity : requires automatic classification
SMFR architecture Annotator Triggering NoSQL All tweets On-demand Geocoder Queue Collector Events metadata Aggregator SQL SQL Selected tweets valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
System infrastructure Architecture based on a “facade” REST SERVER and micro services which expose start/stop operations. Asynchronous persistence to Cassandra leveraging on Kafka queues. Development phases and deployment are based on containers. We use an internal Docker SWARM of 4 nodes.
Collector Annotator Aggregator Geotagger Collector NUTS-lev2 Rapid Risk Assessment NUTS-name Cities bounding_box Flood probability <48 hrs in advance valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Collector Annotator Aggregator Geotagger Collector NUTS-lev.2 EU = ADM-lev.2 GADM valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Text classification: first attempts Collector Annotator Aggregator Geotagger Annotator _loc_? Diverse training sets flooding @user? ? Crowdsourced annotations @user? 29% 56% 36% Multiple annotators/tweet #? RT @user? Typically 80%-85% accurate _url_? flood? 22% 12% Other methods (e.g., SVM) 17% 14% 9% 8% Example Decision Tree valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
CNN for text classification Collector Annotator Aggregator Geotagger D = m x d = S = 50 C = 5 300 5 x 128 Annotator Max-Pooling S x D embedding Convolutions Size m x d Initialized w/ word2vec Width C • Maximum • Word • Width of • Size of max- sequence embeddings convolutions pooling length in words dimensionality flood warning yes due to heavy no rain valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
CNN for text classification Collector Annotator "was having a rough day "photos of students "reeds beach restoration till i saw tops pics helping families clean up aims to improve water Aggregator flooding my social Geotagger their flooded..." flow, reduce flooding" media" Annotator • 99% YES • 39% YES • 2% YES flood Embedding warning Convolution Max pooling due 85% 1K 30’ Convolution to • Accuracy • Training • Training samples time Max pooling heavy Hidden (2) dense in/out rain YES NO ... valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Word embeddings Collector Annotator Aggregator Geotagger Annotator MUSE – Facebook – language agnostic valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Geocoding implicit geo reference Collector Annotator Aggregator Geotagger Geo coding We try to use mordecai for geolocating the most comprehensive text Text: Ministrul Apelor și Pădurilor în zonele cu SpaCy POS risc la inundații din județul Sibiu In second instance we take “place” and “coordinates”objects from the tweet {'country_conf': If the geolocator 0.837, cannot find lat,lon, we 'country_predicted': 'ROU', 'geo': {... ElasticSearc + geonames do not assign the 'lat': '45.8', 'lon': NER tweets to the collection '24.15' , 'place_name': 'Sibiu'} ... valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Aggregating tweets per collection Collector Annotator Aggregator Geotagger Aggregator valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
This Talk case study future dev Deployment valerio.Lorini@ec.europa.eu Integrating Social Media into a Pan-European Flood Awareness System
Recommend
More recommend
