A Real-Time Emergency Event Detection and Location Prediction Framework for Twitter Streams

Abstract

This paper presents a methodology that combines the spatio-temporal metadata of Twitter posts with the analysis of their contents to filter the stream data for emergency situation reporting and visualization. Twitter messages are monitored in real time for a number of small scale emergency events by streaming and analyzing the contents of the posts using Kafka-enabled technology and convolutional neural network deep learning classification models, respectively, to determine whether or not they are incident-related. All the incident-related posts are further classified into the emergency categories they belong to and then geocoded to determine their locations by considering tweet geotags (GPS location), tweet place, textual content for mentioned location and users’ profile location as the main location-related elements. This results in a significantly higher rate of tweets with associated location information, and hence enables tweet location analysis and visualization for smaller events. As an application, we developed an incident detection, notification and reporting system that detects emergency events promptly and sends notification e-mails to the appropriate agencies for timely response.