Real-time Disaster Crisis Mapping based on Classification and Geo-location Recognition in Tweets.

Generally, one finds a large percentage of social media data, such as Tweets or Instagram, lack Geo-tagged location in their metadata, minimizing their use in generating Crisis Maps during natural and human caused disasters. In the following work, we will determine the 'at risk' areas for particular geographical locations(New York State for this current research) through post- disaster events such as Hurricane Sandy by the analysis of all tweets originating from the Geo-location area under consideration through exact string matching of location entities in tweet texts. In this study, we employ the 8 Million Twitter data set collected by Aulov, Price and Halem stored in Couch DB. We use a Named Entity Analysis algorithm, based on the Sultanik and Fink, to obtain locations of places mentioned tweets without geo -location tags, thus increasing spatial information relevant to developing real-time Crisis Maps of the affected disaster areas being impacted under hurricane events or other related extreme natural events. The algorithm for Geo-location recognition is based on forming N-Gram tokens extracted from text in the tweet which are further mapped against a location gazetteer to obtain the coordinates of the locations or places through exact string matching in the gazetteer. The location gazetteer contains key-value pairs of the name and alternate names of the places, belonging to New York State as 'key' and their coordinates as 'value'. Once all the locations are found, an augmented Crisis Map consisting of both Geo-tagged and inferred locations is shown to increase the observations of the impacted areas. We show that based on an increase in frequency of locations, `at-risk' areas can be distinguished from `impacted' areas