Extraction of actionable information from crowdsourced disaster data

Natural disasters cause enormous damage to countries all over the world. To deal with these common problems, different activities are required for disaster management at each phase of the crisis. There are three groups of activities as follows: (1) make sense of the situation and determine how best to deal with it, (2) deploy the necessary resources, and (3) harmonize as many parties as possible, using the most effective communication channels. Current technological improvements and developments now enable people to act as real-time information sources. As a result, inundation with crowdsourced data poses a real challenge for a disaster manager. The problem is how to extract the valuable information from a gigantic data pool in the shortest possible time so that the information is still useful and actionable. This research proposed an actionable-data-extraction process to deal with the challenge. Twitter was selected as a test case because messages posted on Twitter are publicly available. Hashtag, an easy and very efficient technique, was also used to differentiate information. A quantitative approach to extract useful information from the tweets was supported and verified by interviews with disaster managers from many leading organizations in Thailand to understand their missions. The information classifications extracted from the collected tweets were first performed manually, and then the tweets were used to train a machine learning algorithm to classify future tweets. One particularly useful, significant, and primary section was the request for help category. The support vector machine algorithm was used to validate the results from the extraction process of 13,696 sample tweets, with over 74 percent accuracy. The results confirmed that the machine learning technique could significantly and practically assist with disaster management by dealing with crowdsourced data.