Improving Automatic Weather Observations with the Public Twitter Stream

Online social networks such as Twitter and Facebook have become a fixture in the lives of millions of people worldwide. Not only are people communicating with those in their social network, but applications like Twitter allow people to publicly broadcast information relevant to them. For the most part initial weather observations are done automatically, however, some aspects of the weather are still better observed by human eyes. In this paper, we argue that citizens report on the weather they are experiencing through social media tools such as Twitter. Citizen reporting through the Twitter stream will be less accurate than trained observers; however, we posit that the information can be accurate enough to overall improve reports of localized weather activity when contextually related through complex event processing. We develop a method to accurately mine weather events from the public twitter stream that detects primitive weather events from individual users tweets. The method will further detect clusters of all users primitive weather event tweets spatiotemporally and thus infer a real-world weather event. These real-world weather events mined from the Twitter stream are then used to improve automated weather observations within the same spatiotemporal region. We implement the proposed method using Streambase [1] and then evaluate the usefulness of the method. Unfortunately, our results indicate that the Twitter stream does not contain sufficient contextual information to be an ideal source for such spatiotemporal relationships and can not practically benefit reports of localized weather activity.