Tracing the German centennial flood in the stream of tweets: First lessons learned
Social microblogging services such as Twitter result in massive streams of georeferenced messages and geolocated status updates. This real-time source of information is invaluable for many application areas, in particular for disaster detection and response scenarios. Consequently, a considerable number of works has dealt with issues of their acquisition, analysis and visualization. Most of these works not only assume an appropriate percentage of georeferenced messages that allows for detecting relevant events for a specific region and time frame, but also that these geolocations are reasonably correct in representing places and times of the underlying spatio-temporal situation. In this paper, we review these two key assumption based on the results of applying a visual analytics approach to a dataset of georeferenced Tweets from Germany over eight months witnessing several large-scale flooding situations throughout the country. Our results confirm the potential of Twitt er as a distributed 'social sensor' but at the same time highlight some caveats in interpreting immediate results. To overcome these limits we explore incorporating evidence from other data sources including further social media and mobile phone network metrics to detect, confirm and refine events with respect to location and time. We summarize the lessons learned from our initial analysis by proposing recommendations and outline possible future work directions.