The utility of social media for both collecting and disseminating information during natural disasters is increasingly recognised. The rapid nature of urban flooding from intense rainfall means accurate surveying of peak depths and flood extents is rarely achievable, hindering the validation of urban flood models. This paper presents a real-time modelling framework to identify areas likely to have flooded using data obtained only through social media. Graphics processing unit (GPU) accelerated hydrodynamic modelling is used to simulate flooding in a 48-km2 area of Newcastle upon Tyne, with results automatically compared against flooding identified through social media, allowing inundation to be inferred elsewhere in the city with increased detail and accuracy. Data from Twitter during two 2012 flood events are used to test the framework, with the inundation results indicative of good agreement against crowd-sourced and anecdotal data, even though the sample of successfully geocoded Tweets was relatively small.