Harnessing social sensing for real-time flood event reconstruction: A digital autopsy of the 2024 Valencia DANA

Extreme rainfall and flash-flood events in the Western Mediterranean pose persistent challenges for real-time monitoring and emergency response. Traditional hydrometeorological networks often provide limited spatiotemporal coverage during rapidly evolving events, leading to localized blind spots. This study reconstructs the impacts of the October 2024 catastrophic floods in Valencia using more than 156,000 geolocated messages from X (formerly Twitter) as a form of social sensing. These citizen-generated data were integrated with hydrological and hydraulic modeling to contextualize reported flood dynamics and response gaps. Advanced Natural Language Processing (NLP) techniques were applied to extract spatiotemporal patterns of public perception, emergency reactions, and infrastructure disruptions. In addition, a Retrieval-Augmented Generation (RAG) system was implemented to transform unstructured messages into structured event summaries, providing actionable insights during the disaster window. The analysis revealed a clear temporal alignment between social media activity and flood progression, with online reporting peaking shortly after critical impact thresholds. Cross-validation with hydrological simulations confirmed the potential of social sensing to detect localized effects often missed by conventional monitoring, while also exposing the limitations of relying solely on proxy digital traces. Overall, the Valencia case demonstrates the feasibility of integrating human-sensed information with physically based models to enhance situational awareness and support adaptive, data-driven disaster risk reduction in urban environments.