Ungauged Basin Flood Prediction Using Long Short-Term Memory and Unstructured Social Media Data

Floods are highly perilous and recurring natural disasters that cause extensive property damage and threaten human life. However, the paucity of hydrological observational data hampers the precision of physical flood models, particularly in ungauged basins. Recent advances in disaster monitoring have explored the potential of social media as a valuable source of information. This study investigates the spatiotemporal consistency of social media data during flooding events and evaluates its viability as a substitute for hydrological data in ungauged catchments. To assess the utility of social media as an input factor for flood prediction models, the study conducted time-series and spatial correlation analyses by employing spatial scan statistics and confusion matrices. Subsequently, a long short-term memory model was used to forecast the outflow volume in the Ui Stream basin in South Korea. A comparative analysis of various input factor combinations revealed that datasets incorporating rainfall, outflow models, and social media data exhibited the highest accuracy, with a Nash-Sutcliffe efficiency of 94%, correlation coefficient of 97%, and a minimal normalized root mean square error of 0.92%. This study demonstrated the potential of social media data as a viable alternative for data-scarce basins, highlighting its effectiveness in enhancing flood prediction accuracy.