Assessment of urban flood disaster responses and causal analysis at different temporal scales based on social media data and machine learning algorithms

Abstract

With the widespread popularity of social media, leveraging this real-time data for flood disaster information acquisition and emergency management has become increasingly important. This study conducts a statistical analysis of urban flooding in Changsha from 2017 to 2024, utilizing social media data and machine learning algorithms. The findings indicate a temporal correlation between social media activity and rainfall, with a significant increasing trend in response rates over time. Short-term discussions on social media focus on the immediate impacts of flooding, while long-term conversations emphasize disaster prevention measures, infrastructure improvements, and governmental responsibilities. In a short-term analysis, the case of the "6.30 Rainstorm" identified 134 flood point locations through text recognition and 52 depth information points via image recognition, revealing 51 matches with official data, as well as 30 additional at-risk sites, including those in schools and residential communities. In a long-term analysis, the study identified continuous location data for 850 flood points over eight years and conducted statistical and kernel density trend analyses. Results show that flooding predominantly occurs during the flood season, with an increased likelihood of flooding in newly developed urban areas. Contribution analysis using an optimized gradient boosting decision tree identified population density, elevation, and seasonal rainfall as the primary factors contributing to flooding, while vegetation cover was found to mitigate flooding risk. The study highlights the significant potential of social media for assessing urban flood impacts and recommends integrating diverse data sources and further exploring image information to enhance analytical comprehensiveness and accuracy.