Integrating social media and deep learning for real-time urban waterlogging monitoring

Swat district, Khyber Pakhtunkhwa (KPK) Province, Pakistan. With the rise of social-media data, there is an increasing need to promptly and precisely identify content related to disasters, such as urban waterlogging. Social-media data, being cost-effective and abundant, can offer valuable insights into geographic phenomena by analyzing human behavioral patterns, making it a powerful resource for detailed waterlogging (WLG) analysis in urban settings. This research introduces a novel framework for precise information retrieval and real-time extraction of WLG points and fine-grained information in disaster-affected areas using the Facebook platform. First, topic modeling and transfer learning techniques were developed to examine the spatiotemporal dynamics of WLG locations. Second, water depth data from textual content and visual representations were extracted using various deep learning frameworks and integrated through decision-making processes. Third, a unique fine-grained location corpus tailored to urban flooding scenarios was created using the named entity recognition (NER) model. Finally, the BERT-BiLSTM-CRF model was employed to extract WLG points accurately. Using the Swat district as a case study, the extracted WLG points covered at least 79 % of the officially documented WLG points and were primarily located near roadways, especially in low-elevation areas. This framework provides a viable approach for enhancing situational awareness and conducting spatiotemporal analysis of urban floods and WLG disasters at the municipal level in real-time. • Advanced artificial intelligence models extract precise water depth and enhance up-to-the-minute flood insights. • Real-time flood data from social media outperforms official sources in precision. • Novel approach integrates social media for real-time urban waterlogging analysis. • Deep learning models achieve over 79 % accuracy in waterlogging detection.