Abstract：Against the backdrop of climate change, the increasing frequency of extreme rainfall poses severe challenges to
urban safety. Traditional monitoring networks based on physical sensors suffer from large coverage gaps and high
maintenance costs, making it difficult to meet the demand for real-time disaster perception in resilient city building.
Adopting a social sensing perspective, this study develops a multimodal transfer learning model that fuses visual
features from images and sentiment features from text to mine social media (Weibo) data for precise localization of
urban waterlogging points and assessment of their severity. Using the July 21, 2012 Beijing rainstorm as a case study,
the results show that the proposed model achieves a recall rate of about 60% for waterlogging points released by the
transportation department, while effectively identifying additional hidden waterlogging locations beyond the official
monitoring network and thus complementing its spatial and informational blind spots. Based on these findings, the
paper proposes strategies such as building an integrated “physical–social” dual sensing network, optimizing the layout
of existing drainage facilities, and establishing people-centered emergency response mechanisms, providing a scientific
basis for enhanced disaster prevention and mitigation and refined resilience-oriented urban governance.
