AI-Enhanced Multimodal Volunteered Geographic Information for Assessing and Maintaining Urban Stormwater System Under Extreme Weather

Extreme weather events, intensified by climate change, are increasingly threatening urban environments worldwide, necessitating the development of resilient cities as prioritized by the United Nations’ Sustainable Development Goals. Hong Kong, to align its efforts with these goals, faces an urgent need for next-generation interventions for extreme weather events following the “Black Rainstorm of the Century” in 2023. The city's Urban Stormwater System (USS) is currently under strain, facing insufficient capacity to handle escalated rainfall levels. Additionally, operational inefficiencies, such as aging infrastructure, further impede drainage performance. Consequently, how to effectively assess the working conditions of USS and develop maintenance strategies to withstand extreme weather has become a critical issue.⁠⁠⁠⁠⁠⁠⁠Current assessment of USS typically relies on routine inspections, such as scheduled inlet and pipeline inspections. However, these approaches suffer from limited spatial coverage and retrospective observations, and therefore may fail to identify the actual USS failure points during extreme precipitation events. To address this gap, the project proposes leveraging Volunteered Geographic Information (VGI) from social media platforms for the assessment of USS. In this vision, VGI offers a complementary approach: citizens act as distributed sensors, generating geo-tagged flood posts that directly correspond to locations where the USS is not functioning properly, particularly in areas of high public concern. When properly extracted and analyzed, such flood-related VGI could provide timely, failure-tagged data for effective USS assessment.The research flow is designed as follows: (i) An AI-empowered toolbox will be developed to identify, geolocate, and quantify real-time flood-related data from multimodal VGI,including text, images, and videos. This toolbox will quantify flood extents, identify flood locations, and assess public sentiment to inform maintenance strategies. (ii) The vulnerability of local drainage units will be evaluated using the extracted flood extents,which will also serve as labels to train a Bayesian Graph Neural Network (BGNN). The BGNN will probabilistically reveal how vulnerability propagates across the entire USS network, especially in areas lacking VGI coverage. (iii) USS failure risk across the network will be quantified by integrating failure probabilities (from the BGNN) with estimated failure consequences, encompassing social, economic, and public sentiment impacts. Targeting risk mitigation, an explainable LLM-driven agent for USS management will be developed. Following comprehensive validation and evaluation in real-world city settings, the developed AI agent is expected to enhance both short-term emergency response and long-term maintenance planning, and, more broadly, to support the development of climate-resilient cities.