Abstract: To enhance the monitoring and early warning capabilities for waterlogging in megacities, this study addressed the low practicality and insufficient real-time performance of existing waterlogging detection methods through the use of a high-density network of over 90, 000 public surveillance cameras in Guangzhou. A road waterlogging image recognition method was established based on the RTMDet model, a deep-learning-based object detection algorithm. Multithreading technology was employed to efficiently acquire images from a large number of cameras. The RTMDet model with instance segmentation capabilities enables rapid detection and identification of waterlogging areas. A set of 8463 images, collected between August 15, 2018, and May 23, 2020, was used to develop the recognition model. The model was then validated and evaluated using 6106 images collected between June 25, 2020, and September 10, 2022. The results indicate that during intense precipitation, the model can accurately identify images of obvious waterlogging and the locations of these waterlogged areas. After data cleaning, the overall recognition accuracy of the model was 86.60%. Light interference and image blurring due to precipitation were the two primary factors causing false alarms. Currently, this algorithm has been implemented in Guangzhou's meteorological impact forecasting verification and meteorological decision-making support platform, providing effective support for automatic monitoring and early warning of urban waterlogging. It also offers valuable insights for water management, transportation, and other relevant sectors.