<p>Nighttime urban waterlogging detection remains challenging because weak illumination, glare, and water-surface reflections often obscure the boundaries and features of waterlogged road regions. In addition, the limited availability of annotated nighttime samples restricts the generalization ability of data-driven models. To address these issues, this study develops a vision-based edge intelligence framework for nighttime urban waterlogging detection and segmentation. A diffusion-based augmentation strategy is first used to expand nighttime urban waterlogging samples. Then, a lightweight instance segmentation model, named Nocturnal-YOLOv11, is constructed by embedding an adaptive low-light enhancement module into the YOLOv11 framework. The model is further converted, quantized, and deployed on the RDK X3 edge computing platform using an INT8 inference scheme. Experiments on an independent real nighttime test set demonstrate that the proposed method improves low-light visual perception of waterlogged areas under weak illumination and reflective conditions while maintaining real-time edge inference capability. These results indicate its potential for practical nighttime waterlogging perception, urban water management, and emergency response applications.</p>