Urban flooding is a major threat to safety and city infrastructure. Early warning and rapid detection are vital to minimise damage and protect lives. However, distinguishing dangerous floods from harmless puddles remains challenging. This paper presents a real-time detection system for urban floods and puddles, using advanced computer vision and edge networking. The system is built on the YOLOv8 object detection model, which identifies and classifies floods and puddles in street-level scenes. A Milesight 5G-enabled camera streams high-resolution, low-latency video to the model, which enables prompt detection across urban areas. The model was developed using a curated dataset comprising 9,467 labeled images, which were sourced from the Roboflow platform. Careful selection and inclusion of non-flood backgrounds enhanced the model’s ability to distinguish between flood events and ordinary water pooling. Training was structured with 70% of data for training, 20% for testing, and 10% for validation. The YOLOv8 model achieved a precision of approximately 90.6% in identifying flood scenes, with few false positives. Recall for water classes was 72.4%, indicating that smaller puddles may occasionally be missed. The system was tested with the Savonia University of Applied Sciences’ private 5G network and Milesight camera infrastructure. This confirmed seamless real-time operation and demonstrated its suitability for integration with existing urban surveillance systems. These results show that combining YOLOv8 with 5G connectivity offers a robust, scalable approach to urban flood monitoring, supporting emergency response and infrastructure protection.
Baig et al. (Wed,) studied this question.