Climate change has led to an increase in the intensity and frequency of localised heavy rainfall, resulting in a higher risk of flooding and inundation. Intense rainfall occurring within a short duration, which cannot be adequately drained, accumulates and flows on the roads, causing abrupt flooding in urban areas. Road flooding in urban areas leads to traffic immobilization, rapidly impairing urban functionality.
Efforts are being made to install sensors such as flood detectors to forecast road flooding. However, expanding these installations challenges to establishing a reliable road flooding monitoring system due to installation costs, maintenance, and management problems, as well as measurement errors caused by transported debris such as fallen leaves, has been complicated. Recently, with the advancement of deep learning algorithms, image processing techniques based on deep learning algorithms have gained significant attention as highly useful tools, instead of sensors, for analysing catastrophic readiness.
In this study, we utilised the You Only Look Once (YOLO) platform, a widely used computer vision platform based on deep learning algorithms, to detect vehicles and determine the depth of submersion in road flooding images. Through web crawling from social media and Seoul City's traffic CCTV footage, we built a dataset consisting of 3,625 sample images. The dataset was annotated with labels indicating the depth of car submersion, which were categorised into five levels: Level 0 representing no water on the road surface and Levels 1 to 4 representing different degrees of wheel submersion. A total of 4,368 labels were assigned to the vehicles in the sample images. We developed a car submersion depth classification model using You Only Look Once version 8 (YOLOv8). The model was trained on a dataset of 2,106 samples, comprising a total of 3,610 labels. We then evaluated the model's performance using 903 validation samples, which had a total of 1,519 labels. The performance of the model was evaluated using the mean Average Precision (mAP), which is a widely used metric for assessing the accuracy of object localization and class probability classification. The validation results showed a mAP of 0.96 at an Intersection over Union (IOU) threshold of 0.5.
The deep learning-based road flooding depth analysis model developed in this study can automatically detect vehicles from CCTV videos and images, classify the flooding level of each detected vehicle into flood levels, and analyse the flooding status of roads in real setime (taking approximately 1/1000 second per frame). Without the need for additional measurement equipment, it can analyse the flooding status of Seoul's roads from thousands of CCTV videos capturing the city's roads, enabling real-time identification of potential road flooding risks.
This model can be utilised as a real-time automatic warning system to secure the golden time for citizens to evacuate in the event of heavy rainfall. It also has various applications, such as in underground tunnel flooding early warning systems and real-time flooding road information provision. Furthermore, it can be applied as a key technology for Seoul's AI-based integrated flood disaster response and management system, covering smart warning systems, disaster map updates, and flood forecast alerts. To implement this model effectively for Seoul's AI-based integrated flood disaster response system, it is crucial to have a strong determination and active collaboration from the Seoul Metropolitan Government.