Automatic Detection of Urban Flood Risk Level with Yolov8

Abstract

Urban flooding poses significant threats to lives, properties and transportation facilities, and accurate and timely flood information is critical to city decision-making.With the widespread use of 5G and surveillance cameras in cities, video image is becoming a new data source and showing great potential in urban flood monitoring. This study introduces a methodology for assessing the risk of flooding by recognizing the submerged state of cars in video images. Initially, the vehicle flooding status is categorized into five distinct classes, each corresponding to a specific urban flood risk level. A dataset comprising 2,000 images featuring 6,000 vehicles was compiled and annotated for this purpose. Subsequently, four variations based on the YOLOv8 framework were evaluated in recognizing the car submerged state. This investigation also entails determining the optimal model configurations for YOLOv8 across varying environmental conditions, facilitated through rigorous training evaluations and the verification of detection accuracy in complex scenarios. Furthermore, this research examines the influence of sample size and annotation precision on model training outcomes and overall performance. Our study demonstrates the effectiveness and applicability of our method in urban flood monitoring, and provides new ideas for video image-based flood detection research.

Keywords: Urban flood, Disaster risk, Automatic detection, YOLOv8, Deep learning