Edge computing and server-based high-precision flood level classification system

Abstract

Urban flooding and the resulting road water accumulation have become a significant threat to public transportation safety and the stability of municipal infrastructure. Traditional monitoring networks based on physical water level sensors suffer from low deployment density, high maintenance costs, and lagging response times. To address these shortcomings of traditional water accumulation monitoring systems, this study proposes an edge-computing intelligent monitoring system based on collaborative inference between the edge end (You Only Look Once version 5, YOLOv5) and the server end (Transform Vision Detection, TrVDet). A dual-modal perception architecture of “edge-end triggering and server-end precise analysis” has been constructed. At the edge end, the YOLOv5 model is deployed on embedded devices to achieve efficient preliminary screening of water accumulation, reducing dependence on the central server, lowering latency, and enhancing real-time response capabilities. On the server end, multi-object segmentation is performed on the detected water accumulation images, including roads, cars, motorcycles, and bicycles. Finally, a series of logical judgments is applied to determine the water accumulation level based on reference objects within the water. Since there is no publicly available dataset for target object recognition in flooded areas, we employed professional annotators to perform pixel-level labeling on the collected and organized flood data and constructed a multi-class target flood dataset (City Flood Segmentation, CityFloodSeg). Given the scarcity of moderate and severe water accumulation samples, we optimized the instance segmentation model TrVDet under the (A Visual Representation for Neon Genesis, EVA-02) framework and applied five data augmentation methods, including Mosaic and Flip, to expand the diversity of the dataset. Moreover, based on domain expert standards, we designed a logical judgment rule algorithm for model inference of water accumulation levels to classify the levels of water accumulation. Experimental results show that the server-end processing delay is stable within 0.4 s, capable of accurately judging different water accumulation risk levels. This provides centimeter-level real-time situational awareness for urban flood control decision-making and promotes the development of intelligent municipal infrastructure towards higher reliability and universality.