Floodet: A Lightweight Edge AI Model for Intelligent Flood Detection

In cities, road traffic is often affected by rainy season precipitation, resulting in flood disasters and costly economic losses. How to detect floods economically, quickly, and accurately is an important problem for smart urban management. To meet the above challenges, we propose Fooldet, a lightweight edge AI model for intelligent flood detection that can run on resource-constrained devices such as UAVs and smart cars. Floodet is composed of a Ponding Prediction Module (PPM), a Flood Semantic Segmentation Module (FSSM), and a Resource Adaptive Balancing Module (RABM). Firstly, through the PPM, we can identify the Weather change, and select the pictures that may have ponding. Then, on the premise of ensuring accuracy, FSSM segments and extracts the flooding area. Secondly, to evaluate the performance of Floodet, we collect a rich set of actual flood images from both the real world and the Internet, and propose a semantic segmentation dataset for flood detection. Finally, we implement Floodet using an Nvidia Jetson Xavier NX to emulate edge devices with constrained computing resources. Extensive evaluations demonstrate the validity and practicability of Floodet, achieving 40.13% speed improvement upon existing baselines.