<p>Accurately and quickly identifying the inundation area of urban flooding is of great significance for disaster assessment and disaster prevention and mitigation. However, limited availability of remote sensing satellite data and shortcomings of traditional identification methods often hinder accurate determination of urban flood water body extents. Compared with remote sensing methods, deep learning techniques can realize more efficient and accurate flood area identification. Traditional studies have mainly used semantic segmentation to determine the flood extent, but compared to instance segmentation, it is deficient in fine-grained identification and water body boundary accuracy, which can easily lead to errors. This study pioneers the application of YOLOv8-seg for instance segmentation of video imagery related to urban flood inundation, enabling intelligent identification of inundation extents in urban environments. Initially, a novel urban flood inundation dataset was constructed, encompassing diverse weather conditions, illumination levels, and scene complexities. Subsequently, K-fold cross-validation, sample size comparison experiments, and complex scene detection experiments were conducted on different network variants of YOLOv8-seg, systematically evaluating their applicability and performance differences in the urban flood inundation identification task. In the complex scene detection experiments, YOLOv8-seg was further benchmarked against other mainstream segmentation models, confirming its effectiveness and robustness under challenging environmental conditions. The results confirm YOLOv8-seg’s superior robustness and accuracy, especially in challenging conditions, with model deployment strategies tailored to resource constraints. These findings have significantly advanced urban flood monitoring technology, providing practical and efficient solutions for real-time disaster management.</p>