Identification of nighttime urban flood inundation extent using deep learning

Abstract. With the acceleration of urbanization, the disaster of urban flooding has had a serious impact on urban socio-economic activities and has become one of the important factors restricting social development in China. Accurate and timely identification of urban flooding extents is crucial for decision-making. Traditional remote sensing technologies are often limited by environmental factors, making them less suitable for application in complex urban terrains. With the increase in urbanization and the development of emerging technologies, video imagery has become a significant data source with great potential for urban flood identification. However, existing research has primarily focused on flood extent identification in daytime scenarios, often neglecting the nighttime, a period of high flood occurrence. In this study, we propose an efficient model (NWseg) to identify flood extents in nighttime scenes. Initially, we constructed a nighttime flood inundation dataset consisting of 4000 images. Subsequently, MobilenetV2 and ResNet101 networks were used to replace the DeepLabv3+ backbone network and compared with the NWseg model. Next, the NWseg model was compared with ResNet50-FCN, LRASPP, and U-Net models to evaluate the performance of different models in nighttime urban flooding extent identification. Finally, we verified the applicability and performance differences of each model in specific environments. Overall, this study successfully demonstrates the effectiveness of the NWseg model for nighttime urban flooding extent identification, providing new insights for nighttime flood monitoring in cities.