Urban flooding due to intense rainfall events and increased impervious surfaces has emerged as a pervasive challenge affecting the United States. Traditional flood prediction, grounded in physical and hydrological principles, exhibit constraints in real-time applications. Consequently, there has been a discernible shift toward data-driven methodologies, harnessing the capabilities of artificial intelligence and deep learning techniques. This study introduces a Near Real-Time (NRT) flood prediction model using Convolutional Neural Networks (CNN), Sentinel-1 imagery, geo-spatial data, and gauge measurements. U-Net architecture is modified and used for image segmentation to develop the flood prediction model. The model combines digital elevation models, imperviousness, hydrologic soil group and rainfall gauge measurements with calibrated backscatter values derived from Sentinel-1 imagery, thereby facilitating the generation of highly accurate flood predictions. This model is applied to predict flood inundation in Miami-Dade County, Florida, employing near real-time rainfall data. The inundation extent results are cross-referenced with historical flood data obtained from both online repositories and local records. The NRT flood prediction yields the potential to provide rapid and adequately accurate flood inundation at a spatial resolution of 10 m. The findings underscore the model's capacity to effectively capture the vast majority of historical floods, thereby enhancing predictive accuracy. It is found that the choice of polarization affects backscatter behavior and prediction accuracy. The NRT flood prediction model holds substantial promise as a valuable resource for a diverse range of stakeholders, including emergency management agencies, infrastructure management organizations, and urban planners.
