Computer vision-based monitoring of river flooding in urban environments is a cost-effective and low-maintenance method for flood risk management. Research in this area often employs image processing or deep learning techniques to determine the water level, frequently incorporating a water segmentation step. While these methods have been shown to perform reliably during daylight, assessments for nighttime images, in which reduced visibility complicates the use of computer vision, are limited as they do not consider enhancements provided by foundational segmentation models. This study presents a novel methodology that combines the Segment Anything Model (SAM) for water region segmentation and a deep learning-based classification architecture for assessing water depth based on images of an urban creek illuminated only by streetlights at night. The model relies on a hybrid representation in which the segmentation masks are merged with the original images, maintaining contextual information and enhancing feature extraction. Experiments demonstrate that this hybrid enhancement strategy improves classification accuracy by up to \(2.56\%\) . Among the classification architectures evaluated, Xception achieved the highest performance, with an accuracy of \(94.87\%\) and an F1-Score of \(94.60\%\) . These findings highlight the potential of the proposed approach to tackle environmental monitoring challenges using computer vision, even under varying lighting conditions.