Floods represent one of the most devastating natural disasters worldwide, causing substantial human and economic losses, particularly in vulnerable urban and agricultural areas. The advent of high-resolution satellite data and deep learning techniques has significantly improved flood detection and monitoring capabilities. This study explores the efficacy of three advanced deep learning models—PAN, PSPNet, and MANet—for semantic segmentation of flooded areas using Sentinel-1 synthetic aperture radar (SAR) imagery. The ETCI 2021 competition dataset, comprising VV and VH polarization SAR data and corresponding flood masks, was employed to train and evaluate the models. Comprehensive experiments revealed that VH polarization outperforms VV across all models, yielding higher accuracy, precision, recall, F1-score, and Intersection over Union (IoU). Among the tested architectures, MANet demonstrated superior performance with an IoU of 84.21% and F1-score of 91.11%, attributed to its multi-scale and mutual attention mechanisms. These findings affirm the value of SAR imagery, particularly VH polarization, combined with deep learning for accurate flood detection and underscore the potential for real-time disaster response applications.