Cascaded feature fusion of SAR polarization for deep learning-based waterbody detection

Water is an invaluable resource with significant economic and social importance, but it distribution can be life threatening if not properly monitored. With the presence of deep learning, researchers are at liberty to explore Synthetic Aperture Radar (SAR) in several ways for many purposes including water body segmentation. In this study, we introduce a novel Cascaded Feature Fusion Module (CFFM) integrated with a Deep UNet architecture to enhance the detection of water bodies in dual-polarization SAR imagery. Our extensive experiments on GaoFen-3 and Sentinel-1 datasets demonstrate that the proposed CFFM significantly improves the baseline Deep UNet model’s performance by effectively fusing polarization SAR features. This integration leads to superior image quality, reduced noise levels, and increased accuracy in detecting both large and narrow water bodies. Quantitative analysis shows that our model achieves a high mean Intersection over Union (mIoU), surpassing other state-of-the-art models such as BiSeNe, NFANet, and DCFNet. It also exhibits competitive Recall, Precision, and F-Score metrics, indicating its balanced and robust performance. Qualitative analysis further confirms the efficacy of our model, accurately segmenting complex and less homogeneous water regions with minimal noise. These results underscore the model’s potential for industry applications, owing to its lightweight, time-efficient, and versatile nature.