The U-Net architecture has proven to be effective in shoreline segmentation tasks, especially with remote sensing and Synthetic Aperture Radar (SAR) data. However, none of these studies have used high-resolution Unmanned Aerial Vehicle (UAV) imagery for semantic segmentation. In this study, we compare the performance of the vanilla U-Net and two of its variants, namely the Attention U-Net and the ResU-Net, for semantic segmentation on a high-resolution UAV imagery dataset of coastal images. The dataset, consisting of 800 images showcasing various coasts of Mauritius, was split into 60% training, 30% validation, and 10% testing sets. U-Net, serving as the baseline model, achieved a mean Intersection over Union (mIoU) of 0.771. Attention U-Net showed a slight improvement with a mIoU of 0.773, while ResU-Net achieved a mIoU of 0.716, performing best on images with vegetative areas. Results indicate that U-Net and Attention U-Net are effective for semantic shoreline segmentation based on aerial imagery, particularly for distinguishing between wet and dry sand. This study demonstrates the potential of these models for accurate shoreline mapping in Mauritius, even with a relatively small dataset.