<p>This paper presents a novel approach, F-Unet, to tackle the challenge of low segmentation accuracy in sea-land boundary regions when performing sea-land segmentation on high-resolution remote sensing images. The F-Unet network structure improves upon the Unet model by replacing the concatenation part and one of its convolutional blocks with the feature fusion module from the BiSeNet network. This enhancement enables F-Unet to have a wider receptive field, facilitating the effective extraction and fusion of both local detail features and global semantic features. Furthermore, a new loss function, the boundary region enhancement loss, is introduced in this paper. This loss function aims to enhance the network’s ability to learn sea-land boundary regions and improve the accuracy of sea-land boundary region prediction. The experimental results show that on the Gaofen-2 image dataset constructed in this paper, F-Unet with the boundary region enhancement loss outperforms both the original Unet model and Unet using the proposed loss function in terms of overall sea-land segmentation and segmentation accuracy in sea-land boundary regions.</p>