DA-MiTUNet: A Mix Transformer with dual attention embedding in unet for Land-Sea segmentation of remote sensing images

• A novel network architecture (DA-MiTUNet) is proposed to improve the accuracy and robustness of land-sea segmentation. • An encoder with a dual-attention mechanism is designed to achieve feature extraction and enhancement. • An Adaptive Feature Fusion Module (AFFM) is developed to better restore land-sea boundary details. • DA-MiTUNet maintains stable segmentation performance in complex scenes, providing support for dynamic coastline monitoring. Automatic extraction of coastlines from remote sensing images is of great practical importance for coastal risk assessment, ecological environmental protection, and marine economic development. However, the highly dynamic nature of coastlines and the complex, diverse characteristics of land–sea boundaries make precise coastline extraction a challenging task. Although traditional deep learning methods have demonstrated good performance in this respect, they still face numerous shortcomings when dealing with high computational costs and the need to fully utilize multiscale features. In this paper, to address these problems, we propose a novel and efficient land–sea segmentation model for remote sensing imagery based on a classical U-shaped network structure, named DA-MiTUNet. On the one hand, we introduce the convolutional block attention module into the Mix Transformer (MiT), forming a dual-attention encoder in conjunction with an efficient self-attention mechanism. This integration ensures comprehensive extraction of global context and local information, thereby enabling more precise determination of complex land–sea boundary features. On the other hand, we propose an adaptive feature fusion module to further promote the effective fusion of features across different hierarchical levels, achieving more refined land–sea boundary segmentation. Experimental results on the Gaofen-1 Hainan Coastline Dataset (GF–HNCD) and Benchmark Sea–Land Dataset (BSD) datasets demonstrated that the proposed DA-MiTUNet model outperforms other comparative models in terms of both the average F1 score and the mean Intersection over Union value, while achieving excellent segmentation results with relatively low computational complexity, thereby reflecting the potential of our model for dynamic coastal monitoring during extreme sea level events.