Research on Composite shoreline detection based on U-Net

Sea-land segmentation, as one of the core tasks in intelligent interpretation of remote sensing imagery, directly impacts the effectiveness of marine applications such as coastline change monitoring and intertidal zone resource surveys. To address issues like blurred boundaries and misclassification of small features in existing deep learning methods when processing complex shoreline scenarios, this paper proposes an improved U-Net model incorporating a boundary-aware attention mechanism. The model adopts ResNet50 as its backbone network and introduces a Boundary-Aware Attention Mechanism in the encoder to enhance feature representation in key regions and improve the model’s focus on weak boundaries. During the decoding phase, skip connections are employed to compensate for information loss, while deconvolution layers and Softmax are used for refined segmentation. Experimental results demonstrate that compared to traditional U-Net, SegNet, and DeepLabV3, the proposed model achieves superior performance across metrics including mIoU, mPA, and F1-Score. Particularly in complex mixed shoreline environments, it effectively prevents boundary blurring and misclassification, thereby improving both the accuracy and robustness of land-sea segmentation.