基于遥感图像语义分割的海岸线提取与类型判别方法OA

In view of the complex terrain,significant target variations,and diverse backgrounds in coastal line remote sensing images,this study proposes a U-shaped image semantic segmentation network based on the Swin Transformer(SS-UNet),with the aim of enhancing the accuracy of coastal line extraction and classification.The model achieves performance optimization based on the following advancements:incorporating the Swin Transformer module into the U-Net encoder-decoder architecture to replace traditional pooling and upsampling operations,thereby improving the capability to extract global features;introducing a channel attention mechanism(SENet)within the skip connections to optimize multi-scale feature fusion;designing a loss function that includes a penalty term derived from a distance map to enhance the learning of boundary regions.Experiments demonstrate that in comparison with the benchmark models such as U-Net,UNet++,and DeeplabV3+,the SS-UNet improves mean accuracy by over 1.6%,and can perform automatic differentiation of coastal line types,so it provides an effective solution for large-scale coastal line monitoring.