Aiming at the difficulty in fine segmentation of irregular boundaries in coastal remote sensing images, this paper proposes an Asymmetric Multi-path Decoding Network for Coastline Segmentation (AMDNet). Taking Deeplabv3+ as the backbone network, the network uses EfficientNet-B0 as the feature extractor to significantly reduce the computational load of the network. Additionally, the D-LKA module is introduced into the improved ASPP to add extra offsets for adjusting the sampling positions of standard convolution, allowing the convolution kernel to flexibly adjust the sampling grid. Combined with DUpsampling technology to achieve high-precision restoration during the upsampling process, the accuracy of image segmentation is improved. The accuracy, sensitivity, Dice and Jaccard of the AMDNet model on the Aerial photo-maps dataset reach 96.77%, 93.03%, 90.42% and 86.67% respectively, showing a significant performance improvement.