Multiscale Sea-Land Segmentation Networks for Weak Boundaries

Sea-land segmentation based on remote sensing images is one of the most important means to realize dynamic monitoring of coastal zones. However, the traditional models adopt the same design and simple stacking, resulting in equal attention to each region in scenarios mixing strong and weak boundaries, and even less attention to weak boundaries. In addition, due to the complexity of the ecological environment, there is a large intra-class gap and a small inter-class gap in the weak boundary coastal zone. Based on the above problems, this paper proposes a new semantic segmentation network based on remote sensing images: A2RDNet. By replacing standard convolution with cam-conv, this model integrates position information into channel attention, enhancing focus on weak boundary features. Furthermore, the resUNet-6 is proposed to recover substantial location information lost in the initial stages and provide deeper semantic information. Utilizing the DAC module inside the decoder to realize the pyramid structure of multi-scale dense fusion, which enhances the semantic information while fusing the image detail features. A2RDNet was evaluated using a set of Landsat-8 remote sensing images, containing weak boundaries of different types and regions. The experimental results show that the MIoU, mPA, Accuracy, and F1_Score of the proposed method have reached 98.92, 99.46, 99.46, and 99.46. respectively verifying the validity and feasibility of the proposed method. The A2RDNet model has higher segmentation accuracy compared to other methods.