Abstract: Addressing the impact of coastal features on the refined segmentation of land-sea boundaries in high-resolution remote sensing images, this paper proposes a land-sea segmentation algorithm for remote sensing images based on DGIONet. In the encoding stage, the network is designed with a multi-scale spatial attention feature extraction module based on rectangular strip convolution, which utilizes rectangular strip convolutions constructed vertically at different scales to achieve a multi-scale large kernel convolution effect. Relying on the extracted multi-scale features and point convolutions within the module, it implements a spatial attention mechanism, effectively enhancing the network's ability to focus on large scale land-sea features, and enabling feature extraction of land-sea global information and contextual information. In the decoding stage, the network incorporates a dual-path global information optimization decoder. Within the decoder, it relies on a depthwise separable dilated convolution information optimization module and a “Hamburger” global feature restoration module to leverage the extracted global and contextual information for feature restoration. Meanwhile, it fuses the stage features extracted from the encoding stage with the stage-restored features to achieve better information optimization. To verify the effectiveness of the proposed method, this paper constructs a high-resolution remote sensing image land-sea segmentation dataset with a resolution better than 0.3m. Comparative experiments conducted on this dataset demonstrate that the proposed method improves pixel accuracy by 3.01% and mean intersection over union(mIoU) by 10.51% compared to the current mainstream semantic segmentation method, Vision Transformer. This proves the advantages of the proposed method in the refined land-sea segmentation of high-resolution remote sensing images.

Key words: high-resolution remote sensing images, sea-land segmentation, deep learning, convolutional neural network, attention mechanism