A Fusion Method Incorporating Dual-Attention Mechanism and Transfer Learning Into UNet++ for Remote Sensing Image Coastline Extraction

The segmentation of land and sea in remote sensing imagery is of great significance for coastline extraction and dynamic monitoring. Traditional coastline recognition and extraction methods based on spectral features and image processing can only generate limited image feature results when facing the complex textures and spatial distributions of high-spatial resolution remote sensing images, leading to low accuracy in segmentation outcomes. This paper applies a deep convolutional neural network to the problem of sea-land segmentation in high-spatial resolution remote sensing images and innovates upon the classic encoder-decoder architecture. Firstly, to enhance the network’s ability to distinguish coastlines, a dual attention mechanism is introduced into the UNet++ network to improve the learning capacity of coastline features while suppressing the learning of non-coastline features. Secondly, an improved joint loss function is adopted to enhance training effectiveness, thereby significantly improving the accuracy of semantic segmentation. Lastly, transfer learning is utilized to strengthen the detailed features of coastlines and enhance the network’s ability to identify them. Experimental results using the GID dataset for coastline segmentation demonstrate that compared to the latest algorithms such as PSPNet, CS-Deeplab v3+, and UNet++, the improved UNet++ network achieves lower boundary blurriness and more accurate segmentation results for coastlines, with fewer missed and false detections. Amidst the proliferation of high-spatial resolution remote sensing image data, the utilization of the enhanced UNet++ model for coastline extraction has demonstrated remarkable abilities in preserving boundary information and achieving superior semantic segmentation performance. This advancement enables a more refined extraction of spatial distributions, textures, and spectral features from these images, ultimately contributing to an improvement in classification accuracy.