Multi-scale Graph Convolutional Neural Network for Marine Aquaculture Extraction Based on SAR Images

Synthetic aperture radar (SAR) image segmentation is one of the hotspots in the field of remote sensing, and studying pixel level information and its relationships is of great significance. However, the coherent speckle noise and high sea state conditions caused by factors such as wind speed in SAR images pose challenges to feature extraction. Due to the fixed shape of convolutional kernels in convolutional neural networks (CNNs), it is difficult for the kernels to fully cover the marine aquaculture areas in SAR images. In this paper, we utilize the structural exploration and information propagation capabilities of graph convolutional networks (GCNs) between different pixels to complete this pixel-level task. In order to reduce the connection burden between different SAR pixels and the high computational cost of the entire SAR image, this the paper uses the superpixel segmentation method to segment SAR images, reducing the previously high pixel-level computational cost. However, modeling at a single scale can only extract a single feature. To break this limit and extract More representative of the characteristics, this paper proposes multi-level Superpixel structure diagram based on U-net combination GCN is used for SAR image semantic segmentation. In the meantime, The network constructs a new residual convolution attention Module (RCA). RCA modules include CNN and Beyond Self-attention. First, through CNN convolution, which provides a detailed initial space for the model Characteristic information; Then in Features to make it more accurate. This hybrid U-Net can be utilized The characteristics of SAR images are analyzed from multi-scale and hierarchical perspectives In perspective, its performance has been proven. Extensive experiments have been carried out on data sets of aquaculture methods such as floating rafts and cages in different areas. The effectiveness of the proposed method is verified, and the Accuracy and other precision of the segmentation results are significantly improved, and the qualitative and quantitative performance of the proposed method is better than that of other advanced models.