Coastal Aquaculture Area Extraction Based on Self-Attention Mechanism and Auxiliary Loss

With the development of deep learning in satellite remote sensing image segmentation, convolutional neural networks have achieved better results than traditional methods. In some full convolutional networks, the number of network layers usually increases to obtain deep features, but the gradient disappearance problem occurs when the number of network layers deepens. Many scholars have obtained multiscale features by using different convolutional calculations. We want to obtain multiscale features in the network structure while obtaining contextual information by other means. This article employs the self-attention mechanism and auxiliary loss network (SAMALNet) structure to solve the above problems. We adopt the self-attention strategy in the atrous spatial pyramid pooling module to extract multiscale features while considering the contextual information. We add auxiliary loss to overcome the gradient disappearance problem. The experimental results of extracting aquaculture areas in the Jiaozhou Bay area of Qingdao from high-resolution GF-2 satellite images show that, in general, SAMALNet achieves better experimental results compared with UPS-Net, SegNet, DeepLabv3, UNet, DeepLabv3+, and PSPNet network structures, including recall 96.34%, precision 95.91%, F1 score 96.12%, and MIoU 92.60%. SAMALNet achieved better results extracting aquaculture area boundaries than the other network structures listed above. The high accuracy of the aquaculture area can provide data support for the rational planning and environmental protection of the coastal aquaculture area and promote more rational usage of the coastal aquaculture area.