DEGAN: A Deep Learning Model for SAR Image Segmentation in Aquacultural Raft Monitoring

Offshore aquacultural raft information extraction from synthetic aperture radar (SAR) images is crucial for large-scale marine resource exploitation and environmental protection. In this paper, a deep learning model named Detail-Enhancing Generative Adversarial Network (DEGAN) is proposed for SAR image segmentation to monitor aquacultural rafts. DEGAN incorporates several key modules to enhance segmentation performance, including <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$f$</tex>-divergence to improve the discrim-inator's ability to handle noisy data, and the Convolutional Block Attention Module (CBAM) for more effective spatial and channel attention. Within CBAM, multi-scale feature fusion is employed to capture both fine and coarse details across different resolutions. Additionally, the Simple, Parameter-Free Attention Module (SimAM) is introduced to enhance feature refinement by deriving neuron importance through an energy function. These modules, when integrated into the GAN framework, work together to improve noise suppression and detail preservation, critical for accurate segmentation in challenging SAR images. Experimental results demonstrate that DEGAN outperforms five baseline models in four performance metrics, making it highly effective for aquacultural raft monitoring in complex marine environments.