Deep-Learning-Based Marine Aquaculture Zone Extractions From Dual-Polarimetric SAR Imagery

Efficient monitoring of marine aquaculture zones (MAZs) is crucial for facilitating coastal resource management. To achieve this, we developed a specialized deep convolutional neural network tailored for extracting MAZs from synthetic aperture radar (SAR) imagery, integrating prior analytical knowledge of MAZ imaging features. A total of 47 Sentinel-1 dual-polarized (VV and VH) SAR images spanning 2016–2023 in China's Subei Sandbanks along the Yellow Sea coast were collected due to appropriate tidal level and acquired time. We first comprehensively analyzed of normalized radar cross section (NRCS) values for MAZs under varying tidal levels and aquaculture facility structures. Rising tide-induced submergence resulted in a significant mean NRCS reduction of 7.01 dB (VV) and 4.54 dB (VH), causing MAZ signals to resemble seawater. In addition, during low tide, volume scattering from the net screen on the aquaculture rafts increased VH-polarized image recognizability, with a smaller NRCS overlap (64%) between MAZs and tidal flats compared to VV-polarized images. Hence, VH-polarized images taken during low tide with intact aquaculture facilities were selected for dataset construction due to their reliability in characterizing MAZs. Building upon the classical U-Net framework, we introduced four modifications informed by our imaging characteristics analysis to enhance the model's performance. Testing experiments demonstrated an impressive F1-score of 94.77%, highlighting the effectiveness of incorporating prior knowledge into refining deep learning models. Applying the model to SAR images from 2016 to 2023 revealed concentrated MAZs in the relatively flat southeastern Subei Sandbanks, with a noticeable scale decline post-2021 resulting in a 67.65% reduction over the years.