Automatic Extraction of Marine Aquaculture Areas Based on the Semantic Segmentation Network from GF-3 SAR Data

In recent years, mariculture has been extensively developed in China. However, this rapid and uncontrolled expansion poses risks to the coastal environment, economic development and biodiversity conservation. Synthetic aperture radar (SAR) remote sensing can overcome the limitation of weather and has the advantage of all-day observing, which can effectively image the floating aquaculture area (FAR). In this study, we applied various deep learning methods to extract FAR by high-resolution Gaofen-3 (GF-3) SAR data. The trained and predicted datasets were constructed and trained by using U-Net, FCN, and SegNet networks. The results show that the U-Net and SegNet networks are more suitable than FCN, SegNet, and DeeplapV3+ networks for mariculture raft extraction in different scenarios. The best overall accuracy, F1 measure and kappa coefficient best overall accuracy are 95%, 0.84 and 0.80, respectively.