In recent years, the development of China’s marine aquaculture has brought serious challenges to the marine ecological environment. Therefore, it is significant to classify and extract the aquaculture zone and spatial distribution in order to provide a reference for aquaculture management. However, considering the complex marine aquaculture environment, it is difficult for traditional remote sensing technology and deep learning to achieve a breakthrough in the extraction of large-scale aquaculture zones so far. This study proposes a method based on the combination of piecewise linear stretching and R³Det to classify and extract raft aquaculture and cage aquaculture zones. The grayscale value is changed by piecewise linear stretching to reduce the influence of complex aquaculture backgrounds on the extraction accuracy, to effectively highlight the appearance characteristics of the aquaculture zone, and to improve the image contrast. On this basis, the aquaculture zone is classified and extracted by R³Det. Taking the aquaculture zone of Sansha Bay as the research object, the experimental results showed that the accuracy of R³Det in extracting the number of raft aquaculture and cage aquaculture zones was 98.91% and 97.21%, respectively, and the extraction precision of the area of the aquaculture zone reached 92.08%. The proposed method can classify and extract large-scale marine aquaculture zones more simply and efficiently than common remote sensing techniques.

The remote sensing semantic segmentation task for offshore floating aquaculture (OFA) areas faces several challenges, including dense spatial distribution of aquaculture facilities, class imbalance, omission of small targets, and limited computational efficiency. To address these issues, this study constructs an NDVI-GB dataset based on the Optimum Index Factor (OIF) and proposes a Vision Transformer with Multi-scale Feature enhancement and Dynamic Attention fusion (ViT-MFDA). Based on the Swin Transformer framework, the model designs a Spatially Separable Self-Attention Vision Transformer (SSAVT) as the backbone network. By incorporating the Spatially Separable Self-Attention (SSSA) mechanism from Twins-SVT to optimize both local and global feature extraction, and integrating Conditional Positional Encoding (CPE) to enhance positional awareness, the model achieves efficient extraction of global context and local detailed information. Additionally, a Multi-scale Dynamic Weighted Attention Decoder (MDWAD) module is designed. This module integrates multi-scale contextual information through the Pyramid Pooling Module (PPM) and the Feature Pyramid Network (FPN), and introduces a dual channel-spatial attention mechanism to adaptively calibrate target features, which effectively mitigates false and missed detection in complex marine environments. Furthermore, ViT-MFDA employs a weighted joint loss function to alleviate class imbalance issues, improving the completeness and accuracy of segmentation results. Experimental results demonstrate that ViT-MFDA achieves mIoU, mF1-score, and mPrecision scores of 90.51%, 94.95%, and 95.34%, respectively, outperforming mainstream CNN and Transformer models in comprehensive performance, particularly excelling in dense target segmentation and model generalization. The method advances efficient large-scale aquaculture monitoring with real-world applicability.