Aquaculture Areas Extraction Around Saltpans Based on Local-Attention Fusion Network

Aquaculture areas extraction based on remote sensing (RS) is vital for coastal resource utilization and production management. However, studies focusing on coastal aquaculture areas remain limited due to spectral and spatial similarities between aquaculture areas and evaporation ponds in saltpans. To tackle this issue, a local fusion network (LAFNet) for extracting coastal aquaculture around saltpans is proposed. To capture horizontal, vertical contextual relations and global semantic information, local window-multihead self-attention (LW-MSA) containing three specialized attention layers is designed. Moreover, we designed the multilevel upsampling (MLU) decoder to allow the model to retain more information and details of higher level features from the upsampling pyramid. Our datasets are derived from the multispectral Sentinel-2 imagery satellite data around the Changlu Hangu, Huaibei, and Yinggehai salt fields and introduced for the first time following recommendations from domain experts. Experimental results indicate that LAFNet achieves superior performance compared with other state-of-the-art methods on test datasets. Specifically, it attains an intersection over union (IoU) of 88.80%, surpassing the highest scores of other experimental methods by 0.53% points. Furthermore, T-SNE results prove that the proposed method can effectively distinguish between saltpans and aquaculture areas.