Progressive Semantic-Guided Network for the Extraction of Raft Aquaculture Areas from Remote Sensing Images

Accurate monitoring of raft aquaculture areas is particularly important for raft aquaculture planning and management. However, due to natural and human factors such as tidal changes and crop harvesting, the spectral response in some aquaculture areas is weak, leading to omissions and incompleteness in extraction results. To address this problem, we propose a progressive semantic-guided network (PSGNet) for accurate extraction of raft aquaculture areas from remote sensing images. Specifically, inspired by the human visual system, we introduce a feature enrichment module (FEM) with parallel dilated convolution to capture more discriminative features of aquaculture areas, then a partial decoder is used to aggregate high-level features and generate an initial semantic map. In addition, We propose a semantic-guided module (SGM) that progressively suppresses background information and enhances feature response in aquaculture areas through dual-branch semantic guidance. The experimental results on the GF-1 aquaculture area dataset have shown that the proposed PSGNet performs better than other models in extracting raft aquaculture areas and significantly reduces the omissions and incompleteness of extracted aquaculture areas, with the F1-score reaching 0.91, which is 2% higher than SOTA models. Furthermore, the experimental results on the Sentinel-2, GF-2 and Dongtou datasets verify the generalization ability of PSGNet.