FGANet: Fuzzy-Embedded Semantic Decoupling Network for Sea–Land Segmentation in Remote Sensing Images

Sea-Land Segmentation (SLS) of remote sensing imagery plays a vital role in coastal change analysis, intelligent ocean development, and marine resource management. However, it still faces numerous challenges, such as the difficulty in segmenting narrow waterways, spectral similarity between land and sea, and uncertainty in sea-land boundaries. To effectively overcome these challenges, we propose a fuzzy-embedded semantic information decoupling network (FGANet). First, we design a contrast-driven fusion module (CDFA) that guides multi-level feature fusion through decoupled deep features, enabling precise sea-land discrimination. Second, we develop a global prior guided fusion module (GPM) to reduce misclassification in spectrally similar regions. Finally, we introduce a fuzzy connection (FC) unit that mitigates uncertainty at the sea-land boundary through a two-dimensional Gaussian blurring approach. Comprehensive experiments are conducted on publicly available SLSD and SLBD datasets, with results confirming that the proposed FGANet outperforms other advanced existing approaches in the field.