Fuzzy-Aware Multiview Aggregation Network for Sea–Land Segmentation of Remote Sensing Images

Accurate sea–land segmentation (SLS) of remote sensing images is crucial for monitoring coastline dynamics, which is vital to coastal regions. In remote sensing images, due to the ambiguity of the sea-land boundary, the complexity of the scene, and object scale variation, precise SLS faces significant challenges. In response to the above challenges, we propose a fuzzy-aware multi-view aggregation network (FMANet) for accurate SLS of remote sensing images. Specifically, FMANet divides the input image into global view and local view, and integrates their features into a single data stream. Multi-view feature aggregation module (MFAM) is presented to establish long-distance and deep-level visual interactions among multiple views. Fuzzy-aware module (FAM) is proposed to perceive and refine category boundaries by establishing fuzzy relationships between adjacent pixels and effectively modeling the fuzzy regions of the sea-land boundary. Coarse-fine feature fusion module (CFFM) is put forward to alleviate the misalignment issue of adjacent local views by fusing the coarse segmentation results in the final prediction stage. Comprehensive experimental results on three benchmark SLS datasets show that the proposed FMANet outperforms existing leading methods, confirming its superiority. Additionally, the ablation experiments further verify the effectiveness of each proposed module.