ECPNet: A Multimodal Sea–Land Segmentation Network With Edge and Confidence Priors

Sea-land segmentation in remote sensing imagery is of great significance for dynamic coastline monitoring. However, single-modality remote sensing images often struggle to capture the complex structure of coastlines effectively. To address this challenge, we propose ECPNet, a multi-modal sea-land segmentation network that integrates optical and SAR images. The complex and variable nature of coastlines, coupled with dynamic factor disturbances such as waves and tides, further exacerbate the inherent feature misalignment between optical and SAR modalities due to their fundamentally different imaging mechanisms. Moreover, optical images are easily affected by weather conditions, while SAR images are affected by speckle noise,both of which severely affect the confidence of modal features and limit segmentation accuracy. To overcome these challenges, we propose the Edge-Guided Deformable Alignment Module (EDAM), which uses edge information as a structural prior to guide fine-grained, pixel-level adaptive alignment and boundary enhancement of shallow optical and SAR features. In addition, we introduce the Triple Confidence-guided Fusion Module (TCFM), which dynamically evaluates the reliability of features from each source and adaptively adjusts their fusion weights. This mechanism effectively suppresses the adverse effects of clouds, haze, and speckle noise. Experimental results on dataset based on Sentinel data demonstrate that our method achieves more accurate segmentation and sharper boundary delineation compared to existing approaches, under both clear and cloud-covered conditions. The dataset is publicly available at https://huggingface.co/datasets/cuibinge/Multimodal-Sea-Land.