关键词: 水体分割, 深度学习, 全局与局部语境信息, 高分辨率遥感, 注意力机制

Abstract: [Purposes] This study proposes a deep learning-based water body segmentation algorithm integrating global-local contextual information,aiming to improve the accuracy of water body extraction in complex environments. [Methods] A high-resolution remote sensing image dataset containing multiple types of water bodies was constructed,and a PCT-Net model was developed.By introducing an interactive attention-based feature fusion mechanism and a global-local contextual information fusion strategy,the proposed model enhances discrimination capability and segmentation stability in complex water body scenarios,including dark backgrounds,dense pond distributions,and water-background confusion.In addition,a semantic-aware dynamic upsampling module was designed to effectively alleviate boundary blurring and insufficient detail recovery of small-scale water bodies in high-resolution remote sensing imagery. [Findings] Experimental results demonstrate that the proposed PCT-Net model outperforms representative segmentation models such as U-Net,SeaFormer,DeepLabV3+,and PSPNet,achieving an overall accuracy of 99.08%, a Kappa coefficient of 0.981 3,a false positive rate of 0.91%, and a false negative rate of 0.93%. [Conclusions] The proposed method can be widely applied in water resource monitoring,flood disaster assessment,and related fields,providing technical support for regional sustainable development．

Key words: water body segmentation, deep learning, global-local contextual information, high-resolution remote sensing, attention mechanism