【摘要】 针对滨海湿地语义变化检测中存在地物破碎化、边界模糊以及因样本不足导致地物可分性差等问题，研究提出一种语言引导的空间关系感知网络GeoLingua。该方法利用基于斑块的图卷积网络实现非局部上下文信息聚合，增强对破碎区域语义一致性的建模能力；引入边界引导的注意力机制提升对模糊边界的敏感性与识别精度；通过情景化语义引导模块，构建动态语义约束以引导视觉特征学习，从而弥补因训练样本不足导致的地物可分性差等问题。在黄河口滨海湿地数据集上开展实证分析表明，变化检测任务交并比超过97%，语义分割任务平均交并比超过93%。该方法在复杂生态环境中具备优异的语义变化感知能力，并验证了融合结构建模与语义约束的有效性。更多还原

【Abstract】 Aiming at the challenges of landscape fragmentation, blurred boundaries, and limited discriminability caused by sample scarcity in coastal wetland semantic change detection, this paper proposes a language-guided spatial topology-aware network, termed GeoLingua. First, a patch-based Graph Convolutional Network(GCN) is utilized to aggregate non-local contextual information, which enhances the modeling capability of semantic consistency in fragmented regions. Second, a boundary-guided attention mechanism is introduced to improve the sensitivity and recognition accuracy for fuzzy boundaries. Furthermore, a contextualized semantic guidance module is designed to construct dynamic semantic constraints that guide visual feature learning, effectively compensating for the poor discriminability of ground objects under insufficient training data. Empirical analysis conducted on the Yellow River Delta coastal wetland dataset demonstrates that the proposed method achieves an Intersection over Union(IoU) exceeding 97% for the change detection task and a mean IoU(mIoU) exceeding 93% for the semantic segmentation task. The results indicate that GeoLingua exhibits superior semantic change perception in complex ecological environments, validating the effectiveness of integrating structural modeling with semantic constraints.更多还原