计算机应用研究2026，Vol.43Issue(3)：917-923,7.DOI:10.19734/j.issn.1001-3695.2025.05.0225
基于异质图和Mamba的跨模态遥感语义分割
Crossmodal remote sensing semantic segmentation based on heterogeneous graph and Mamba
摘要
Abstract
To address the significant crossmodal feature heterogeneity and inefficient deep semantic interaction in semantic segmentation of visible remote sensing images,this study proposed a crossmodal heterogeneous graph-guided Mamba network(CHGMNet).The method designed a crossmodal heterogeneous feature alignment module(CHFAM),which constructed learn-able feature similarity metrics via heterogeneous graph convolution to adaptively align spectral and geometric features within a shared semantic space,effectively alleviating dimensional mismatches across modalities.Meanwhile,it introduced a novel multi-path fusion Mamba module(MPFM)that captured multi-level fused features through a linear-complexity state space model.By integrating a multi-path adaptive architecture,the module significantly improved computational efficiency while maintaining global context modeling capability.Experimental results on two large-scale high-resolution remote sensing datasets,Vaihingen and Potsdam,demonstrate that CHGMNet significantly outperforms existing mainstream methods in mIoU,mF1,and OA me-trics,validating its superiority in crossmodal remote sensing interpretation tasks.关键词
跨模态/图卷积/视觉状态空间模型/遥感语义分割Key words
crossmodal/graph convolution/visual state space model/remote sensing semantic segmentation分类
信息技术与安全科学引用本文复制引用
叶志伟,冯青阳,刘明明,王苑,高榕,严灵毓..基于异质图和Mamba的跨模态遥感语义分割[J].计算机应用研究,2026,43(3):917-923,7.基金项目
国家自然科学基金资助项目(U23A20318,62376089,62472149) （U23A20318,62376089,62472149）
湖北省高等学校优秀中青年科技创新团队计划资助项目(T2023006) （T2023006）
湖北省科技计划立项项目(2023BEB024) （2023BEB024）
