Abstract: [Purposes] To address blurred object boundaries,fragmented small targets,and regional semantic inconsistencies commonly encountered in high-resolution remote sensing image segmentation,this study introduces a fusion method that incorporates boundary priors with region-level semantic consistency. [Methods] The proposed approach employs class-agnostic geometric masks generated by SAM(segment anything model)as boundary priors and aligns them with pixel-level semantic probabilities derived from UNetFormer on a unified raster grid.At the connected-component level,a region-semantic feature cost matrix is constructed,and the region-to-class assignment is formulated as a binary linear integer programming problem solved under global semantic consistency constraints. [Findings] Experiments on three remote sensing datasets,including UAVid,show that the proposed method significantly outperforms single-model baselines in mean intersection-over-union (mIoU)and mean F1 (mF1),with notable gains in boundary completeness and semantic consistency for classes such as buildings and vehicles. [Conclusions] The method is decoupled from backbone architectures and requires no retraining,enabling its integration as an independent semantic fusion module within existing segmentation pipelines.It offers an effective solution for enhancing fine-grained semantic interpretation in high-resolution remote sensing imagery．

Key words: high-resolution remote sensing image, semantic segmentation, segment anything model, region semantic aggregation, binary integer linear programming, global consistency constraints