Scholarlink

<p>Accurate surface-water segmentation in high-resolution remote-sensing imagery remains challenging due to cluttered backgrounds, ambiguous land–water boundaries, and the elongated, anisotropic structure of river networks. We propose GC-MambaWater, a hybrid encoder-decoder framework that combines linear-complexity visual state space modeling with connectivity aware dense mask refinement for surface-water extraction. Its encoder adopts a gated context visual state space (GC-VSS) design, where a global response normalization (GRN)-based directional router adaptively reweights horizontal and vertical scan responses to better capture long-range dependencies in slender water bodies. On the decoder side, a Mask2Former-based backbone organizes multi-scale features, followed by SPCII refinement on the final mask feature and a Connect head for continuity-aware structural modeling. Experiments on the Satellite Images of Water Bodies and 2020 Gaofen Challenge datasets achieve mIoU scores of 86.69% and 86.77%, respectively. Under the same 512 × 512 protocol, GC-MambaWater improves mIoU over SegFormer-B5 by 0.26 and 0.17 points on the two datasets while reducing GFLOPs by 10.64%; compared with Mask2Former-Swin-S, the gains are 0.35 and 0.22 points with 16.11% lower GFLOPs. These results indicate that direction adaptive encoding and continuity-aware mask refinement provide a favorable balance between segmentation accuracy, structural completeness, and computational cost for high-resolution surface-water segmentation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Direction adaptive visual state-space encoding with connectivity aware dense mask refinement for high-resolution surface-water segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Mingrun Zhuang,
  • Binbin Li,
  • Zhengbo Yu,
  • Dazhi Pan

摘要

Accurate surface-water segmentation in high-resolution remote-sensing imagery remains challenging due to cluttered backgrounds, ambiguous land–water boundaries, and the elongated, anisotropic structure of river networks. We propose GC-MambaWater, a hybrid encoder-decoder framework that combines linear-complexity visual state space modeling with connectivity aware dense mask refinement for surface-water extraction. Its encoder adopts a gated context visual state space (GC-VSS) design, where a global response normalization (GRN)-based directional router adaptively reweights horizontal and vertical scan responses to better capture long-range dependencies in slender water bodies. On the decoder side, a Mask2Former-based backbone organizes multi-scale features, followed by SPCII refinement on the final mask feature and a Connect head for continuity-aware structural modeling. Experiments on the Satellite Images of Water Bodies and 2020 Gaofen Challenge datasets achieve mIoU scores of 86.69% and 86.77%, respectively. Under the same 512 × 512 protocol, GC-MambaWater improves mIoU over SegFormer-B5 by 0.26 and 0.17 points on the two datasets while reducing GFLOPs by 10.64%; compared with Mask2Former-Swin-S, the gains are 0.35 and 0.22 points with 16.11% lower GFLOPs. These results indicate that direction adaptive encoding and continuity-aware mask refinement provide a favorable balance between segmentation accuracy, structural completeness, and computational cost for high-resolution surface-water segmentation.