Scholarlink

<p>Semantic segmentation of high-resolution remote sensing imagery plays a key role in land-cover mapping, urban planning, precision agriculture, and environmental monitoring. Despite recent progress, existing approaches still struggle with three persistent challenges: (1) insufficient feature expressiveness when relying solely on spatial-domain representations, (2) severe intra-class variance and low inter-class separability across large geographical spans, making it difficult to balance global semantic consistency with local detail recovery, and (3) inaccurate delineation of complex land-cover topologies and continuous transitions, exacerbated by upsampling-induced blur. To address these issues, we propose FSBNet, a U-shaped encoder–decoder framework built on a ConvNeXt-S backbone that integrates three plug-and-play modules. First, a Frequency–Spatial Interactive Attention (FSIA) module explicitly leverages frequency-domain cues via adaptive frequency decomposition and cross-domain attention, improving robustness to heterogeneous textures and repetitive patterns. Second, a Bidirectional Interactive Global Aware (BIGA) module introduces class-prototype–driven global constraints through a forward projection to category centers and a backward reconstruction to pixel features, further stabilized by confidence-weighted fusion. Third, a Boundary-Aware Upsampler (BAU) decouples boundary discovery from upsampling, applying region-adaptive interpolation to preserve sharp contours while maintaining smooth interiors. Extensive experiments on the ISPRS Vaihingen and Potsdam benchmarks and the LoveDA dataset demonstrate consistent improvements over strong CNN-, Transformer-, and hybrid baselines. FSBNet achieves 84.37% and 87.58% mIoU on Vaihingen and Potsdam, respectively, and 67.08% mIoU on LoveDA, indicating strong accuracy and generalization under diverse land-cover scenes.</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

FSB-Net: Frequency-spatial-boundary interactive network for remote sensing semantic segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月4日

查看全文
  • Chengfan Luo,
  • Qikai Lin,
  • Yunhong Ding,
  • Jingyu Liu

摘要

Semantic segmentation of high-resolution remote sensing imagery plays a key role in land-cover mapping, urban planning, precision agriculture, and environmental monitoring. Despite recent progress, existing approaches still struggle with three persistent challenges: (1) insufficient feature expressiveness when relying solely on spatial-domain representations, (2) severe intra-class variance and low inter-class separability across large geographical spans, making it difficult to balance global semantic consistency with local detail recovery, and (3) inaccurate delineation of complex land-cover topologies and continuous transitions, exacerbated by upsampling-induced blur. To address these issues, we propose FSBNet, a U-shaped encoder–decoder framework built on a ConvNeXt-S backbone that integrates three plug-and-play modules. First, a Frequency–Spatial Interactive Attention (FSIA) module explicitly leverages frequency-domain cues via adaptive frequency decomposition and cross-domain attention, improving robustness to heterogeneous textures and repetitive patterns. Second, a Bidirectional Interactive Global Aware (BIGA) module introduces class-prototype–driven global constraints through a forward projection to category centers and a backward reconstruction to pixel features, further stabilized by confidence-weighted fusion. Third, a Boundary-Aware Upsampler (BAU) decouples boundary discovery from upsampling, applying region-adaptive interpolation to preserve sharp contours while maintaining smooth interiors. Extensive experiments on the ISPRS Vaihingen and Potsdam benchmarks and the LoveDA dataset demonstrate consistent improvements over strong CNN-, Transformer-, and hybrid baselines. FSBNet achieves 84.37% and 87.58% mIoU on Vaihingen and Potsdam, respectively, and 67.08% mIoU on LoveDA, indicating strong accuracy and generalization under diverse land-cover scenes.