Scholarlink

<p>Accurate extraction of water bodies from high-resolution remote sensing imagery underpins flood monitoring, water resource management, and ecological assessment. Two structural challenges persist in this task: maintaining long-range structural continuity of river and canal networks, which local convolutions fail to preserve across large spatial extents; and fine-grained boundary discrimination at water–land interfaces, where spectral similarity between water and surrounding backgrounds causes spatial-domain features to over-smooth shorelines and narrow channels. We propose a spatial–frequency dual-domain feature enhancement network for water body extraction that addresses both challenges by integrating an axial gated depthwise convolution module at deep encoder features and a frequency-domain boundary enhancement module at mid-resolution features. The axial gated module captures long-range spatial context at low-resolution encoder features, while the frequency-domain module performs band-selective spectral gain modulation at mid-level features to preserve fine boundary detail. Both modules are integrated into a ConvNeXt-Tiny encoder–decoder architecture and evaluated on three public benchmarks. Our method achieves Dice scores of 0.894 on DeepGlobe, 0.878 on Gaofen (+4.6&#xa0;pp over DiffRoad, the strongest Gaofen baseline), and 0.897 on LoveDA (water/non-water setting, +3.0&#xa0;pp). It ranks first in Dice on Gaofen and LoveDA, where complex urban waterways and shadow interference pose severe challenges, and remains competitive with the strongest transformer-based models on DeepGlobe. Bootstrap significance tests confirm that the observed gains over all convolutional baselines are statistically significant (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(p&lt;0.001\)</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>p</mi><mo>&lt;</mo><mn>0.001</mn></mrow></math></EquationSource></InlineEquation>). Ablation experiments further show that the two modules exhibit complementary precision–recall behavior when combined, supporting the dual-domain design as a practical framework for high-resolution water body mapping.</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

A spatial–frequency dual–domain network for remote sensing water body segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月19日

查看全文
  • Pingan Zhang,
  • Binbin Li,
  • Zhengbo Yu,
  • Dazhi Pan

摘要

Accurate extraction of water bodies from high-resolution remote sensing imagery underpins flood monitoring, water resource management, and ecological assessment. Two structural challenges persist in this task: maintaining long-range structural continuity of river and canal networks, which local convolutions fail to preserve across large spatial extents; and fine-grained boundary discrimination at water–land interfaces, where spectral similarity between water and surrounding backgrounds causes spatial-domain features to over-smooth shorelines and narrow channels. We propose a spatial–frequency dual-domain feature enhancement network for water body extraction that addresses both challenges by integrating an axial gated depthwise convolution module at deep encoder features and a frequency-domain boundary enhancement module at mid-resolution features. The axial gated module captures long-range spatial context at low-resolution encoder features, while the frequency-domain module performs band-selective spectral gain modulation at mid-level features to preserve fine boundary detail. Both modules are integrated into a ConvNeXt-Tiny encoder–decoder architecture and evaluated on three public benchmarks. Our method achieves Dice scores of 0.894 on DeepGlobe, 0.878 on Gaofen (+4.6 pp over DiffRoad, the strongest Gaofen baseline), and 0.897 on LoveDA (water/non-water setting, +3.0 pp). It ranks first in Dice on Gaofen and LoveDA, where complex urban waterways and shadow interference pose severe challenges, and remains competitive with the strongest transformer-based models on DeepGlobe. Bootstrap significance tests confirm that the observed gains over all convolutional baselines are statistically significant (\(p<0.001\)p<0.001). Ablation experiments further show that the two modules exhibit complementary precision–recall behavior when combined, supporting the dual-domain design as a practical framework for high-resolution water body mapping.