Scholarlink

<p>Remote sensing image segmentation is fundamental for high-precision urban planning and environmental monitoring. Despite recent advances, deep learning models still struggle with severe semantic overlapping between land-cover categories with high spectral similarity as well as boundary uncertainty caused by complex spatial textures. To address these issues, we propose a Dual-branch Uncertainty-aware Haar-enhanced Network (DUHA-Net). Our framework synergistically integrates a ConvNeXt branch to capture fine-grained local structural details and a Swin Transformer branch to model long-range global context. A Haar Enhancement Module (HEM) is introduced to refine boundary accuracy in the frequency domain by reinforcing high-frequency spatial information. Furthermore, a Semantic-Guided Decoupling Aggregation (S-GDA) module is designed to explicitly mitigate inter-class confusion by decoupling overlapping semantic features. To enhance the reliability of feature fusion, an Uncertainty-aware Dual-branch Training (UDT) strategy is employed to adaptively weight the contributions of different branches. Extensive experiments on the ISPRS Vaihingen and Potsdam datasets demonstrate that DUHA-Net achieves state-of-the-art performance, reaching 58.89% and 65.90% mIoU, respectively, with superior capability in handling semantic ambiguity and preserving fine object boundaries. The results demonstrate the effectiveness and robustness of DUHA-Net, making it well-suited for practical tasks such as urban land-cover mapping, building extraction, and environmental change analysis. The source code is available at: <a href="https://github.com/gy0616/my-project">https://github.com/gy0616/my-project</a>.</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

Enhanced remote sensing image segmentation via dual-branch uncertainty-aware haar-enhanced network

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Weiya Shi,
  • Yue Guo,
  • Saiyang Liu,
  • Mingxin Yang,
  • Dezhong Pei

摘要

Remote sensing image segmentation is fundamental for high-precision urban planning and environmental monitoring. Despite recent advances, deep learning models still struggle with severe semantic overlapping between land-cover categories with high spectral similarity as well as boundary uncertainty caused by complex spatial textures. To address these issues, we propose a Dual-branch Uncertainty-aware Haar-enhanced Network (DUHA-Net). Our framework synergistically integrates a ConvNeXt branch to capture fine-grained local structural details and a Swin Transformer branch to model long-range global context. A Haar Enhancement Module (HEM) is introduced to refine boundary accuracy in the frequency domain by reinforcing high-frequency spatial information. Furthermore, a Semantic-Guided Decoupling Aggregation (S-GDA) module is designed to explicitly mitigate inter-class confusion by decoupling overlapping semantic features. To enhance the reliability of feature fusion, an Uncertainty-aware Dual-branch Training (UDT) strategy is employed to adaptively weight the contributions of different branches. Extensive experiments on the ISPRS Vaihingen and Potsdam datasets demonstrate that DUHA-Net achieves state-of-the-art performance, reaching 58.89% and 65.90% mIoU, respectively, with superior capability in handling semantic ambiguity and preserving fine object boundaries. The results demonstrate the effectiveness and robustness of DUHA-Net, making it well-suited for practical tasks such as urban land-cover mapping, building extraction, and environmental change analysis. The source code is available at: https://github.com/gy0616/my-project.