Scholarlink

<p>This paper proposes a deep learning framework for land-type segmentation in optical remote sensing images, addressing challenges such as low boundary resolution, poor contrast, and environmental noise. Based on PSPNet with a ResNet-50 backbone, the model integrates deformable convolutions and the CBAM attention mechanism to enhance multi-scale feature extraction and boundary modeling. A hybrid loss combining cross-entropy and Dice loss is introduced to mitigate class imbalance and improve convergence. Experiments on two benchmark datasets demonstrate superior accuracy, especially in edge delineation and minority class segmentation. Ablation studies confirm the individual contributions of each component, and the proposed method shows strong robustness in complex land cover scenarios.</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 Remote Sensing Land Type Segmentation Network Combined with Deformable Convolution

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Yiqun Ma,
  • Wenhan Ji,
  • Yanfeng Wu,
  • Changdi Li

摘要

This paper proposes a deep learning framework for land-type segmentation in optical remote sensing images, addressing challenges such as low boundary resolution, poor contrast, and environmental noise. Based on PSPNet with a ResNet-50 backbone, the model integrates deformable convolutions and the CBAM attention mechanism to enhance multi-scale feature extraction and boundary modeling. A hybrid loss combining cross-entropy and Dice loss is introduced to mitigate class imbalance and improve convergence. Experiments on two benchmark datasets demonstrate superior accuracy, especially in edge delineation and minority class segmentation. Ablation studies confirm the individual contributions of each component, and the proposed method shows strong robustness in complex land cover scenarios.