Scholarlink

<p>Remote sensing imagery plays a crucial role in Earth observation, supporting applications from ecological research to urban planning. Despite advancements, challenges such as jagged boundaries and loss of small-object details persist in semantic segmentation. This study proposes SGDC-DeepLab, a lightweight model based on the DeepLabv3 + framework, integrating a synergistic Gaussian–dilated convolution (SGDC) module and a lightweight feature fusion module (LFFM). The SGDC module leverages Gaussian filtering to suppress edge artifacts and enhance robustness to noise, while the LFFM efficiently integrates multi-scale features. The adaptive focal loss (AFL) function dynamically adjusts class weights to improve fitting ability. Experimental results on the LoveDA dataset show a mIoU improvement of 5.9% over the state-of-the-art D2LS model, with a 90.11% reduction in model parameters. These findings demonstrate SGDC-DeepLab’s superiority in boundary refinement and small-object segmentation. The source code for this study is open-sourced at: <a href="https://github.com/Sophie0825-hubu/SGDC-DeepLab">https://github.com/Sophie0825-hubu/SGDC-DeepLab</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

Enhancing remote sensing image segmentation with SGDC-DeepLab: a lightweight approach using Gaussian filters

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月8日

查看全文
  • Yafei Zhou,
  • Zhiyong Tian

摘要

Remote sensing imagery plays a crucial role in Earth observation, supporting applications from ecological research to urban planning. Despite advancements, challenges such as jagged boundaries and loss of small-object details persist in semantic segmentation. This study proposes SGDC-DeepLab, a lightweight model based on the DeepLabv3 + framework, integrating a synergistic Gaussian–dilated convolution (SGDC) module and a lightweight feature fusion module (LFFM). The SGDC module leverages Gaussian filtering to suppress edge artifacts and enhance robustness to noise, while the LFFM efficiently integrates multi-scale features. The adaptive focal loss (AFL) function dynamically adjusts class weights to improve fitting ability. Experimental results on the LoveDA dataset show a mIoU improvement of 5.9% over the state-of-the-art D2LS model, with a 90.11% reduction in model parameters. These findings demonstrate SGDC-DeepLab’s superiority in boundary refinement and small-object segmentation. The source code for this study is open-sourced at: https://github.com/Sophie0825-hubu/SGDC-DeepLab.