Scholarlink

<p>Remote sensing images, characterized by complex backgrounds, large-scale variations, and significant intra-class differences, pose challenges for traditional segmentation methods. This paper introduces a lightweight network, LSC-DeepLab, based on DeepLabv3+ for spatial–frequency domain information fusion. Utilizing MobileNetv2 as the backbone, the network incorporates a Simplified Strip Pooling Module (SSPM) to enhance local feature segmentation and a frequency domain branch based on the Haar wavelet transform to extract and integrate frequency components. Simultaneously, a spatial attention mechanism captures spatial position information. The Cross-Domain Attention Fusion (CAF) module integrates spatial and frequency domain information. Comprehensive experiments on two datasets demonstrate that LSC-DeepLab outperforms mainstream general and remote sensing semantic segmentation methods in accuracy, achieving a 79.15% mIoU and 88.05% mPA on the Potsdam dataset. Our code and datasets can be available at <a href="https://github.com/tqx0717/LSC.">https://github.com/tqx0717/LSC.</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 through spatial–frequency domain fusion: a lightweight deep learning approach

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Qixuan Tang,
  • Ming Zhang,
  • Dahua Yu,
  • Yinghao Fan,
  • Xuanwen Liu,
  • Liuye Yu

摘要

Remote sensing images, characterized by complex backgrounds, large-scale variations, and significant intra-class differences, pose challenges for traditional segmentation methods. This paper introduces a lightweight network, LSC-DeepLab, based on DeepLabv3+ for spatial–frequency domain information fusion. Utilizing MobileNetv2 as the backbone, the network incorporates a Simplified Strip Pooling Module (SSPM) to enhance local feature segmentation and a frequency domain branch based on the Haar wavelet transform to extract and integrate frequency components. Simultaneously, a spatial attention mechanism captures spatial position information. The Cross-Domain Attention Fusion (CAF) module integrates spatial and frequency domain information. Comprehensive experiments on two datasets demonstrate that LSC-DeepLab outperforms mainstream general and remote sensing semantic segmentation methods in accuracy, achieving a 79.15% mIoU and 88.05% mPA on the Potsdam dataset. Our code and datasets can be available at https://github.com/tqx0717/LSC.