Scholarlink

<p>Remote sensing image change detection (CD) plays a critical part in environmental monitoring, urban expansion, and disaster assessment. However, the complexity of background variations, multi-scale target changes, and noise interference offers considerable obstacles to conventional approaches. To address these challenges, this research offers a Multiscale Siamese Transformer Attention Network (MSTANet), which blends multi-scale feature extraction, attention mechanisms, Transformer design, and Haar wavelet downsampling to boost both the accuracy and robustness of change detection. Firstly, the proposed multi-scale feature extraction module leverages depthwise separable convolution kernels of various sizes to collect multi-scale change information, while a soft-thresholding-based attention mechanism enhances the representation of change regions. Secondly, the introduction of a Transformer-based attention mechanism successfully filters critical features, suppresses irrelevant background interference, and boosts the model’s ability to capture long-range connections. Finally, Haar wavelet downsampling method preserves crucial spatial features while reducing computational complexity, considerably limiting the information loss associated with previous downsampling strategies. Experimental results on two commonly used high-resolution datasets, LEVIR-CD and WHU-CD, reveal that MSTANet achieves F1-scores of 91.37<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation> and 93.99<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation> on LEVIR-CD and WHU-CD correspondingly with just 7.90M parameters, hence verifying its superior performance in change detection tasks.</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 multiscale siamese transformer attention network for remote sensing image change detection

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Qinsheng Du,
  • Shiyan Zhang,
  • Ningbo Zhang,
  • Chao Shen,
  • Zuosheng Du,
  • Xin Guo,
  • Jian Zhao

摘要

Remote sensing image change detection (CD) plays a critical part in environmental monitoring, urban expansion, and disaster assessment. However, the complexity of background variations, multi-scale target changes, and noise interference offers considerable obstacles to conventional approaches. To address these challenges, this research offers a Multiscale Siamese Transformer Attention Network (MSTANet), which blends multi-scale feature extraction, attention mechanisms, Transformer design, and Haar wavelet downsampling to boost both the accuracy and robustness of change detection. Firstly, the proposed multi-scale feature extraction module leverages depthwise separable convolution kernels of various sizes to collect multi-scale change information, while a soft-thresholding-based attention mechanism enhances the representation of change regions. Secondly, the introduction of a Transformer-based attention mechanism successfully filters critical features, suppresses irrelevant background interference, and boosts the model’s ability to capture long-range connections. Finally, Haar wavelet downsampling method preserves crucial spatial features while reducing computational complexity, considerably limiting the information loss associated with previous downsampling strategies. Experimental results on two commonly used high-resolution datasets, LEVIR-CD and WHU-CD, reveal that MSTANet achieves F1-scores of 91.37 \(\%\) and 93.99 \(\%\) on LEVIR-CD and WHU-CD correspondingly with just 7.90M parameters, hence verifying its superior performance in change detection tasks.