Scholarlink

<p>This paper presents a novel Cascaded Transformer with Double Branch Co-Attention UNet (CTDB-UNet) for Semantic Change Detection (SCD) in Remote Sensing Images (RSIs). Existing methods often suffer from poor correlation between dual-module architectures, leading to suboptimal accuracy in detecting changes in land surface and cover. Our approach enhances the correlation between semantic feature maps and attention units, addressing these limitations. The Cascaded Transformer Block (CTB) efficiently extracts rich semantic features, which are processed by the Double Branch Co-Attention UNet (DBCoA-UNet) to identify critical features and clarify edge information. Additionally, the Co-Attention mechanism resolves ambiguities among feature maps from both transformers. From the experimental results, the proposed model attains the overall accuracy (OA) of 83.89%, 91.25% and 87.84% for SECOND, HR-SCD and NAFZ datasets respectively. Moreover, the proposed method achieves 72.45% mean Intersection over Union (mIoU) and 22.68% Sek on the HR-SCD dataset, outperforming state-of-the-art techniques like USSFC-Net and FWA-BiLSTM by up to 4.9% in overall accuracy, demonstrating its effectiveness in SCD.</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

CTDB-UNet: cascaded transformer with double branch co-attention U-net for semantic change detection

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • C. Priya,
  • K. Sundara Krishnan

摘要

This paper presents a novel Cascaded Transformer with Double Branch Co-Attention UNet (CTDB-UNet) for Semantic Change Detection (SCD) in Remote Sensing Images (RSIs). Existing methods often suffer from poor correlation between dual-module architectures, leading to suboptimal accuracy in detecting changes in land surface and cover. Our approach enhances the correlation between semantic feature maps and attention units, addressing these limitations. The Cascaded Transformer Block (CTB) efficiently extracts rich semantic features, which are processed by the Double Branch Co-Attention UNet (DBCoA-UNet) to identify critical features and clarify edge information. Additionally, the Co-Attention mechanism resolves ambiguities among feature maps from both transformers. From the experimental results, the proposed model attains the overall accuracy (OA) of 83.89%, 91.25% and 87.84% for SECOND, HR-SCD and NAFZ datasets respectively. Moreover, the proposed method achieves 72.45% mean Intersection over Union (mIoU) and 22.68% Sek on the HR-SCD dataset, outperforming state-of-the-art techniques like USSFC-Net and FWA-BiLSTM by up to 4.9% in overall accuracy, demonstrating its effectiveness in SCD.