Scholarlink

<p>Change detection in remote sensing is critical for monitoring environmental dynamics, urban expansion, and disaster response, yet existing methods often struggle to balance spatiotemporal feature representation with computational efficiency. To address this, we propose&#xa0;TEWA-Net, a novel encoder-decoder architecture integrating Convolutional Long Short-Term Memory (ConvLSTM) with spatial-channel attention mechanisms and a&#xa0;Temporal Evolution-Weighted Attention (TEWA)&#xa0;module. The TEWA module dynamically weights temporal features based on their evolutionary significance, while the hybrid spatial-channel attention mechanism enhances discriminative feature extraction across complex landscapes. The proposed architecture is evaluated on three benchmark optical datasets: LEVIR-CD, S2Looking, SZTAKI AirChange (SZADA and TISZADOB), and OMBRIA. On LEVIR-CD, the proposed architecture incorporating TEWA and attention mechanisms achieves up to 99.17% accuracy, significantly outperforming the baseline. On S2Looking, the proposed architecture maintains robust accuracy (~ 99.2%) and strong precision, with TEWA alone showing comparable performance to the proposed architecture. For the SZTAKI dataset, the architecture achieves up to 97.87% accuracy, with TEWA alone achieved the same accuracy as the proposed architecture. On the large-scale OMBRIA benchmark, TEWA-Net using Sentinel-2 imagery attains up to 92.5%-pixel accuracy and 84.5% frequency-weighted IoU, outperforming multimodal Sentinel-1 + 2 methods based on OmbriaNet and recent U-Net variants. TEWA-Net is evaluated under synthetic cloud occlusion; while all methods declined with increasing obscuration, TEWA-Net showed the smallest accuracy loss and highest precision, supporting its robustness for operational use in cloud-affected imagery. Overall, TEWA-Net demonstrates strong generalization across diverse change detection scenarios and achieves state-of-the-art accuracy and precision.</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

TEWA-Net: A spatiotemporal attention-enhanced ConvLSTM architecture for remote sensing change detection

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Menna M. Elkholy,
  • Marwa S. Moustafa,
  • Dina ElSayad,
  • Hala M. Ebied,
  • Mohamed F. Tolba

摘要

Change detection in remote sensing is critical for monitoring environmental dynamics, urban expansion, and disaster response, yet existing methods often struggle to balance spatiotemporal feature representation with computational efficiency. To address this, we propose TEWA-Net, a novel encoder-decoder architecture integrating Convolutional Long Short-Term Memory (ConvLSTM) with spatial-channel attention mechanisms and a Temporal Evolution-Weighted Attention (TEWA) module. The TEWA module dynamically weights temporal features based on their evolutionary significance, while the hybrid spatial-channel attention mechanism enhances discriminative feature extraction across complex landscapes. The proposed architecture is evaluated on three benchmark optical datasets: LEVIR-CD, S2Looking, SZTAKI AirChange (SZADA and TISZADOB), and OMBRIA. On LEVIR-CD, the proposed architecture incorporating TEWA and attention mechanisms achieves up to 99.17% accuracy, significantly outperforming the baseline. On S2Looking, the proposed architecture maintains robust accuracy (~ 99.2%) and strong precision, with TEWA alone showing comparable performance to the proposed architecture. For the SZTAKI dataset, the architecture achieves up to 97.87% accuracy, with TEWA alone achieved the same accuracy as the proposed architecture. On the large-scale OMBRIA benchmark, TEWA-Net using Sentinel-2 imagery attains up to 92.5%-pixel accuracy and 84.5% frequency-weighted IoU, outperforming multimodal Sentinel-1 + 2 methods based on OmbriaNet and recent U-Net variants. TEWA-Net is evaluated under synthetic cloud occlusion; while all methods declined with increasing obscuration, TEWA-Net showed the smallest accuracy loss and highest precision, supporting its robustness for operational use in cloud-affected imagery. Overall, TEWA-Net demonstrates strong generalization across diverse change detection scenarios and achieves state-of-the-art accuracy and precision.