Scholarlink

<p>With the rapid development of remote sensing technology, multi-modal remote sensing images have been widely applied in fields such as environmental monitoring and urban planning. However, differences in imaging mechanisms among different sensors lead to radiometric, geometric, and texture differences in images, posing significant challenges to cross-modal image matching. Traditional matching methods rely on manually designed features or supervised training. They require a large amount of labeled data and perform poorly in cross-modal, multi-source heterogeneous tasks. To address this issue, this study proposes an adaptive matching method for multi-modal remote sensing images based on Hierarchical Self-Supervised Contrastive Learning (HSSCL). Its core innovations are as follows: Extract low-level, middle-level, and high-level features of multi-modal images through a deep neural network (DNN), and design a multi-level contrastive loss function to achieve accurate cross-modal feature alignment. Introduce a graph structure feature association strategy to enhance geometric consistency, and construct an adaptive feature alignment mechanism that integrates local and global information. Use a Graph Neural Network (GNN) to model the structural consistency of cross-modal images and optimize matching relationships. Experiments are conducted on four public datasets: SEN12MS (Sentinel-1/2 multispectral data), WHU-SAR-Optical (SAR and optical images), BigEarthNet (multispectral data for 43 land cover classes), and OpenSARShip (SAR ship detection data). The evaluation metrics included accuracy, recall, F1-score, response time, generalization ability, data processing efficiency, and matching accuracy in video defect analysis. The results show that compared with advanced models such as CMCNet and DFFN, the proposed method achieves the following improvements: Accuracy increased by more than 20.3%. Recall increased by 20.7%. F1-score increased by more than 20.5%. Response time optimized by more than 20.6%. Matching accuracy in video defect analysis increased by more than 43.8%. Overall generalization ability increased by 31.2%. Data processing efficiency increased by 27.5%. This method effectively captures the association of cross-modal multi-level features, significantly improves matching accuracy, robustness, and stability. It provides a new technical path for cross-modal remote sensing image matching and offers important references for the application of self-supervised learning (SSL) in remote sensing image analysis.</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

Application of hierarchical self-supervised contrastive learning in domain adaptation matching of multimodal remote sensing image

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月28日

查看全文
  • YiQiang Li,
  • ZhenBao Luo,
  • Ge Zhu,
  • Tao Chen,
  • Hui Zhao,
  • ChaoZe Zhong,
  • DaiZhong Jin,
  • YanSheng Dang,
  • Fan Yang,
  • Xiang Li

摘要

With the rapid development of remote sensing technology, multi-modal remote sensing images have been widely applied in fields such as environmental monitoring and urban planning. However, differences in imaging mechanisms among different sensors lead to radiometric, geometric, and texture differences in images, posing significant challenges to cross-modal image matching. Traditional matching methods rely on manually designed features or supervised training. They require a large amount of labeled data and perform poorly in cross-modal, multi-source heterogeneous tasks. To address this issue, this study proposes an adaptive matching method for multi-modal remote sensing images based on Hierarchical Self-Supervised Contrastive Learning (HSSCL). Its core innovations are as follows: Extract low-level, middle-level, and high-level features of multi-modal images through a deep neural network (DNN), and design a multi-level contrastive loss function to achieve accurate cross-modal feature alignment. Introduce a graph structure feature association strategy to enhance geometric consistency, and construct an adaptive feature alignment mechanism that integrates local and global information. Use a Graph Neural Network (GNN) to model the structural consistency of cross-modal images and optimize matching relationships. Experiments are conducted on four public datasets: SEN12MS (Sentinel-1/2 multispectral data), WHU-SAR-Optical (SAR and optical images), BigEarthNet (multispectral data for 43 land cover classes), and OpenSARShip (SAR ship detection data). The evaluation metrics included accuracy, recall, F1-score, response time, generalization ability, data processing efficiency, and matching accuracy in video defect analysis. The results show that compared with advanced models such as CMCNet and DFFN, the proposed method achieves the following improvements: Accuracy increased by more than 20.3%. Recall increased by 20.7%. F1-score increased by more than 20.5%. Response time optimized by more than 20.6%. Matching accuracy in video defect analysis increased by more than 43.8%. Overall generalization ability increased by 31.2%. Data processing efficiency increased by 27.5%. This method effectively captures the association of cross-modal multi-level features, significantly improves matching accuracy, robustness, and stability. It provides a new technical path for cross-modal remote sensing image matching and offers important references for the application of self-supervised learning (SSL) in remote sensing image analysis.