Scholarlink

<p>High-resolution reconstruction of seafloor topography holds significant importance for marine science and engineering. Existing super-resolution studies relying on digital bathymetric models (DBMs) as singular labels face limitations in validation against real multibeam data, and current DBM resolution and accuracy require further improvement. To address these issues, this paper proposes a seafloor topography background field super-resolution optimization scheme based on window-shifting multi-source data fusion and constructs a Residual U-Net Attention model. By incorporating satellite altimetry gravity field parameters, the model is applied to the 1’resolution GEBCO background field across two distinct study areas: the Xisha Islands and the Hawaiian-Emperor Seamount Chain. Through dynamic sliding windows and positional encoding for local feature fusion, it achieves background field optimization at target resolutions of 30″ and 15″. Comparisons with high-resolution multibeam ground truth data excluded from the training process demonstrate that the output error of the Residual U-Net Attention model is significantly lower than that of the original GEBCO background field and mainstream baseline models. Specifically, in the 15″ task, the root mean square error (RMSE) for the Xisha and Hawaiian-Emperor Seamount Chain regions substantially decreased by approximately 40.3% and 60.3%, respectively, compared to the background field. Furthermore, the model exhibits superior power spectrum fitting characteristics within the 0.1–1 cycles km<sup>−1</sup> frequency band. Additionally, a comparative analysis of eight fundamental neural network architectures regarding their super-resolution optimization capabilities confirms that multi-scale contextual modules and channel attention mechanisms effectively enhance the parallel modeling of global trends and local details while significantly suppressing multi-source data noise. This study holds practical value for improving the accuracy and multi-resolution generalization stability of seafloor topography reconstruction, providing a feasible pathway for topography reconstruction driven by multi-source physical observations.</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

Window-shifted multi-source data fusion method: Residual U-Net attention network for seafloor topography background field super-resolution optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Xingyu Chai,
  • Hongze Leng,
  • Junqiang Song,
  • Peikun Xiao,
  • Senliang Bao,
  • Jun Zhao,
  • Jincai Li,
  • Kefeng Deng,
  • Jinhui Yang

摘要

High-resolution reconstruction of seafloor topography holds significant importance for marine science and engineering. Existing super-resolution studies relying on digital bathymetric models (DBMs) as singular labels face limitations in validation against real multibeam data, and current DBM resolution and accuracy require further improvement. To address these issues, this paper proposes a seafloor topography background field super-resolution optimization scheme based on window-shifting multi-source data fusion and constructs a Residual U-Net Attention model. By incorporating satellite altimetry gravity field parameters, the model is applied to the 1’resolution GEBCO background field across two distinct study areas: the Xisha Islands and the Hawaiian-Emperor Seamount Chain. Through dynamic sliding windows and positional encoding for local feature fusion, it achieves background field optimization at target resolutions of 30″ and 15″. Comparisons with high-resolution multibeam ground truth data excluded from the training process demonstrate that the output error of the Residual U-Net Attention model is significantly lower than that of the original GEBCO background field and mainstream baseline models. Specifically, in the 15″ task, the root mean square error (RMSE) for the Xisha and Hawaiian-Emperor Seamount Chain regions substantially decreased by approximately 40.3% and 60.3%, respectively, compared to the background field. Furthermore, the model exhibits superior power spectrum fitting characteristics within the 0.1–1 cycles km−1 frequency band. Additionally, a comparative analysis of eight fundamental neural network architectures regarding their super-resolution optimization capabilities confirms that multi-scale contextual modules and channel attention mechanisms effectively enhance the parallel modeling of global trends and local details while significantly suppressing multi-source data noise. This study holds practical value for improving the accuracy and multi-resolution generalization stability of seafloor topography reconstruction, providing a feasible pathway for topography reconstruction driven by multi-source physical observations.