Scholarlink

<p>Remote sensing water depth inversion is not restricted by geographical location and human factors, offering advantages over traditional ship bathymetry methods in shallow sea waters. Recent studies have increasingly applied machine learning to this task. Geographically, the shallow seabed's topography is continuous, with local correlation in remote sensing imagery between neighboring and central pixels. However, most previous methods are based on the inversion paradigm of ‘reflectivity (point) → water depth value (point)’, which separates the spatial correlation among the pixels of remote sensing image. Furthermore, the size of water depth inversion images is significantly smaller than the natural images to improve the geographic location accuracy, posing challenges for constructing deep neural networks for water depth inversion. To address these issues, we propose a multi-scale spatial aware neural network (MSAN) for shallow water depth inversion based on the inversion paradigm of ‘remote sensing image (map) → water depth value (point)’. MSAN introduces the convolutional block attention module (CBAM) and feature multiplexing realized by the skip connection, enhancing the model's ability to extract the water depth features within the limited network depth. Finally, Overall Accuracy with Offset (OAO) is proposed in this article to evaluate the classification accuracy based on the magnitude of the difference between predicted and actual sample labels. Experimental results on the WorldView-2 shallow water depth inversion dataset for Mischief Reef and Taiping Island show that MSAN outperforms several representative machine learning methods, and can further improve the accuracy of shallow water depth inversion.</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

Multi-scale Spatial Aware Neural Network Based on Neighboring Information for Inversion of Shallow Water Depth

  • 期刊论文

  • 发表日期：

    2025年2月14日

查看全文
  • Zecheng Li,
  • Guizhou Zheng

摘要

Remote sensing water depth inversion is not restricted by geographical location and human factors, offering advantages over traditional ship bathymetry methods in shallow sea waters. Recent studies have increasingly applied machine learning to this task. Geographically, the shallow seabed's topography is continuous, with local correlation in remote sensing imagery between neighboring and central pixels. However, most previous methods are based on the inversion paradigm of ‘reflectivity (point) → water depth value (point)’, which separates the spatial correlation among the pixels of remote sensing image. Furthermore, the size of water depth inversion images is significantly smaller than the natural images to improve the geographic location accuracy, posing challenges for constructing deep neural networks for water depth inversion. To address these issues, we propose a multi-scale spatial aware neural network (MSAN) for shallow water depth inversion based on the inversion paradigm of ‘remote sensing image (map) → water depth value (point)’. MSAN introduces the convolutional block attention module (CBAM) and feature multiplexing realized by the skip connection, enhancing the model's ability to extract the water depth features within the limited network depth. Finally, Overall Accuracy with Offset (OAO) is proposed in this article to evaluate the classification accuracy based on the magnitude of the difference between predicted and actual sample labels. Experimental results on the WorldView-2 shallow water depth inversion dataset for Mischief Reef and Taiping Island show that MSAN outperforms several representative machine learning methods, and can further improve the accuracy of shallow water depth inversion.