| 摘要: |
| 近年来遥感影像技术和深度学习技术发展迅速，其中遥感图像中包含复杂多样的地理信息，深度学习技术也在日趋成熟。为了从遥感图像中获取其中包含的海洋内波信息，提出了一种优化的TransUNet框架的条纹分割算法，验证了该算法在海洋内波SAR图像明暗条纹和明条纹分割中的有效性。将优化的TransUNet框架的条纹分割算法与原始的TransUNet和UNet的分割效果进行对比，发现原始的TrasUNet和UNet存在部分条纹无法分割或分割不清楚的情况，优化的TransUNet算法的分割结果基本都比较完整且没有断裂。这表明优化过的TransUNet算法具有更好的分割效果；而且可以通过改变模型的复杂度以适应不同的数据规模。利用所提出的算法可为海洋内波的反演研究奠定基础，在海洋内波条纹分割方面具有良好的应用潜力。 |
| 关键词: 内波条纹 SAR TransUNet 深度学习 分割 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金项目 |
|
| Segmentation of internal ocean wave stripes based on deep learning network models |
|
wu xu yun,zhang yu,jia hai qing
|
|
Shanghai Marine Monitoring and Forecasting Center
|
| Abstract: |
| Remote sensing image technology and deep learning technology have been developed rapidly in recent years. Remote sensing images contain complicated and diverse geographical information, and deep learning technology is also becoming mature increasingly. In order to obtain the information of internal ocean waves contained in remote sensing images, an optimized stripe segmentation algorithm based on the TransUNet framework is proposed, and the effectiveness of the algorithm in the segmentation of bright and dark stripes and bright stripes in SAR images of internal ocean waves is verified . By comparing the segmentation effects of the optimized stripe segmentation algorithm based on the TransUNet framework with those of the original TransUNet and UNet, it is found that the original TransUNet and UNet cannot segment or cannot segment clearly some internal oceanic wave strips, while the segmentation results from the optimized TransUNet algorithm are basically complete and without fractures. This indicates that the optimized algorithm has a better segmentation effect. Moreover, the optimized TransUNet algorithm can adapt to different data scales by changing the complexity of the model. The proposed algorithm can lay a foundation for the inversion research of internal ocean waves and has great potential in the segmentation of internal ocean stripes. |
| Key words: Internal wave stripes SAR TransUNet Deep learning Segmentation |
