College of Ocean Science and Engineering, Shanghai Maritime University, Shanghai 201306, China
2.
Translational Research Institute of Brain and Brain-Like Intelligence, Shanghai Fourth People's Hospital, School of Medicine, Tongji University, Shanghai 200434, China
Funds:
The National Natural Science Foundation of China (NSFC) under contract No. 51679132; the Science and Technology Commission of Shanghai Municipality under contract Nos. 21ZR1427000 and 17040501600.
The development of oceanic remote sensing artificial intelligence has made possible to obtain valuable information from amounts of massive data. Oceanic internal waves play a crucial role in oceanic activity. To obtain oceanic internal wave stripes from synthetic aperture radar (SAR) images, a stripe segmentation algorithm is proposed based on the TransUNet framework, which is a combination of U-Net and Transformer, which is also optimized. Through adjusting the number of Transformer layer, MLP channel, and Dropout parameters, the influence of over-fitting on accuracy is significantly weakened, which is more conducive to segmenting lightweight oceanic internal waves. The results show that the optimized algorithm can accurately segment oceanic internal wave stripes. Moreover, the optimized algorithm can be trained on a microcomputer, thus reducing the research threshold. The proposed algorithm can also change the complexity of the model to adapt it to different date scales. Therefore, TransUNet has immense potential for segmenting oceanic internal waves.
Figure 5. The effect of Transformer layer on loss rate.
Figure 6. The effect of MLP channel on loss rate.
Figure 7. Test set segmentation results of different MLP channels. The original image (a), MLP channels are 128 (b), 256 (c), 512 (d), 768 (e), and 1024 (f).
Figure 8. The effect of Dropout on loss rate.
Figure 9. Model performance analysis of the original TransUNet (a) and the optimized TransUNet (b).
Figure 10. Qualitative comparison of different approaches by visualization. The original image (a), different approaches by original TransUNet (b), optimized TransUNet (c) and U-Net (d).
Figure 11. Results of the entire SAR image segmentation (4903×5151).
Figure 12. Small-scale SAR image segmentation results, (b) Resolution size is 334 × 305 (c) resolution size is 282 × 348.
Figure 13. Results of the entire SAR image segmentation（9501×7915ALOS-ALPSRP 055520130-L1.1）.
DownLoad:
