Simulated Analysis of Processing Satellite Laser Ranging Data Using Neural Networks Trained by DeepLabCut

With the development of high-repetition rate laser ranging, huge amount of laser ranging data are generated. DeepLabCut is a novel automatic annotation method for markerless motion capture from big data. With the advantage of only small size of training dataset needed, it has been successfully applied to various research fields. However, few researches can be found related to data processing for satellite laser ranging using DeepLabCut method. In this paper, different satellite laser ranging data are simulated by two polynomials according to the characteristics of echoes and noise. Secondly, two extraction strategies of time-drift and global-uniform are proposed for key points extraction to generate training datasets as ground truth. And two training datasets including 50 key points from 10 frames and 5 frames are generated, respectively. Then, deep neural networks are trained using DeepLabCut based on the training datasets. Finally, satellite laser ranging data as videos are tested with the trained neural networks. Results show that the key points suffered from drift and mismatching without uniform distribution, which indicates that DeepLabCut is not an applicable method based on the two proposed extraction strategies for satellite laser ranging data processing. Possible reasons including image textures, indiscrimination of echoes and noise are concluded. The simulation analysis in this paper is useful for deciding whether to apply DeepLabCut to process satellite laser ranging data.