Xin Yee, University of Colorado; Phan Dao, Applied Optimization; David Strong, Strong EO Imaging, Inc; Charles Wetterer, KRB; Benjamin Roth, U.S. Air Force Academy; Francis Chun, U.S. Air Force Academy
Keywords: Machine learning classification, spectroscopic data, grating, non-resolved signatures, GEO
Abstract:
The United States Air Force Academy (USAFA) operates the Falcon Telescope Network (FTN) to support its research program in the Space Situational Awareness (SSA) utility of satellite optical signatures and its mission to train USAFA cadets. The FTN sensors are test beds for developing measurement techniques and collecting photometric, multispectral and polarimetric data. Operated in campaign mode, the system collects data to support the development of techniques to characterize space objects using multi-modal signatures. Specifically, the sensors are equipped with diffraction grating elements to operate as slit-less spectrographs and collect the signatures analyzed in this paper.
FTN spectroscopic data has been used to demonstrate that it can effectively distinguish different stable geosynchronous satellites (GEO) using Machine Learning (ML) classification techniques. We would like to show that ML classifiers can identify a GEO using a single spectroscopic measurement and achieve that with high accuracy. The classifiers are trained with the signatures of all candidate GEOs and the incoming signature is to be classified. Because the signature of a GEO varies with time or solar phase angle, all representative signatures for each candidate GEO have to be in the training data set. Also, because of the long-term evolution of the spectral signature-effect caused by the change of the solar declination angle the training data has to be refreshed every few days. Earlier analyses showed that hundreds of signatures (per GEO) are sufficient as a training set for a successful classification.
To evaluate the accuracy of the ML classifier for 20 different GEOs, we reduce each spectrum to a vector of features which has a smaller dimension than the original spectrum. The dimension reduction is accomplished by keeping the most significant Principal Component Analysis (PCA) components or by smoothing and decimating each spectrum according to an expected spectral resolution. With the first approach, principal components that contain the largest variation of the spectral data are used. The second approach preserves as much information as allowed by the effective spectrographs resolution. The classifiers performance is evaluated by randomized test data and cross validation for the two dimension reduction approaches using all common ML classifiers. K-Nearest neighbor classification, Linear Support Vector Machine, and an ensemble classifier are used to assess the classification accuracy. The classification of spectroscopic signatures from the 20 satellites achieved accuracy as high as 48% using only five nights of spectral data. This is a drastic improvement over a 5 % accuracy if we were to pick from a uniformly distributed set of 20 labels. It is important to note that the accuracy is limited by the possibility that GEOs with similar buses can produce identical spectral signatures. We showed using a learning curve analysis that if we increase the number of training data, the classification accuracy will increase. This is a strong indicator that we can improve our classification accuracy once we collect more training data.
Date of Conference: September 19-22, 2023
Track: Satellite Characterization