Machine Classification and Sub-classification Pipeline for GEO light curves
Abstract
We have developed a pipeline of algorithms that use statistical measures of light curves to classify GEOs with minimal human intervention. This machine-based processing pipeline first determines whether the inspected object is a three-axis stabilized GEO (TAG) or a tumbling object (TO). If the object is determined to be a TO, it will be further sub-classified as either a rocket body or tumbling satellite. For light curves classified as TAG, sub-classification into signature types will ensue. The classification process for TAGs and TOs is described in this paper.
Our technique is a hybrid of assessments from (1) a Random Forest (RF) classifier, (2) a test for periodicity, (3) a test for the presence of aliasing, and (4) a test to evaluate the quality of fit of a spline model to the data. The RF uses statistical features calculated from light curves to classify an incoming object. The periodicity assessment is based on best practices in the astronomy community to determine the statistical significance of each period found. To detect aliasing, we test whether the residuals of a spline fit to the light curve are normally distributed. A lack of normality can imply aliasing caused by under sampling the periodic light curves outside the Nyquist limit. The fourth test compares the coefficient of determination (R2) of a light curve's spline fit to the empirical distributions of R2 values from objects with known stability modes. Finally, a Hidden Markov Model takes the four previous test assessments and combines the gathered evidence in a dynamic Bayesian network that classifies the light curves accordingly. After this top-level classification, the objects determined to be TO are then sub-classified to determine the object's type (e.g., an unstable TAG or a rocket body). This sub-classification is performed with a RF that is trained on light curves with known object type. For light curves already classified as TAG, sub-classification into signature types will proceed. This sub-classification is useful for analysts because the presence of signature features are highly correlated to physical structures of the satellites. The primary light curve features are narrow and broad brightness peaks, and the primary structures are the main payload enclosure (bus), the solar panels, and large communications antennas. The latter are often located on the eastern and/or western sides of the satellite and tend to break the N/S symmetry typically found in the bus and solar panels. This asymmetry is projected into the reflected light observed in photometric signatures. We have developed the SigML (Signature Machine Learning) algorithm to aid analysts in assessing the on-orbit status of GEOs. The classification scheme for signature type includes six assessments. For three longitudinal phase angle regions, the presence of narrow features and broad features are assessed separately. Narrow features include narrow peaks whereas a broad feature may either be a broad peak or a shoulder. We extract the feature information via wavelet decomposition and wavelet coefficient ratios. A high correlation between our expert-classified data and assessments produced by our automated pipeline of algorithms has been demonstrated. This paper presents a summary of the methods employed for classifying GEO objects and the test results on a multi-year dataset of selected space objects.- Publication:
-
Advanced Maui Optical and Space Surveillance Technologies Conference
- Pub Date:
- September 2019
- Bibcode:
- 2019amos.confE..53D
- Keywords:
-
- automated classifier;
- GEO light curves;
- wavelet;
- random forest