Kimmy De Alba, U.S. Space Force, Space Systems Command; Enrique De Alba, EO Solutions; Alexander Cabello, EO Solutions; Zach Gazak, U.S. Space Force, Space Systems Command; Justin Fletcher, U.S. Space Force, Space Systems Command
https://doi.org/10.64861/GSML3153
Keywords: Aperture photometry, astronomical image processing, robust magnitude estimation, feature fusion, sim-to-real transfer, interpretable machine learning, satellite photometry, domain adaptation.
Abstract:
Photometric light curves provide a passive and comparatively low-cost means of characterizing unresolved resident space objects (RSOs). They can extend space situational awareness (SSA) and space domain awareness (SDA) beyond orbit determination by supporting inferences about attitude, configuration, activity, and anomalous behavior. Their operational value, however, depends on more than accurate magnitude estimates in isolated frames. A useful extraction pipeline must also preserve the temporal structure and pose-dependent information contained in the full brightness sequence, even when the tracked target position is imperfect. The goal of this research is to determine when different photometric methods produce light curves that support the same physical interpretation and when apparently informative differences are instead caused by extraction artifacts. To address this problem, we develop an end-to-end evaluation framework that connects source-level photometric performance with downstream pose-template discrimination under controlled localization uncertainty.
The study uses 5,040 synthetic electro-optical images of an International Space Station model generated with the Computational Omni-Analysis Simulation Toolkit and SatSim. The dataset contains twelve canonical initial attitudes. Each attitude is represented by a seven-minute linear pass sampled at 1 Hz, producing 420 frames with known visual magnitude, target centroid, and generating pose. Four complete photometric pipelines are compared: tuned circular-aperture photometry, tuned satellite point-spread-function photometry, a depth-first-search and interquartile-range method known as DFS+IQR, and Aperture-X, a supervised fusion model that combines classical and adaptive photometric features. Development and testing are conducted with six-fold grouped cross-validation at the pass level. Complete attitude passes are assigned to training, validation, or test sets so that temporally adjacent frames cannot leak across splits.
Performance is evaluated using complementary metrics. Valid-estimate coverage measures whether each method produces a usable satellite magnitude. Frame-mean mean absolute error (MAE) measures photometric accuracy while weighting each frame equally. Downstream utility is assessed with the Pose Template Separability Index (PTSI), which summarizes pairwise distance among attitude-dependent light curves, and the Light-Curve Smoothness (LCS) score, which measures temporal coherence relative to the ground truth. These measures are combined in the Pose Template Utility Index (PTUI) to balance discrimination and physical plausibility. Full pairwise matrices are also retained to identify specific attitude pairs that remain difficult to distinguish. Robustness is tested with five seeded centroid-jitter distributions containing offsets of 0, 1, 2, 4, or 8 pixels. Perturbations are applied independently along each image axis at held-out test time. Twenty deterministic replicates are used for each distribution, and the same offsets are shared by all estimators. This paired design isolates sensitivity to localization error without allowing retraining on perturbed data.
The expected outcomes, supported by the current results, show that clean-condition accuracy is not sufficient for selecting an operational photometric pipeline. Without centroid jitter, Aperture-X and DFS+IQR achieve frame-mean magnitude errors of 0.0074 and 0.0079 mag, respectively, while all four methods show similar average template separability. Under localization error, their behavior diverges. DFS+IQR retains full coverage, nearly constant photometric error, and stable pose-template utility across all tested conditions. Its robustness is supported by a local target-relocking mechanism that can recover from moderate errors in the supplied centroid. Aperture-X retains full coverage, but its error increases to 0.1545 mag under light jitter, 0.3814 mag under moderate jitter, and 0.5581 mag under heavy jitter. Satellite-PSF photometry is more fragile, reaching 2.4733 mag error and 0.8669 coverage under heavy jitter. Circular-aperture photometry remains numerically stable, but only at a substantially biased baseline error near 0.449 mag.
A central contribution is the demonstration that template separability alone can be misleading. In the satellite-PSF case, centroid jitter increases pairwise template distance even as the recovered light curves become severely corrupted. Artificial roughness can therefore appear to improve discrimination while reducing physical fidelity. Combining separability with temporal smoothness and coverage provides a more reliable measure of whether the extracted signal remains useful for SSA/SDA inference. The resulting framework can guide estimator selection, tracker-photometry integration, confidence assessment, and test design for operational optical systems. It also supports future applications in attitude monitoring, maneuver assessment, tumbling detection, anomaly identification, and behavioral characterization. More broadly, the research establishes a reproducible path for evaluating whether photometric pipelines preserve actionable information under the localization errors expected in real tracking architectures. Future validation on larger spacecraft sets, dimmer targets, varied sensors, correlated tracking errors, and calibrated on-orbit observations will be required before the method can support direct operational deployment.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications