Kimmy De Alba, Space Systems Command A&AS; Kevin Phan, EO Solutions; Alex Cabello, EO Solutions; Zach Gazak, Space Systems Command A&AS; Justin Fletcher, Space Systems Command A&AS
Keywords: Aperture Photometry, Machine Learning, Satellites, Stars, Autonomous Telescope Networks, Light Curves, Anomaly Detection
Abstract:
Accurate visual magnitude estimation is essential for Space Domain Awareness (SDA) tasks, including collision avoidance, anomaly detection, and satellite material-state monitoring. Given the recent growth of orbital populations, the scalability and robustness of photometric methods have become critical operational concerns. To address these needs, we present a comprehensive benchmark evaluating four physics-derived aperture techniques: fixed-radius, full width at half maximum (FWHM), depth-first search with interquartile range (DFS+IQR), and an adaptive Gaussian point spread function (PSF). These techniques are assessed within both analytical and interpretable learned magnitude-estimation frameworks. Using 1.7 million synthetic image patches (1.5M stars and 0.2M satellites), our zero-shot evaluation on real imagery containing 148k stars and 3k synthetic-injected satellites achieves mean absolute errors (MAE) of 1.28 for stars and 1.21 for satellites. Subsequent fine-tuning on 2.6M stars and 19k satellites reduces MAEs to 0.92 and 0.88, respectively. Our study also establishes baselines for sample efficiency through training-set sweeps from 100 to 1,600 images, quantifying both error reduction and computational efficiency. Importantly, we introduce stray-light contamination and nearest-source distance as critical input features, significantly improving model robustness in mixed observing conditions. Our study also systematically injects synthetic satellites into real imagery, quantifying the sim-to-real domain gap for large-scale photometric analysis. By explicitly addressing source crowding, stray-light conditions, sample-limited scenarios, and source-agnostic estimation without calibration frames, our method sets a new standard for scalable SDA photometry.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications