Aperture-X: Physics-informed aperture feature learning for robust photometry

Accurate visual magnitude estimation is central to astronomy and space traffic management (STM). With thousands of active satellites in Earth orbit, photometry must scale and remain robust to stray light and other distribution shifts where classical pipelines degrade. We present Aperture-X, an interpretable XGBoost-based photometry model designed for source morphologies encountered across telescope tracking modes, including compact point sources from actively tracked target satellites and elongated streaks from non-rate-tracked target satellites. Aperture-X fuses four physics-derived aperture measurements–fixed radius, FWHM (Full Width at Half Maximum), DFS＋,IQR (Depth-First Search+Interquartile Range), and adaptive Gaussian PSF–into a single magnitude estimate with feature-level interpretability for model auditing. Trained on 1.2M synthetic patches spanning both streaked stars and satellite point sources (868k stars, 293k satellites), Aperture-X reduces star and satellite mean absolute error (MAE) by up to 41% and 48% under simulated stray light compared to analytical baselines and modern learning-based baselines (MLP, sf-CNN, ResNet, 50, ConvNeXt-Tiny). MAE is used as the primary accuracy metric because it directly measures the average visual-magnitude prediction error in magnitudes. For sim-to-real evaluation, we introduce a three-stage curriculum that narrows the sim-to-real gap for real star photometry (43k stars), and we propose a proxy satellite injection benchmark (44k injected point sources) to assess transfer for satellite-like point sources in real backgrounds. Finally, we report Space Traffic Management (STM) oriented metrics, Outlier Rate (error > 0.3 mag) and Characterizable Rate (error < 0.1 mag). Aperture-X reduces simulated satellite outliers to below 0.1% and increases the characterizable fraction of real stars by 12% after adaptation.