Alexander Cabello, EO Solutions; Claire Acke, EO Solutions; Enrique De Alba, EO Solutions; Jeff Houchard, EO Solutions; Kimmy De Alba, SSC/SYD 85/BI; Zach Gazak, SSC/SYD 85/BI A&AS; Virginia Wright, AFRL
https://doi.org/10.64861/OZOP2845
Keywords: Photometry, Atmospheric Clouds, Simulation, Photometric Calibration, Space Domain Awareness, Electro-Optical, Atmospheric Transmission
Abstract:
Ground-based electro-optical (EO) sensors for Space Domain Awareness (SDA) routinely operate through thin and partial cloud cover. Photometric measurements support satellite characterization, stability monitoring, and anomaly detection, but clouds attenuate starlight unevenly across the field of view, so the calibration stars establishing a photometric zeropoint may be attenuated differently than the target itself. The result can be a biased magnitude estimate with no obvious warning. Quantifying that error, and developing calibration robust to it, is difficult with real data because the true line-of-sight transmission is not directly known. This work addresses that gap by generating cloud-impacted EO imagery with exact transmission truth, measuring how photometric accuracy degrades, and testing calibration and reliability methods for partially cloudy conditions._x000D_
_x000D_
We introduce a physically parameterized synthetic cloud model in SatSim, an open-source GPU-accelerated space scene simulator. The model renders spatially correlated cloud transmission fields with configurable morphology and consistently attenuates stars, targets, and the pre-cloud sky background while adding cloud-scattered glow. Randomized observing conditions include a lunar scattering model with randomized phase and boresight separation and urban sky glow spanning dark to bright sites. Critically, the simulation exports truth that real data cannot provide: a per-pixel transmission plane for every cloudy frame and per-object annotations recording the flux-weighted effective transmission applied to each source, together with the aperture-area mean transmission and its dispersion across the measurement aperture._x000D_
_x000D_
Using this capability, we generate a large annotated dataset spanning seven EO sensor models with varied apertures, fields of view, pixel scales, and point-spread functions. Scenes are rendered in sidereal track with synthetic GEO-like targets streaking at 15 arcseconds per second over 3-second exposures. Five cloud morphology presets, from thin uniform veils to broken cellular decks and thick sheets, are organized into seven cloudy regimes. Sky glow is randomized across all scenes and inherited by the clear twins, so clear does not mean pristine. Every cloudy scene is paired with a clear-sky twin whose deterministic content differs only by the cloud layer; detector and photon noise are independently realized. Each frame therefore has a matched noisy clear-sky reference that enables end-to-end verification that the exported truth matches the attenuation applied by the renderer. The render comprises 6,200 cloudy scenes, their 6,200 clear twins, and 50 six-frame wind sequences: 12,450 rendering tasks producing 12,700 frames. Scenes carry a random split plus two structural holdouts, one sensor and one cloud morphology, excluded from every methodology choice._x000D_
_x000D_
The processing chain performs forced aperture photometry using circular apertures for calibration stars and pill-shaped apertures following each target streak. No centroid refinement is applied, so cloud-induced flux loss is measured rather than chased. Every instrumental magnitude, for calibrators and targets alike, is first flat-field corrected for the sensor’s configured gain map. From the star measurements, we fit a ladder of zeropoint models per frame: a whole-frame sigma-clipped global zeropoint as the baseline, cross-validated Chebyshev polynomial surfaces, a local grid with bilinear interpolation, and thin-plate spline fits. Two truth-assisted references, defined on a ten percent subset retaining full noiseless truth planes, bracket the problem: a transmission-only reference that pairs perfect cloud knowledge with an instrument calibration anchored on independent clear scenes, and a source-specific oracle that isolates the measurement-noise floor. For evaluation, we construct a truth-assisted per-target uncertainty estimate from signal-to-noise, local zeropoint uncertainty, and transmission dispersion; an observable-only variant is evaluated separately. Internal consistency checks verify the chain at the render and measurement layers._x000D_
_x000D_
On the held-out test split, comparing models on identical target rows, the cross-validated local polynomial outperforms the global baseline in all three transmission bins below 0.8. Untrimmed root-mean-square error falls from 0.26 to 0.17 magnitudes at aperture transmissions of 0.5 to 0.8, from 0.90 to 0.62 at 0.2 to 0.5, and from 1.73 to 1.11 below 0.2, with scene-level bootstrap confidence intervals excluding zero in every case. The global baseline is not merely noisier but systematically biased, reaching plus 0.61 magnitudes at transmissions of 0.2 to 0.5 and plus 1.08 below 0.2, consistent with calibration stars being attenuated differently than the target. Measured against each scene’s matched clear-sky twin, local calibration removes 42 to 56 percent of the whole-frame baseline’s paired cloud-induced excess squared error. Above transmission 0.8, the ordering reverses by a small margin, so local calibration is a tool for partial to deep cloud rather than a uniform improvement; the global model also retains a 5 to 14 percentage-point availability advantage. A reliability screen whose operating point is selected on validation data alone flags 91 percent of gross errors at a 26 percent false-positive rate on the test split, and an observable-only variant retains nearly the same gross-error recall. Among test targets passing the accuracy filters, the clipped residual core has 0.07 to 0.11 magnitudes of scatter below transmission 0.5, while untrimmed errors remain heavy-tailed. On the truth subset, an independently anchored transmission-only reference reduces RMSE by a factor of 2.8 in the flat-fielded pipeline. Omitting the gain correction roughly doubles its RMSE, showing that the full benefit of auxiliary transmission sensing depends on comparable instrument calibration. All results are simulation upper bounds under known astrometry, with photometry forced at truth positions, so detection, astrometric, and association errors are out of scope._x000D_
_x000D_
Beyond photometry, the pixel-level truth can support future training and evaluation of cloud-detection and segmentation models; the paper also includes qualitative real/synthetic cloud comparisons. The work delivers an open-source simulation capability, a quantified cloud-photometry error budget, and calibration and reliability methods that recover usable measurements in simulation. Real-cloud validation remains future work._x000D_
Date of Conference: September 15-18, 2026
Track: Atmospherics/Space Weather