Widy Wang, University of Toronto; Sean Wolfe, University of Toronto; M. Reza Emami, University of Toronto
https://doi.org/10.64861/CUXU8679
Keywords: Temporarily bound objects, Near-Earth asteroids, Spaceborne object detection, transformer network, Machine learning, Cislunar space domain awareness
Abstract:
Temporarily Bound Objects (TBO), colloquially referred to as minimoons, constitute a dynamically transient population of small bodies that become gravitationally bound to the Earth-Moon System (EMS) for finite durations. TBOs are scientifically valuable, as they provide insight into small-body population statistics, Earth-Moon capture dynamics, and potential low-delta-v mission targets. From the perspective of Space Domain Awareness (SDA), TBO detection poses a significant challenge, as their sub-meter sizes, high apparent angular rates, and low Signal-to-Noise Ratios (SNR) place them near the limits of current observation systems. Thus, accurate identification of TBOs requires high-performance sensing systems and advanced AI-based detection algorithms._x000D_
_x000D_
For detecting TBOs, spaceborne platforms offer several advantages compared to ground-based observatories, including the absence of atmospheric extinction, reduced and more stable background brightness, diffraction-limited image quality, flexible observation geometry, and continuous operation independent of weather and daylight constraints. However, microsatellite-class platforms impose strict mass, power and onboard computational limitations. Under these constraints, AI-based detection methods are attractive, because computationally intensive training can be performed a priori on the ground, while lightweight inference can be executed onboard in real time. Among such methods, video transformer networks operating on short-exposure image sequences (tracklets) are of particular interest, as they exploit correlations across both spatial and temporal dimensions. Furthermore, transformer networks inherently filter out common streak-like artifacts, such as cosmic rays, which exhibit motion inconsistent with orbital dynamics._x000D_
_x000D_
Classical detection pipelines, such as matched filtering and synthetic tracking, address motion-induced signal degradation, commonly referred to as trailing loss, where the apparent motion of an object during an exposure spreads its flux across multiple pixels, reducing peak SNR. Matched filtering increases SNR within a single image by correlating the data with a streak-shaped template that models motion during the exposure. Synthetic tracking mitigates motion-induced signal loss by shifting and co-adding multiple short exposures along hypothesized velocity vectors, requiring search across velocity space and leading to substantial computational complexity. Both approaches rely on the assumption that apparent motion can be approximated as linear over the relevant observation window. For spaceborne observers in cis-lunar space, relative acceleration between observer and target can introduce parallax-induced curvature over extended observation intervals, potentially challenging the linear motion assumption for nearby objects such as TBOs._x000D_
_x000D_
It has been demonstrated that video transformer architectures trained on synthetically generated Near-Earth Asteroid (NEA) tracklets can detect genuine NEAs. Extension of these methods to the more extreme TBO regime, however, introduces several limitations. The existing frameworks for synthetic tracklet generation frequently assume independence between phase angle, observer-object distance, and apparent angular rate when estimating apparent magnitude, neglecting correlations intrinsic to capture dynamics. Additionally, prior implementations replicate grayscale frames across R, G, and B input channels, inefficiently utilizing input dimensionality. Finally, transformer-based detection performance has not been evaluated under microsatellite spaceborne constraints, where exposure time directly influences smear, SNR distribution, and classification difficulty._x000D_
_x000D_
In order to enable increased detection rates of TBOs under realistic SDA constraints, this work addresses these limitations within a unified sensing and classification framework, and verifies the improved algorithm in both space-based and ground-based scenarios. A Commercial Off-The-Shelf (COTS) microsatellite optical payload model is developed, incorporating aperture diameter, plate scale, quantum efficiency, read noise, dark current, exposure time, inter-frame dead time, and trailing-loss effects. Per-frame SNR is computed using a matched-filter-consistent formulation in which smear is parameterized by apparent angular rate, exposure time, and point spread function (PSF) width. For ground-based comparison, a Sierra Remote Observatories (SRO) survey-class configuration is modeled alongside a microsatellite platform based on the Sun–EMS L1 to enable direct comparison under consistent photometric assumptions._x000D_
_x000D_
An N-body-integrated synthetic TBO population exceeding 20,000 members is used to derive observer-dependent joint distributions of minimum apparent magnitude and apparent angular motion. Sampling at the epoch of minimum apparent magnitude preserves physical brightness–geometry correlations. These distributions are combined with the payload model to compute SNR statistics across observing configurations. An exposure-time study evaluates how cadence influences smear, SNR distribution, and transformer-based classification performance._x000D_
_x000D_
Synthetic tracklets are generated by injecting photometrically scaled streak kernels into genuine background image data, with photon-counting statistics and detector noise applied to both source and background. The generation pipeline preserves coupled geometric and photometric dependencies. Transformer inputs are restructured by assigning distinct temporal frames to separate input channels, increasing effective tracklet depth without modifying network architecture._x000D_
_x000D_
The results indicate that the COTS microsatellite platform produces a more favorable SNR distribution relative to the SRO ground-based configuration, owing to the improved observation geometry and absence of atmospheric attenuation. The spaceborne observer yields a lower mean apparent magnitude across the synthetic TBO population, shifting a larger fraction of objects into the detectable regime. The exposure-time study demonstrates that cadence selection significantly influences classification performance. Channel-restructured inputs improve detection threshold by enabling longer effective tracklets. Finally, the combined COTS spaceborne payload and improved transformer-based detection framework result in a higher fraction of the synthetic TBO population being classified as detectable compared to the ground-based equivalent._x000D_
_x000D_
By integrating physically consistent sensing models with improved transformer-based sequence classification, this work establishes a systems-level methodology for evaluating AI-enabled faint-object detection at the limits of current SDA capability. While motivated by minimoons, the framework is directly applicable to other faint, high-angular-rate objects in cis-lunar space, and informs both algorithmic development and small-satellite architecture trade studies for future SDA systems._x000D_
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications