Christopher Grant, Hyperion Vector Systems
https://doi.org/10.64861/GFRI3526
Keywords: detection algorithms, neural networks, quantization, classical processing, DCNN, space domain awareness, edge deployment, model optimization, processing architecture comparison, infrared detection, dim-object detection
Abstract:
As space-based and multi-mission Space Domain Awareness (SDA) architectures proliferate, operators repurpose sensors designed for other missions, deploy detection software across heterogeneous processing hardware, and perform detection onboard within throughput constraints that preclude full-frame downlink. These realities impose evaluation criteria that single-instrument, ground-processed benchmarks cannot address. Extending the SDA-specific baseline of Merry et al. [2] across three sensing domains (ground-based visible, space-based visible, and space-based infrared), we evaluate which detection architecture families degrade least when sensor modality, processing precision, and training-data fidelity change simultaneously. The evaluation consumes the labeled sets produced by a companion archival training database [35], so every measured number traces to a physical frame and a truth record, and the false-alarm rate is computed against that database’s calibrated per-pixel noise. We define an evaluation across three dimensions not previously assessed together: sensor-modality portability, software portability under quantization, and synthetic-to-real transfer robustness. The space-infrared cell carries the measured results presented here, the visible-spectrum cells are currently in measurement, and quantization is measured at 8-bit precision, with 4-bit evaluation remaining outstanding. We report detection probability against probability of false alarm (Pd/Pfa) at fixed false-alarm-rate points on the star-removed detections. This is a deliberate departure from the average-precision conventions of the detection literature because a classical-versus-neural comparison for operational tasking depends on the false-alarm rate that downstream processing can tolerate while maintaining specified observation and awareness accuracy. Detection and estimation algorithms from adjacent sensing families, including radar, signal intelligence, and optics, can be adapted to SDA and warrant investigation. The transfer is strongest when it is structural, with similar physics mapped to similar physics. In the GEO-infrared cases measured to date, preprocessing selection dominates detector performance. A low-rank background decomposition moved one detector from chance-level performance to near-perfect detectability, while low-signal-to-noise-ratio (SNR) bins remained challenging for every single-frame detector. That limitation represents a single-frame constraint. We therefore evaluated multi-frame accumulation as a separate approach and found that naive cross-frame stacking on a fixed or blind velocity grid does not overcome this limitation. Instead, detection performance degrades with increasing stack depth as co-added streaks become smeared. This decline remains consistent across a twentyfold range of hypothesis counts, indicating that streak morphology is the dominant factor. The recovered signal on an extended streak comes from integrating along the streak within a single frame. Dynamically defining both the search grid and the accumulation process to follow the track appears to be the most promising path to achieving cross-frame gains and remains an area for future work. Within the tested synthetic data range, a smaller attention-based convolutional neural network (CNN) outperformed a larger conventional network, demonstrating that architecture can matter more than model size. However, transfer to real-world imagery remained weak, with the primary limitation being the synthetic-to-real domain gap. By comparison, transferring between the W3 and W2 infrared bands incurred little performance penalty. The results suggest that detection architecture families are generally portable across sensing domains, whereas preprocessing pipelines are not.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications