Taylor Phan, Clemson University, EO Solutions; Kimmy De Alba, SSC/SYD 85/BI; J. Zachary Gazak, SSC/SYD 85/BI A&AS
https://doi.org/10.64861/QTML1282
Keywords: Burst image restoration, multi-frame blind deconvolution (MFBD), physics-informed reconstruction, atmospheric turbulence, satellite identification
Abstract:
Positive satellite identification is a central SDA task, but short-exposure ground-based imagery is often degraded by atmospheric turbulence, limiting both human interpretation and automated recognition. Although burst-based methods such as multi-frame blind deconvolution (MFBD) can exploit temporal diversity in these data, their downstream value for satellite identification has not been well quantified. Our work quantifies how burst-based restoration affects both perceptual image quality and downstream classification performance for RSO identification, relative to operating directly on degraded imagery. We implement a wavefront-based MFBD framework within a GPU-accelerated optimization architecture, incorporating Likelihood-based Uncertainty-Constrained Iterative Deconvolution (LUCID) to stabilize iterative reconstruction. Each short-exposure frame is modeled as a spatially invariant convolution of a static object with a frame-specific point spread function derived from pupil-plane phase aberrations. Within the forward model of LUCID, atmospheric phase is represented using a differentiable parameterization embedded in a wave-optics imaging model consistent with maximum a posteriori MFBD formulations. Object structure and wavefront coefficients are jointly estimated by minimizing a Gaussian negative log-likelihood objective across all frames in the burst, enabling efficient differentiation and optimization on modern GPUs while preserving physical consistency in the forward model. The restoration framework is evaluated on the SILO: A Machine Learning Dataset of Synthetic Ground-Based Observations of LEO Satellites dataset, which provides synthetic wave-optics simulations of 38 unique RSOs observed from the AEOS 3.6-meter telescope on Mt. Haleakala. SILO includes 125 turbulence-degraded short-exposure frames per scenario, with Fried parameters ranging from 5-200 cm, early-morning observing geometries, and targets simulated at 600 km altitude and 800 km range. Each image is assigned a Space-object National Imagery Interpretability Rating Scale (SNIIRS) label based on known object features and turbulence strength. This controlled dataset allows us to measure performance across turbulence regimes and viewing geometries. Image quality improvements are quantified using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learned perceptual image patch similarity (LPIPS), allowing direct comparison between restored imagery and corresponding degraded inputs. Across all evaluated scenarios, physics-based MFBD restoration yields consistent gains in perceptual fidelity relative to unprocessed imagery. The magnitude of improvement varies with turbulence strength, with the largest gains observed in mild-to-moderate turbulence regimes where recoverable spatial information remains present in the burst. To assess operational relevance, restored imagery is propagated through pretrained convolutional neural network classifiers, such as ResNet50, ResNeXt50, and ConvNeXt_Small, for positive satellite identification. Classification accuracy is compared between degraded imagery without post-processing and imagery restored via MFBD. Post-processing yields measurable improvements in identification performance, improving classification accuracy by up to 7% relative to unprocessed inputs, with performance gains correlating with improvements in perceptual image quality metrics. These results demonstrate that restoration-induced gains in image fidelity translate directly into improved downstream recognition performance rather than cosmetic enhancement alone. Because iterative MFBD is computationally expensive, we further evaluate a learned variant that incorporates frame-quality evaluation to reduce latency. A lightweight convolutional neural network assigns per-frame utility scores prior to optimization, enabling low-information or severely degraded frames to be down-weighted or excluded from the likelihood objective. This quality-aware frame selection preserves the underlying physical forward model while reducing effective burst size and computational burden. The hybrid approach retains classification performance close to full MFBD while reducing runtime and peak GPU memory, improving the quality-latency tradeoff. Collectively, these results provide a quantitative, end-to-end assessment of burst-based post-processing for SDA applications, demonstrating that physics-informed image restoration materially improves both perceptual image quality and automated satellite identification performance compared to operating directly on turbulence-degraded imagery. By explicitly characterizing performance, robustness, and computational trade-offs across turbulence regimes, this work informs practical decisions regarding when and how burst-based post-processing should be incorporated into operational RSO identification pipelines.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization