From Sim to 6DOF: Deep Learning for Real-Time Satellite Pose Estimation from Resolved Ground-Based Imagery

This dissertation presents the first practical system for automated six degrees of freedom (6DOF) satellite pose estimation from resolved ground-based adaptive optics (AO) imagery. Addressing a key challenge in Space Domain Awareness (SDA), the proposed approach eliminates the need for human labeling by directly regressing satellite orientation and position from blurry, noisy, and deeply-shadowed imagery. The architecture consists of a multi-stage deep neural network pipeline that localizes the satellite, predicts pose, and optionally smooths predictions over time. Networks are trained exclusively on fully synthetic imagery generated from a 3D CAD model. Despite this, the model generalizes effectively, bridging the Sim2Real domain gap and overcoming the nonexistence of real labeled data. On 137 real, human-labeled test images of Seasat, it achieved a mean rotation error of 5° and a mean image-plane translation error of 21 cm. Independent evaluation of additional real Seasat images rated 178 of 199 predicted poses as “ground truth equivalent” or “high confidence match,” with zero catastrophic failures. The approach was extended to seven degrees of freedom (7DOF) to handle articulating components and demonstrated on the Hubble Space Telescope (HST). On a 249-frame real, labeled pass of HST, applying temporal processing yielded a mean rotation error of 5.9°, a mean image-plane translation error of 48 cm, and a mean symmetry-adjusted solar array rotation error of 6°. The combined 586 real test images of Seasat and HST span multiple decades and were captured under diverse pose, illumination, and atmospheric conditions from separate ground sites. On a high-fidelity wave optics (HFWO) synthetic test set featuring Seasat in varied poses and illumination conditions, the model achieved 8.4° mean rotation error, 34 cm image-plane translation error, and 1.4% range error for a typical atmosphere (r₀ = 6 cm) and mean target range of 1,031 km. The system outperformed a human labeler in both accuracy (48% reduction in rotation error) and speed (800× faster, 7.1 Hz inference), running on consumer-grade hardware. End-to-end data generation and training required less than 40 hours on a single A100 GPU to produce a model suitable for long-term deployment. The approach was also demonstrated on a smaller satellite (ARGOS) with strong geometric symmetry using HFWO synthetic test data. A General Image-Quality Equation (GIQE)-based image quality metric was introduced to forecast pose accuracy, marking the first comprehensive study of directly regressed pose performance as a function of image quality. Key factors affecting pose accuracy were systematically analyzed, including object pose, illumination direction, Sun azimuth and zenith angles, CAD model fidelity, and training set size. Additionally, generalist models such as GPT-4o and Depth Anything V2 failed across most SDA pose estimation tasks. However, vision language models showed rapid improvement, warranting ongoing evaluation by the pose estimation and SDA communities. These results establish a new operational baseline and demonstrate, for the first time, reliable satellite pose estimation from AO SDA imagery in real time.