David Chaparro, EO Solutions; Kevin Phan, EO Solutions; Enrique De Alba, EO Solutions; Zach Gazak, SSC/SYD 85/BI A&AS
https://doi.org/10.64861/JKSD9968
Keywords: Fine Tuning, Simulated Datasets, Satellite Detection, Pretraining, task-informed weight initialization,
Abstract:
Modern Space Domain Awareness (SDA) systems have adopted machine learning for Resident Space Object (RSO) detection in optical imagery. However, rate-tracked RSO detection datasets are costly to acquire and specialized publicly available datasets are inapplicable to all detection tasks, impeding adoption of machine learning on new optical systems. Within existing rate tracked detection datasets, positive RSO samples frequently vary in brightness, relative velocity, and unique noise signature due to sensor specific noise. RSO detection datasets are also frequently under sampled as images may contain 0-3 RSOs per image and consist of less than 1% of the image area. RSOs are also commonly located in the center of an image, under-sampling the signal of the RSO in combination with sensor specific noise. Fortunately, physics informed simulators such as SatSim can generate semi-realistic telescope imagery. These simulators can also generate an arbitrary number of RSO samples, allowing simulated datasets to completely sample the RSO visual state space. Similar to pretraining for large language models, we hypothesize object detection models with weights initialized with a task-informed dataset may reach more optimal local minima. To our knowledge, we present the first ablation study on simulation based pretraining strategies for RSO detection on real RSO-imagery.
Rather than generating physically representative scenes, we construct images containing 50 RSOs per frame, efficiently sampling the RSO appearance space yielding distributions unattainable through real-world observation. We generate three simulated datasets: a realistic scene set of 18,000 rate-track images each containing 1-2 satellites, a set of 18,000 rate-track dense star and satellite fields which contains a realistic star background and 50 RSO’s per image, and a set of 18,000 rate-track dense satellite field containing 50 RSO’s per image without stars. All simulated imagery is generated with a randomly assigned per-pixel field of view (FOV) between 0.5 and 3 arcseconds. First, we train a RetinaNet model for 50 epochs on each simulated dataset to learn a broad distribution of satellite appearances. Then we fine tune each pretrained model on our real dataset consisting of 32,500 images from one telescope with data collected over one year containing 1-2 RSOs per image. We train an additional model on real data using only random weight initialization as a baseline. Our learning rate, weight decay, momentum, and batch size are all held constant at the fine-tuning stage to isolate the effect of weight initialization. We evaluate the performance of each training method by measuring the precision, recall, and F1 at the peak F1 (F1*) where inference boxes are considered true positives if they lie above a set confidence threshold and within a L2 distance of a ground truth target. Our experiments demonstrate that pretraining using realistic scene pretraining results in an F1* of 0.948 +/- 0.004, precision of 0.970 +/- 0.005 and recall of 0.926 +/- 0.007, exceeding baseline F1* by 1.5 percentage points at L2 distance of tf=1 pixel. Both precision and recall drive the increased detector accuracy, suggesting that pretraining can increase learned detection performance.
We establish that realistic scene pretraining is the strongest pretraining strategy, despite the imagery not representing the corresponding telescope configuration. We also discovered that pretraining with dense star and satellite fields does not significantly boost detection recall, and merits further investigation to optimize the number of RSO samples per image. Our work demonstrates that task-informed weight initialization enhances RSO detection performance at high F1 operating regimes where performance gains become increasingly difficult to achieve without requiring additional dataset collection. Our method can retroactively be applied to legacy datasets to reduce training time and boost final performance. Existing pipelines can easily incorporate optimized pretrained models without curating costly new datasets. Leveraging diverse simulated RSO samples consistently provides precision and recall gains establishing pretraining as a practical method to enhance high-performance detection capabilities.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications