Kevin Phan, EO Solutions; David Chaparro, EO Solutions; Taylor Phan, EO Solutions; Justin Fletcher, Space Systems Command A&AS
Keywords: Space Domain Awareness, Sensor Drift, Automated Target Injection, Space Situational Awareness, Object Detection, Transfer Learning, Small Aperture Telescope, Machine Learning, Deep Learning
Abstract:
In the domain of Space Domain Awareness (SDA), one of the key challenges is the variability in data quality and characteristics across different sensors. Changes in hardware, the use of multiple sensor types, and the gradual degradation of existing instruments can all lead to differences in noise profiles, sensitivity, and resolution. These variations can cause detection models to underperform, resulting in false positives or missed detections, particularly for small or dim objects such as faint satellites or space debris. Traditional solutions involve retraining detection models on newly annotated datasets for each sensor, but this requires significant manual effort, expert annotation, and time, making it impractical for large-scale or time-sensitive operations.
In this paper, we present a novel approach for sensor-specific model calibration through automated target injection. Our method leverages high-confidence detections from a heuristic annotator, segments and masks these targets, and re-injects varied versions into verified blank frames from the target sensor. This process generates synthetic training data that reflects the sensor’s native background, noise, and optical characteristics, improving the model’s sensitivity to faint objects while preserving a low false positive rate.
We further integrate this process into retraining pipeline. A baseline RetinaNet model is first trained on high-quality industrial-grade imagery, and then fine-tuned using the created dataset from the target sensor. By utilizing a pre-existing model from another sensor the adaptation process is efficient, reducing both computational cost and training time. This makes the method well-suited for rapid recalibration when switching to a new sensor, upgrading equipment, or compensating for changes in operating conditions.
Our approach eliminates the need for extensive manual annotation for each new sensor, enabling scalable and repeatable calibration across heterogeneous systems. This method significantly improves detection performance of pretrained models on new sensors, reduces false negatives, and maintains stability across varied sensor types, ensuring that SDA systems remain accurate and reliable in dynamic operational environments.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications