Kevin Phan, EO Solutions; Justin Fletcher, Space Systems Command (A&AS)
Keywords: Satellite Detection, Synthetic Tracking, Convolutional Neural Networks, Signal-to-Noise Ratio, Machine Learning, Near-Earth Object Detection, Automation.
Abstract:
Accurate astrometric localization of satellites is crucial for space domain awareness, demanding precise alignment of stars in images with real-world counterparts. However, detecting unknown objects amidst uncertainties in velocities and astrometric data poses a significant challenge. This study proposes a novel approach that integrates deep learning with synthetic tracking techniques to address these challenges effectively.
Synthetic tracking, renowned for reducing noise in short-exposure sidereal images and correcting camera imperfections, forms the foundation for refining astrometric fits. By leveraging synthetic tracking, the study enhances the accuracy of satellite localization. However, to avoid false positives, traditional synthetic tracking approaches often employ a signal-to-noise ratio (SNR) threshold, potentially missing dimmer objects and leading to false negatives (Shao et al 2018).
To mitigate false positives while capturing dimmer objects, the study integrates convolutional neural networks (CNNs). CNNs have had great success at detecting geosynchronous objects in rate tracked images (Fletcher et al 2019). CNNs excel at analyzing complex patterns and relationships in data, allowing for the detection of objects below traditional SNR thresholds with greater precision and accuracy. By driving down the SNR threshold, CNNs expand the scope of detectable objects, significantly improving the robustness of the localization process.
This study conducts a comprehensive examination of synthetic tracking as a fundamental technique for satellite detection. Employing synthetic data generated by SatSim, we assess the efficacy of standalone synthetic tracking in detecting moving objects across varying brightness levels. Performance evaluation is conducted using metrics such as precision, recall, and F1 score, aiming to understand the implications of object brightness on the performance metrics of synthetic tracking. Particularly, emphasis is placed on assessing the detection rates of dimmer objects while maintaining high precision.
Traditionally, human intervention is necessitated for verifying edge cases near the SNR threshold. This study aims to automate this process entirely by integrating Convolutional Neural Networks to screen and validate the detections provided by synthetic tracking. This integration aims to establish a robust pipeline for identifying unknown near-earth objects. CNN models will be trained using datasets from SatSim. Once trained, the CNN model will be leveraged to augment the performance of synthetic tracking on autonomous ground based observatories dedicated to space domain awareness.
Subsequently, utilizing the same datasets employed for baseline assessments, the study will evaluate the performance of the integrated pipeline. Anticipated outcomes include notable enhancements in detection metrics and improved sensitivity in detecting dimmer objects. The incorporation of CNNs is deemed pivotal for elevating detection accuracy and facilitating lower SNR thresholds, while mitigating concerns related to false positives.
The research showcases the effectiveness of this integrated methodology in localizing and detecting unknown objects with remarkable precision and accuracy. By combining synthetic tracking with CNNs, the study presents a robust pipeline for enhancing satellite detection precision and reducing false positives in satellite imagery. This research not only advances the state-of-the-art in satellite detection and identification but also offers a promising framework for applications requiring accurate and reliable object localization in satellite imagery, thereby contributing significantly to space domain awareness.
Date of Conference: September 17-20, 2024
Track: Machine Learning for SDA Applications