Detecting Satellites in Radio-Frequency Data via Semi-Supervised Learning

Researchers present a semi-supervised learning workflow for detecting and classifying satellites in radio-frequency data, combining Non-negative Matrix Factorization with expert interpretation to reduce dependence on large labeled datasets. This approach addresses the challenge of space domain awareness by leveraging unlabeled RF observations to identify patterns in satellite signals, space debris, and ionospheric conditions without extensive manual annotation.

This research tackles a significant problem in space domain awareness by developing a practical machine learning pipeline for RF signal analysis. Traditional supervised deep learning methods require substantial labeled datasets and frequent retraining as environmental conditions shift, creating operational friction for satellite monitoring systems. The semi-supervised approach presented here reduces these constraints by treating unlabeled RF observations as valuable training material, enabling pattern discovery without exhaustive manual annotation.

The methodology combines three complementary techniques: Non-negative Matrix Factorization with automatic model determination (NMFk) identifies underlying cluster structures in unlabeled data, subject-matter experts assign physical meaning to these clusters, and a final classifier learns to predict categories on future observations. This human-in-the-loop design maintains interpretability while improving scalability—a critical advantage when monitoring systems must adapt to changing RF conditions across different geographic regions or orbital environments.

For the broader space technology sector, this advancement strengthens space domain awareness capabilities, which underpins national security, satellite operations, and debris tracking. The reproducible, interpretable methodology could accelerate adoption of machine learning in RF monitoring systems operated by government agencies and commercial space operators. The research demonstrates that limited labeled data need not be a bottleneck when combining unsupervised learning with domain expertise.

• →Semi-supervised learning reduces reliance on large labeled datasets for satellite RF signal detection and classification.
• →Non-negative Matrix Factorization with automatic model determination identifies meaningful clusters in unlabeled RF observations.
• →Expert-guided interpretation ensures detected patterns have physical meaning for space domain awareness applications.
• →The pipeline enables transfer across changing RF conditions without constant model retraining.
• →This approach strengthens space monitoring capabilities for government and commercial satellite operators.