🤖 machine learning

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

This paper presents a semi-supervised workflow for detecting and classifying satellites in radio-frequency data that combines Non-negative Matrix Factorization with automatic model determination (NMFk) and expert-guided cluster interpretation to overcome the challenges of limited labeled datasets and variable RF conditions.

Original authors: Cade W. Trotter, Maksim E. Eren, Justin C. Holmes, J. Brent Parham, David Ewing, Boian S. Alexandrov, Gian Luca Delzanno

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Cade W. Trotter, Maksim E. Eren, Justin C. Holmes, J. Brent Parham, David Ewing, Boian S. Alexandrov, Gian Luca Delzanno

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a radio operator trying to listen for specific voices in a crowded, noisy room. The room is filled with static, wind, and the hum of the building (the background noise). Occasionally, a specific person walks by and says something unique (a satellite passing by).

The problem is that you have hours of audio recording, but no one has told you who is speaking or when they spoke. You can't just ask a human to listen to every single second of that recording; it would take forever. This is exactly the challenge scientists face when trying to track satellites using radio waves.

Here is how the authors of this paper solved that problem, using a three-step "detective" workflow:

1. The "Pattern Finder" (NMFk)

First, the team used a computer program called NMFk. Think of this like a super-smart music producer who listens to a messy recording and says, "I hear three distinct types of sounds here: the wind, the traffic, and a specific type of bird call."

Instead of needing a human to label every sound, NMFk looks at the raw data and automatically groups similar sounds together. It figures out, "Okay, there are 25 different 'types' of radio signals hiding in this mess." It doesn't know what they are yet, but it knows they are different from each other.

2. The "Expert Translator" (Subject-Matter Experts)

Now the computer has 25 piles of radio signals, but they are just numbers. This is where the human experts come in. They act like translators.

The experts look at a few examples from each of the 25 piles and say:

"Ah, this pile sounds like a Starlink satellite passing overhead."
"This pile is just background static from the atmosphere."
"This pile is a different type of satellite."

By doing this, the experts give names and meaning to the computer's piles. They turn "Pile #14" into "Satellite Detected."

3. The "Speed Reader" (XGBoost Classifier)

Once the experts have named the piles, the team trains a second computer program (an XGBoost classifier) to act as a speed reader.

This new program learns the rules: "If the signal looks like Pile #14, it's a satellite. If it looks like Pile #22, it's just static." Because the computer learned from the experts' "named piles," it can now scan through new radio data instantly and shout out, "Satellite!" or "Nothing here!" without needing a human to listen to every second.

The Results

The team tested this system on real radar data collected over three hours.

The Discovery: The system found 25 distinct patterns in the noise.
The Translation: Experts confirmed that several of these patterns were indeed satellites (specifically Starlink satellites), while others were just background space weather.
The Test: When they let the "Speed Reader" program look at new data it had never seen before, it was very good at identifying the satellites. It successfully found about 9 out of 15 expected satellites in the test group, and it matched the experts' labels almost perfectly.

Why This Matters

Usually, to teach a computer to find satellites, you need thousands of hours of data that humans have already labeled (e.g., "This is a satellite," "This is noise"). That is expensive and slow.

This paper shows a smarter way: Let the computer find the patterns first, let a human explain what those patterns mean, and then teach the computer to recognize them on its own. It's like teaching a child to recognize dogs by showing them a few pictures and saying, "That's a dog," rather than showing them every single dog in the world and asking them to guess.

The authors conclude that this "semi-supervised" approach is a practical way to keep track of space debris and satellites, even when we don't have a lot of pre-labeled data to start with. They are now working on making the system even better by looking at different types of radio waves to catch even more objects.

Technical Summary: Detecting Satellites in Radio-Frequency Data via Semi-Supervised Learning

Problem Statement
Radio-frequency (RF) monitoring is essential for Space Domain Awareness (SDA), providing observations of satellites, space debris, and background ionospheric conditions. However, RF datasets are often large, unlabeled, and dynamic. Traditional supervised deep learning approaches (e.g., YOLO) require extensive, pre-labeled datasets and careful retraining to adapt to changing environmental conditions, which is often impractical due to the scarcity of expert-annotated examples and the high cost of subject-matter expert (SME) review. Furthermore, fully supervised methods may fail to generalize when background RF conditions shift. The core challenge is to detect and classify sparse satellite signatures within vast amounts of unlabeled RF data without relying on massive pre-labeled training sets.

Methodology
The authors propose a semi-supervised workflow that bridges unsupervised pattern discovery with expert interpretation and supervised classification. The pipeline consists of three primary stages:

Data Preprocessing and Feature Extraction:
- Source: In-phase and quadrature (I/Q) measurements from the ALTAIR radar (Kwajalein Atoll) collected during the SINTRA program.
- Processing: For 0.05-second pulse intervals, received signals are convolved with transmitted pulses and transformed into the frequency domain via Fourier transform. Spectra are summed over 80 pulses (4-second integration) to resolve the "ion line" and satellite signatures.
- Representation: RF observations are converted into a non-negative feature matrix representing time-series range–Doppler power spectra.
Unsupervised Clustering via NMFk:
- Algorithm: The authors employ Non-negative Matrix Factorization with automatic model determination (NMFk).
- Mechanism: NMFk decomposes the observation matrix $X$ into basis patterns $W$ and coefficients $H$ ( $X \approx WH$ ). Unlike standard NMF, NMFk automatically estimates the optimal number of clusters ( $k$ ) by perturbing the input matrix, performing repeated NMF factorizations, and evaluating cluster stability using Silhouette statistics and reconstruction error (L-statistics).
- Outcome: The process yields a robust set of latent factors representing recurring physical patterns in the unlabeled data without prior labels.
SME Interpretation and Supervised Classification:
- Labeling: Subject-matter experts review the NMFk clusters, analyzing spectral features (range, Doppler, intensity) to assign physical semantic labels (e.g., "Starlink satellite," "non-Starlink satellite," "background ionospheric noise").
- Training: These SME-assigned labels serve as ground truth to train an XGBoost classifier.
- Inference: The trained classifier maps new RF feature vectors to the SME-defined event categories, enabling automated detection of future observations.

Key Contributions
The paper outlines four specific contributions:

The application of a semi-supervised NMFk-based classification framework specifically for RF data in the context of SDA.
A demonstration of how NMFk can estimate latent, recurring physical patterns in unlabeled RF observations.
The training of an XGBoost classifier on SME-interpreted NMFk cluster labels to assign new samples to discovered event categories.
Validation of the NMFk detections and classifier predictions against Two-Line Element (TLE) derived ground truth to assess correspondence with known tracked space objects.

Results

Clustering: NMFk identified an optimal set of 25 clusters from the ALTAIR dataset. SMEs interpreted these clusters, distinguishing between background ionospheric activity, non-Starlink satellites, and Starlink-like signatures.
Classification Performance: An XGBoost classifier trained on the SME-labeled clusters (using an 80/20 chronological train/test split) achieved a micro-averaged precision of 0.875, recall of 0.870, and F1-score of 0.872 on the test set. The macro-averaged F1-score was 0.593, which the authors attribute to the scarcity of samples in minority clusters due to the chronological split.
Detection Validation:
- In the training window, the NMFk+SME workflow detected 44 of 57 known Starlink satellites.
- In the held-out test window, the classifier successfully reproduced the NMFk labeling behavior and detected 9 of the 15 expected Starlink satellites.
- The classifier showed perfect agreement with the SME-labeled NMFk clusters in the test data, indicating successful learning of the latent patterns.
Limitations: The authors note that some clusters were too sparse for reliable learning in the test split. Additionally, some detections did not align with TLE data, suggesting either missed TLE entries or detections of non-Starlink objects with similar signatures.

Significance and Claims
The paper positions this workflow as a practical solution for RF event classification where labeled data is scarce and interpretability is paramount. By leveraging unsupervised factorization to discover structure and SMEs to assign meaning, the method reduces dependence on large pre-labeled datasets. The authors claim this approach allows for the automated detection and characterization of space objects while remaining robust to changing RF conditions.

The work is presented as a "proof-of-concept" that demonstrates a promising methodology. The authors modestly acknowledge that accuracy can be improved, citing ongoing work to address false negatives through hierarchical detection (separating background from signals) and multi-channel approaches utilizing different frequencies and polarizations. The primary significance lies in establishing a semi-supervised pipeline that makes RF monitoring more scalable and interpretable for Space Domain Awareness.