Andrii Bidochko

• Updated: July 4, 2026
• 7 min read

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

Direct Answer

The paper introduces a semi‑supervised workflow that combines Non‑negative Matrix Factorization with automatic model selection (NMFk), expert‑driven cluster labeling, and a downstream classifier to detect and classify satellites in raw radio‑frequency (RF) monitoring data. This matters because it dramatically reduces the need for large, hand‑labeled RF datasets while still delivering reliable satellite detection for space‑domain awareness operations.

Background: Why This Problem Is Hard

Space‑domain awareness (SDA) relies on continuous RF monitoring to track active satellites, catalog debris, and assess ionospheric conditions. In practice, RF sensors generate massive streams of spectrograms that are:

• Highly variable across frequency bands, time of day, and atmospheric conditions.
• Sparsely populated with clear satellite signatures—most samples are background noise or ionospheric events.
• Expensive to label: expert analysts must manually annotate each detection, a process that can take hours per day of observation.

Supervised deep‑learning models have shown impressive performance when trained on abundant, accurately labeled data. However, they suffer from two critical drawbacks in the SDA context:

1. Label scarcity: Acquiring thousands of annotated RF clips for every new sensor or frequency band is infeasible.
2. Domain drift: RF environments evolve with solar activity, hardware upgrades, and regulatory changes, forcing frequent model retraining.

These constraints create a bottleneck that limits the scalability of purely supervised approaches, prompting researchers to explore semi‑supervised techniques that can leverage the wealth of unlabeled observations.

What the Researchers Propose

The authors present a three‑stage semi‑supervised pipeline designed specifically for satellite detection in RF data:

1. Unsupervised factorization: Convert raw spectrograms into a non‑negative feature matrix and apply NMFk, an extension of Non‑negative Matrix Factorization that automatically determines the optimal number of latent components (clusters).
2. Expert‑guided interpretation: Domain specialists examine the resulting clusters, assign physical meanings (e.g., “satellite pass,” “ionospheric scintillation,” “background noise”), and create a labeled reference set without manually annotating individual samples.
3. Classifier training: A lightweight supervised model (e.g., a random forest or shallow neural net) is trained on the expert‑labeled clusters and then used to predict the class of new, unseen RF observations.

This design keeps the heavy lifting—discovering latent structure—in the unsupervised stage, while the human expert only needs to label at the cluster level, dramatically cutting annotation effort.

How It Works in Practice

The workflow can be visualized as a linear pipeline with clear hand‑offs between automated and human components:

1. Data ingestion: Continuous RF streams are segmented into short time‑frequency windows and transformed into magnitude spectrograms.
2. Feature matrix construction: Each spectrogram is flattened into a non‑negative vector; stacking these vectors yields a matrix X where rows correspond to time‑frequency samples and columns to spectral bins.
3. NMFk factorization: The algorithm decomposes X ≈ W·H, where W captures basis spectra (latent patterns) and H encodes their activation over time. NMFk iteratively tests different numbers of components and selects the model that best balances reconstruction error and sparsity.
4. Cluster extraction: Each column of H defines a cluster of samples that share a common spectral signature. The system automatically groups samples by their dominant component.
5. Expert labeling: Analysts review representative spectrograms from each cluster, assign a semantic tag (e.g., “LEO satellite,” “GEO satellite,” “ionospheric burst”), and optionally merge or split clusters based on domain knowledge.
6. Supervised classifier training: The labeled clusters become a training set for a downstream classifier that learns to map raw spectrograms directly to the expert‑defined categories.
7. Operational deployment: New RF data flow through the trained classifier, producing real‑time alerts for satellite detections and contextual tags for downstream SDA systems.

What sets this approach apart is the automatic determination of the latent dimensionality (the “k” in NMFk). Traditional NMF requires a pre‑specified number of components, often leading to under‑ or over‑fitting. NMFk’s data‑driven model selection ensures that the factorization captures the true complexity of the RF environment without manual tuning.

Evaluation & Results

The authors validated the pipeline on a multi‑month RF dataset collected from a ground‑based monitoring station covering 1–2 GHz. Evaluation focused on three questions:

• Cluster fidelity: Do the NMFk‑derived clusters correspond to physically meaningful phenomena?
• Classification accuracy: How well does the downstream classifier predict expert‑assigned labels on a held‑out test set?
• Comparison to fully supervised baselines: Does the semi‑supervised method approach the performance of a model trained on thousands of manually labeled samples?

Key findings include:

• The NMFk factorization consistently identified 7–9 dominant components, each aligning with distinct RF signatures such as LEO passes, GEO passes, ionospheric scintillation, and background thermal noise.
• Expert review confirmed that >92 % of the clusters could be assigned a clear physical interpretation, demonstrating that unsupervised factorization preserved domain‑relevant structure.
• The classifier trained on cluster‑level labels achieved an overall F1‑score of 0.87 on the test set, only 4 % lower than a fully supervised deep‑CNN trained on 5,000 hand‑labeled spectrograms (F1 = 0.91).
• When the dataset was perturbed to simulate solar‑storm conditions, the semi‑supervised model degraded gracefully (F1 ≈ 0.82) whereas the fully supervised baseline dropped below 0.70, highlighting robustness to domain shift.

These results demonstrate that the semi‑supervised pipeline can deliver near‑state‑of‑the‑art detection performance while requiring an order of magnitude fewer human annotations.

Why This Matters for AI Systems and Agents

From an AI‑systems perspective, the paper offers a template for building agents that can learn from scarce, noisy data streams without exhaustive labeling:

• Reduced annotation cost: By shifting the labeling burden to the cluster level, organizations can deploy monitoring agents faster and at lower operational expense.
• Adaptability: The NMFk component can be re‑run periodically, allowing agents to self‑recalibrate as the RF environment evolves, a capability essential for long‑term autonomous SDA platforms.
• Interpretability: Each latent component corresponds to a physically meaningful pattern, enabling agents to explain why a particular detection was raised—a critical requirement for trust in high‑stakes aerospace applications.
• Integration potential: The lightweight classifier can be embedded in edge devices or orchestrated via a cloud‑based workflow engine, feeding alerts directly into downstream decision‑making pipelines.

Practically, teams building UBOS platform overview can leverage the semi‑supervised workflow as a plug‑in module within their Workflow automation studio, enabling rapid prototyping of RF‑based monitoring agents without writing custom deep‑learning pipelines.

What Comes Next

While the study marks a significant step forward, several avenues remain open for exploration:

• Scalability to multi‑sensor networks: Extending NMFk to jointly factorize data from geographically distributed receivers could uncover cross‑site patterns and improve detection confidence.
• Hybrid semi‑supervised schemes: Combining NMFk with contrastive self‑supervised pre‑training may further reduce the need for expert intervention.
• Real‑time deployment: Optimizing the factorization step for GPU or FPGA acceleration would enable sub‑second detection pipelines suitable for on‑board satellite monitoring.
• Broader domain transfer: The same methodology could be applied to other sparse signal domains such as acoustic monitoring of marine vessels or electromagnetic surveillance of UAVs.

Addressing these challenges will require close collaboration between signal‑processing researchers, domain experts, and AI platform builders. Organizations interested in experimenting with the workflow can start by exploring the Chroma DB integration for storing and querying latent component embeddings, or the ChatGPT and Telegram integration for real‑time alert delivery to operational teams.

References

Detecting Satellites in Radio‑Frequency Data via Semi‑Supervised Learning – arXiv preprint, June 2026.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.