It may not be visible to the naked eye, but a silent crisis is unfolding in Earth’s orbit. There are currently more than 11,000 operational satellites in orbit, and that number is expected to reach 30,000 to 60,000 by 2030. Our orbital infrastructure faces 40,500 tracked objects larger than 10 cm, 1.1 million pieces of space debris between 1 cm and 10 cm, and 130 million pieces of space debris between 1 mm and 1 cm. An unprecedented challenge. Traditional space surveillance systems, designed for an era of simpler space operations, are struggling to keep up with the rapid increase in orbital activity and space debris accumulation.
And the danger is great. Even a single collision between two satellites, or between a satellite and debris larger than a centimeter, can trigger a chain of further debris, potentially rendering the entire orbit unusable for decades, as predicted by Kessler syndrome. The threat goes beyond mere risk of collision. From advanced signal interference to potential hostile acts, the security challenges facing space assets are evolving dramatically.
As commercial launches accelerate, microsatellites become cheaper, and megasatellite constellations become a reality, the mathematical complexity of tracking and securing space assets is surpassing human analytical capabilities. Managing on-orbit hazards requires building better space situational awareness (SSA), anomaly detection, and predictability in orbit. The inherent complexity and the need to respond quickly to threats and problems make AI assistance among other things. In my view, the world needs planetary neural networks (PNNs), systems that can manage these challenges for carriers around the world.
AI in the Space Security Framework
Optical and radar ground systems detect objects in orbit by transmitting electromagnetic waves and analyzing the echoes reflected from those objects. Traditional radar signal processing methods have proven to be robust, but reach physical and algorithmic limits when dealing with weak radar cross sections, cluttered signals, or transient detection. As a result, small debris often goes undetected, resulting in incomplete trajectory catalogs and increased risk of collisions.
Recent advances in machine learning and deep learning have opened new possibilities for processing complex radar and optical data. By introducing an AI-based layer into the signal processing chain, the system can enhance detection of weak signals and filter out noise and atmospheric interference. AI can identify and classify objects in orbit, recognize patterns that correspond to shape, spin state, and orbital state, and predict trajectories and likelihood of collision with greater accuracy, even when observational data is incomplete or uncertain.
This integration marks a paradigm shift from purely physics-based models to data-driven intelligence that can not only adapt to dynamic observation conditions in real time, but also detect, confirm, and classify smaller debris.
For example, a convolutional neural network (CNN) trained on thousands of radar echoes can recognize the unique spatial features of a small piece of metal even when the signal is partially masked by noise. This greatly increases the sensitivity and reliability of detection networks such as those operated by emerging SSA providers.
Although CNNs are good at spatial analysis, they do not consider temporal evolution, which is an important aspect of object tracking and trajectory prediction. To address this limitation, researchers combine CNNs with Long Short-Term Memory (LSTM) networks. These recurrent neural networks can learn long-term dependencies on continuous data, making them ideal for analyzing how an object’s radar or optical signature changes over time.
AI capabilities enable continuous tracking even when data is intermittent due to sensor gaps or poor viewing conditions. The system can also disambiguate overlapping trajectories and maintain object identity across multiple sensor networks.
Planetary Neural Network (PNN)
To change the way orbital debris is tracked, cataloged, and managed, I propose a “central nervous system” for orbital awareness. The system integrates multiple data streams, from satellite telemetry, ground-based sensors, and electromagnetic spectrum analysis to social media reporting, typically well beyond the realm of space engineering. All these contribute to a dynamic real-time image of the space environment and SSA.
PNN sounds elegant. However, making it a reality is by no means easy. The path to full-scale adoption remains fraught with difficult technical and operational hurdles.
One of the most persistent challenges is the lack of data interoperability. Satellite carriers, both public and private, use different formats, sampling rates, and labeling conventions. For accurate orbit tracking, global synchronization must reach millisecond accuracy. This is difficult enough on Earth, but it’s even more difficult between orbital nodes.
For AI, it’s like trying to assemble a puzzle, where each piece was cut by a different manufacturer. The result is fragmented data pipelines, slow training, and difficulty deploying models at scale.
For a PNN to be viable, interoperability must be achieved at three levels:
- Standardization of data formats: All observations must be translated into a shared machine-readable schema such as the CCSDS (Consultative Committee on Space Data Systems) standard.
- Unification of coordinates and time: Interoperability is not possible without common spatio-temporal references.
- Harmonizing semantics and metadata: Different sensors use different terminology and units of measurement, even if they are in the same format.
Another technical barrier to using AI is the generation of false positives. False positives occur when a system detects an object that is not there, a random variation, an interference pattern, or a signal that is misinterpreted. Detection of such phantoms is surprisingly common in radar and optical tracking. They can originate from weak radar echoes, continuous wave interference, atmospheric reflections, or simply thermal or electronic noise that mimics the AI’s hypersensitivity as the neural network “sees” patterns in random noise.
Multiply hundreds of sensors around the world, edge case events (rare and extreme scenarios that are most likely to confuse models), and oversensitivity, a side effect of trying to detect the undetectable, and the disruption is not just physical, but digital.
This is where PNN can make a difference. Connecting radar, optical and infrared sensors around the world provides a built-in mechanism for cross-verification. If one radar station reports a new object but no other optical or orbital nodes confirm it, the network can flag the detection as suspicious. Conversely, if multiple sensors independently detect the same signature, confidence increases exponentially. This multi-sensor consensus is one of the most powerful weapons against false positives. PNNs allow the entire planet to see, compare, and agree, rather than relying on one set of “eyes.”
Another line of defense lies in the system’s temporal intelligence. As pointed out above, LSTM-based models analyze not only individual frames but also changes in the signal over time. Real orbital objects follow predictable physics, and their motion across the sky is continuous and consistent. In contrast, false positives appear suddenly and disappear just as quickly. By tracking temporal consistency, LSTMs can learn to reject temporary anomalies by effectively asking, “Is this object behaving like something in orbit, or like a glitch?” This temporal reasoning transforms the raw detection results into stable tracks and removes the momentary noise inherent in single-sensor systems.
To further refine the decision, each detection in the PNN can be assigned a confidence score (a number representing the likelihood that it represents a real object). This score integrates multiple factors, including signal-to-noise ratio, multi-sensor correlation, orbital stability, and agreement between different AI models.
When multiple models, i.e. CNNs, autoencoders, and transformers, evaluate the same data, their results can be combined through ensemble learning. If all models match, the detection is likely genuine. If only one occurs, it’s probably noise. This type of AI committee dramatically reduces false positives and replaces overconfidence in a single model with collective judgment. Despite automation, human expertise remains important. Operators can review ambiguous detections through a visualization dashboard that shows signal strength, trajectory geometry, and sensor correlation.
Each identified false positive helps retrain the model, making it more resilient in the future, thus creating a feedback loop between machine learning and human inference. This is a practical middle ground and is quickly proving its worth in the rapidly evolving space domain.
The future of AI and space security
The trajectory of AI in space security is only pointing in one direction: toward increasingly sophisticated systems capable of both detecting and actively defending against space-based threats.
As machine learning models become more sophisticated and the capabilities of edge computing and TPUs expand, we are moving closer to a future where AI can autonomously predict potential collisions, predict, detect, and classify signal interference in real time, and automatically implement countermeasures against identified threats.
There will be many more applications in the future, from coordinating multiple satellites for distributed threat detection to quantum-resistant security protocols and advanced predictive maintenance capabilities. All this extends the operational life of the satellite and maintains its safety and security integrity.
In other words, AI in space applications is an absolute game-changer for the future of safety and security in space, whether ground-based or space-based, and we can’t wait to see what’s ahead.
Hans Martin Steiner is Terma’s Vice President and Head of Facilities Space.
