Probabilistic Artificial Intelligence Framework for Data Association in Distributed Multi-Sensor Space Object Tracking under Sparse Measurement Regimes

The sky above us is becoming increasingly crowded. Beyond the visible stars and the occasional flash of a satellite, a silent traffic jam of human-made objects orbits the Earth. This population, known as resident space objects, includes active satellites, defunct spacecraft, spent rocket bodies, and countless tiny fragments of debris. While the largest of these are tracked by ground-based radar stations, the sheer volume of objects—numbering in the tens of thousands for cataloged items and millions for smaller, untrackable pieces—creates a dangerous environment. The primary challenge for space agencies is not just seeing these objects, but knowing exactly where they are and predicting where they will go next. This is critical for preventing collisions that could destroy valuable satellites or create even more debris. However, the current methods used to track these objects struggle when the data is incomplete. Radar stations often only catch a brief glimpse of a satellite as it passes overhead, providing a short, sparse arc of information rather than a continuous stream. In these difficult conditions, traditional mathematical tools often fail to connect the dots, leaving gaps in our knowledge of the orbital landscape.

To address this growing complexity, researchers Mehar Taj and Dr. Zabiha Khan from Nitte University in India have developed a new, unified system that uses artificial intelligence to solve three difficult problems at once: identifying which radar blip belongs to which object, deciding which radar should look where next, and calculating the object's path with high precision. Their work moves away from the old way of handling these tasks separately, where one computer program guesses the identity of an object and another decides where to point the radar. Instead, they created a single, integrated framework that learns to do all three simultaneously. This approach is specifically designed for the "sparse measurement" regime, where radar stations only get a few seconds of data on an object before it disappears from view. By combining different types of artificial intelligence, the team built a system that mimics the way an expert human analyst might reason through uncertainty, but with the speed and consistency of a machine.

The core of their solution lies in how the system connects new radar data to existing records of space objects. In a crowded sky, a radar might detect a hundred blips in a single second, but it does not immediately know which blip corresponds to which known satellite. Traditional methods try to solve this by testing every possible combination, a process that becomes impossibly slow as the number of objects grows. The new system uses a graph-based neural network, a type of artificial intelligence that excels at understanding relationships between many different items. It treats the radar measurements and the known satellite tracks as points on a map, drawing lines between them based on how likely they are to match. Crucially, this network is built with an understanding of orbital physics, meaning it knows that objects move in specific, predictable ways governed by gravity. This allows the system to instantly rule out impossible matches without needing to calculate every single possibility. In their tests, this method proved to be vastly faster and more accurate than older techniques, especially when the radar data was very brief.

Once the system knows which blips belong to which objects, it must decide how to use the limited time of the radar stations. Radars cannot watch everything at once; they must choose which objects to track to get the most useful information. The researchers trained a reinforcement learning agent, a type of AI that learns by trial and error, to act as a smart scheduler. This agent looks at the entire catalog of space objects and their current levels of uncertainty. If a satellite's position is poorly known, the agent prioritizes it, directing the radar to gather more data. If a satellite is well-tracked, the agent moves on to others. This dynamic approach ensures that the radar network focuses its energy where it is needed most, rather than following a rigid, pre-set schedule. The results showed that this intelligent scheduling reduced the overall uncertainty of the entire space catalog by up to 75 percent compared to standard, rule-based scheduling methods.

The final piece of the puzzle is determining the exact path of the satellite based on these sparse, short bursts of data. Standard mathematical models often assume that errors in measurement follow a simple, bell-shaped curve, which works well when there is plenty of data. However, when data is scarce, the true uncertainty can be complex and irregular. The researchers replaced these standard models with a neural ordinary differential equation system. This is a machine learning model that is constrained by the laws of physics, ensuring that the predicted paths always make sense in the real world while still learning from the messy, real-world data. This hybrid approach allowed the system to calculate the position and speed of objects with significantly higher accuracy than traditional methods. In simulations involving a fragmentation event, where a satellite breaks into hundreds of pieces, the new system reduced the error in position estimates by 78 percent compared to the best existing standard method.

The researchers tested their framework in five different scenarios, ranging from a quiet sky with few objects to a chaotic environment with hundreds of satellites and debris fragments. In every case, the integrated system outperformed the traditional methods that handle data association, scheduling, and orbit calculation as separate steps. The most dramatic improvements appeared in the most difficult situations, where the number of objects was high and the radar data was extremely brief. In one extreme test involving a fragmentation event with 800 objects, the traditional methods took so long to process the data that they became unusable for real-time operations, while the new system completed the task in a fraction of a second. The study confirms that treating these problems as a single, connected challenge yields better results than solving them one by one. While the work was conducted in a high-fidelity simulation and has not yet been tested on live radar data from the Indian Space Research Organisation, the results suggest a powerful new direction for keeping the skies safe. By unifying the tasks of seeing, deciding, and predicting, this framework offers a way to manage the growing complexity of space traffic with greater speed and reliability.