Improving Space Object Tracking By Calibrating Ground-Station Data Using AI

tldr; I developed a way to improve space object tracking by creating a regression model that takes data from 3 different ground-stations as input to predict more accurate data.

Since 2010, I’ve watched with growing concern as the number of objects in Low Earth Orbit (LEO) has doubled. This surge has dramatically increased the likelihood of collisions, which pose serious threats to our functional satellites and generate even more space debris.

Traveling at tens of kilometers per second, these fragments can trigger a cascade of collisions, potentially filling Earth’s orbit with hazardous debris. This could make LEO impassable, halting space exploration and research as we know it.

The affects of a 14g space debris hitting aluminum at 24, 000 km/h.

Current space surveillance methods, which rely on radar sites, provide only probabilistic assessments of collisions. This uncertainty endangers the safety and operational integrity of important space infrastructure.

Low-Earth Orbit (LEO) Is Becoming Congested

According to NASA, the LEO debris population has reached an unstable tipping point where collisions will become the dominant debris-generating mechanism in the future.

Projections indicate that without active debris removal, the LEO debris population (objects >10 cm) will experience a rapid non-linear increase over the next 200 years, even with strict mitigation measures like the 25-year disposal rule.

Even with 95% compliance to the 25-year rule, the LEO debris population is expected to increase by more than 50% in 200 years, leading to an average of 26 catastrophic collisions during this period according to monte-carlo simulations.

The collision rate between cataloged objects in LEO is predicted to reach approximately 0.14 debris-producing collisions per year, with a 63% probability of at least one additional collision by 2012 based on the rate in 2009.

The Problem

The number of objects in LEO is rapidly increasing, leading to a higher risk of collisions that is triggering a chain reaction known as the Kessler Syndrome.

Kessler Syndrome

Kessler Syndrome is a theoretical scenario where the density of objects in LEO is high enough that collisions between them could cause a cascading chain reaction, rendering space activities and the use of satellites in LEO difficult for many years.

• There are currently around 30,000 objects larger than 10cm and millions of smaller pieces of debris orbiting Earth.
• The relative velocities of objects in LEO can reach up to 15km/s (54,000 km/h).
• A collision between two intact objects at such velocities would generate a significant amount of new debris.
• NASA estimates that there are over 500,000 pieces of debris between 1–10cm in size.
• The International Space Station has to perform avoidance maneuvers about once per year to avoid potential collisions.
• In 2009, the collision between the Iridium 33 and Kosmos-2251 satellites added over 14,000 pieces of trackable debris.
• It is estimated that destructive collisions below 1,000 km altitude occur approximately every 3.9 years over 1,000 years.
• The density of debris is highest in LEO between 600–800 km altitude.
• Kessler predicted in 1978 that by 2000, the debris environment would be self-sustaining due to collisions.

These high velocities involved make even small debris catastrophic to spacecraft, which is why actively mitigating and removing debris is crucial to prevent Kessler Syndrome from becoming a reality.

Current Research Is Focusing On Various Things

Currently there is a lot of on-going research to mitigate the growing concern of collisions. There are many aspects of Kessler’s Syndrome that are being targeted.

Active Debris Removal (ADR):

• The European Space Agency’s ClearSpace-1 Mission, set for 2025, aims to capture and deorbit a piece of space debris using a robotic arm, marking one of the first dedicated debris removal missions.
• Japan’s JAXA and Astroscale’s ELSA-d mission, launched in 2021, demonstrated the capability to capture and remove defunct satellites using magnetic docking.

Improved Satellite Design and End-of-Life Disposal:

• Satellites and rocket stages are now passivated to prevent explosions, a practice recommended by space agencies.
• Deorbiting mechanisms are increasingly standard, with SpaceX’s Starlink satellites designed to deorbit within five years after their missions end.

Collision Avoidance Systems:

• SpaceX’s Starlink satellites use onboard AI to autonomously avoid collisions, an essential feature given the large number of Starlink satellites in orbit.
• The U.S. Space Surveillance Network (SSN) tracks over 30,000 objects larger than 10 cm in orbit, providing vital data for satellite operators to perform collision avoidance maneuvers.

International Guidelines and Cooperation:

• The Inter-Agency Space Debris Coordination Committee (IADC) has established guidelines for debris mitigation, including the recommendation that 95% of satellites in large constellations be deorbited at mission end.
• The United Nations Office for Outer Space Affairs (UNOOSA) promotes international cooperation in space debris mitigation and the adoption of global best practices.

Although fundamentally, everything relies on tracking the spacecrafts.

Improving Space Object Tracking Itself Can Be Big

Currently, the European Space Agency’s Space Debris Office maintains a catalog of over 34,000 objects larger than 10 cm, with tracking accuracy within a few hundred meters for objects in low-Earth orbit.

Enhancing this accuracy can have profound impacts on space safety and sustainability. With higher tracking accuracy, much less uncertainty will be accommodated in collision prediction, so satellite operators can precisely carry out avoidance maneuvers in good time. Better data can help reduce false alarms, and needless maneuvers-which prolong satellite lifetimes.

It will also support efficient active debris removal through higher accuracy in space object tracks. This would increase the targeting and capture of debris, helping greatly in missions like ESA’s ClearSpace-1 and Japan’s ELSA-d.

Status Quo

Ground stations use the principle of radiolocation to determine the position and trajectory of spacecraft and other objects in space. Here’s an in-depth explanation of how it works:

Radiolocation Principle

Radiolocation involves transmitting radio waves from a ground station towards a target object in space and analyzing the reflected or retransmitted signals to extract information about the object’s position, velocity, and other characteristics. This process relies on the principles of electromagnetic wave propagation and the Doppler effect.

Transmission of Radio Waves:

• Ground stations are equipped with powerful transmitters and directional antennas capable of emitting radio waves at specific frequencies.
• The radio waves are transmitted in pulses or continuous waves, depending on the radiolocation technique used.

Reflection or Retransmission:

• When the radio waves encounter a spacecraft or other object in space, they interact with the object’s surface or internal components.
• Depending on the object’s characteristics and the radiolocation technique, the radio waves can be reflected (radar) or retransmitted (transponder-based).

Signal Reception and Analysis:

• The ground station’s receivers and antennas detect the reflected or retransmitted signals.
• By analyzing the characteristics of the received signals, such as the time delay, frequency shift (Doppler effect), and signal strength, various parameters can be calculated.

Position and Trajectory Determination:

• Time Delay: The time it takes for the signal to travel to the object and back provides information about the object’s range or distance from the ground station.
• Doppler Shift: The change in frequency between the transmitted and received signals is caused by the relative motion between the object and the ground station, allowing the calculation of the object’s radial velocity.
• Angle Measurements: By using multiple ground stations or directional antennas, the angle of arrival of the reflected or retransmitted signals can be measured, providing information about the object’s azimuth and elevation.

Orbit Determination:

• By combining range, velocity, and angle measurements from multiple ground stations over time, the object’s orbit or trajectory can be precisely determined using techniques like trilateration and orbital mechanics calculations.

Although, current space object tracking only provides us with approximations on where space objects are, leaving satellite operators with a marginally practical ‘probability of collision’.

A Singular Ground-Station Is Not Reliable

Reliance on a singular ground station for space object tracking is simply unreliable, particularly with the growing population of objects in LEO.

There are many factors that can manipulate the tracking data.

• Angular Coverage & Power: The number of trackable objects is directly linked to the survey zone’s size and transmitted signal power.
• Atmospheric & Light-time Effects: These effects introduce errors ranging from -13.982m to -78.787m for tropospheric correction and -0.304m to 0.103m for light-time correction.
• Time Tag Bias: Errors in signal timing can lead to position errors ranging from -7.671m to 0.491m, for instance, with a 1ms bias.
• Measurement Bias & Noise: Systematic errors like bias and noise significantly affect orbit determination. Even small biases can have a large impact.

Although, that ultimately made me wonder… why don’t we currently use multiple ground-stations to track space objects already?

Well, there are multiple technical challenges.

Ground stations use a mix of radars, optical telescopes, and radio frequency systems, requiring complex data fusion techniques to account for discrepancies in accuracy, resolution, and coverage.

Additionally, delays can also arise from coordinating data collection and processing across geographically dispersed stations. Real-time integration can be difficult due to the high speeds of space objects.

Ground stations tend to also just dynamically adjusts their systems to reduce these errors, making it intricate when trying to make multiple systems work together.

Problem Framing

What methodologies can be utilized to leverage data acquired from a network of ground stations for the refinement of tracking data pertaining to individual space objects, thereby enhancing their positional and orbital accuracy?

I Simulated Satellites To Create A Dataset

To test anything, I need a dataset — for which I simulated the STARLINK-31085 satellite since it travels over North America, and has a well researched orbit.

STARLINK-31085 Satellite in STK

For this project I used STK by AGI (Analytical Graphics, Inc.), a software suite used to model and simulate complex scenarios involving spacecrafts.

I used three geographically separated ground stations across America for the simulation.

• Castle Rock
  (39.2768° N, -104.807° W)
• Tranquillon Peak
  (34.583° N, -120.561° W)
• White Sands
  (33.8131° N, -106.659° W)

The 3 ground-stations across America.

I used the Two-Line Element Set (TLE) data to understand and simulate the orbital trajectory of the Starlink satellite. From which I collected AER (Azimuth, Elevation, and Range) data from each ground station individually, focusing on overlapping regions with simultaneous observations.

STARLINK-31085's Orbit

Lastly, the data from these 3 ground-stations was merged into a single .csv file, creating a tabular dataset.

The AER data from ground-station 1, and 2, visualized on a scatter plot.

I Created A Regression Model With The Dataset

What is Regression?

Regression is a statistical method through which it becomes possible to understand the relationship between variables. Suppose there is a collection of data points; each point represents two things: an input and an output. The aim is to find a mathematical equation that perfectly describes how the input relates to the production. Once that model is obtained, it can then predict future outputs from new inputs.

For example, say you have data that includes the number of hours students study and their corresponding exam scores. You can use regression to determine how changes in the number of hours they study will predict changes in their exam scores.

Regression in Machine Learning

This forms a type of supervised learning model. Supervised here simply means that we will train a model against a dataset where both the inputs and the corresponding outputs are well-defined. Machine learning in regression is learning a function that describes a strong relationship between a set of input variables or features and the corresponding continuous output variable. This learned function must further have the capability to predict the output of new or unseen inputs.

For instance, you have a machine learning model to predict house prices given the size, number of bedrooms, and location as features. These regression problems learn the historical relation between these inputs (features) and the house prices (outputs).

A basic example of polynomial regression.

What I Did with the ML Regression Model

In this project, I built an ML regression model using a recently proposed model architecture optimized for tabular data, called TabNets (source).

Here is how I trained the model step by step:

I wanted to know how well AER data for the ground stations was faring against the actual AER data a simulation would provide. This essentially meant seeing how close the measurements from the ground stations were to the actual position of the object.

The goal was to make the measurements as close as possible to the black point (actual value). This graph also shows the variance in the data collected from the different ground-stations.

The x-values for the model are the AER measurements coming from the three ground stations, while the y-values are the actual AER data provided in the simulation.

Using TabNets, a special kind of neural network architecture fine-tuned explicitly for solving problems with heavy input noise in tabular data, I built a regression model. This model was trained on inferring the relationship between these measurements and accurate AER data.

TabNets use spare feature selection, which sequentially validates the importance of each feature at each step of the training process.

TabNet employs multiple decision blocks that focus on processing a subset of input features for reasoning. Two decision blocks shown as examples process features that are related to professional occupation and investments, respectively, in order to predict the income level. Source: https://arxiv.org/pdf/1908.07442

I then trained the model on the collected dataset, allowing it to learn how to map the input AER measurements to actual AER. During this training phase, the model’s internal weights are updated such that the difference between its prediction and the actual data of AER is minimal.

Then, post-training, I evaluated the performance of the model about how well the model was capable of predicting actual AER data from new measurements at the ground stations. This makes the evaluation metrics measure how well the model learned the relationship.

The model was evaluated using MSE, RMSE, and MAE. The lower the number, the better.

The downward trend over 40+ epochs indicates that the model was actually learning how to predict more precise AER data.

My Model Improved Space Object Tracking

My model was able to accurately improve the reliability of space object tracking by taking multiple ground-stations into account.

As an example, the following input:

# Input data
example_input = pd.DataFrame({
    'azimuth_1': [50],
    'elevation_1': [3],
    'range_1': [100],
    'azimuth_2': [40],
    'elevation_2': [2],
    'range_2': [150],
    'azimuth_3': [60],
    'elevation_3': [1],
    'range_3': [200]
})

Can give an output like:

Predicted azimuth: 15.968437
Predicted elevation: 5.441539
Predicted range: 168.35468

Conclusion

Satellite operators can benefit financially from enhanced accuracy in tracking by up to millions of dollars saved yearly. For instance, a single collision avoidance maneuver costs anywhere between $100,000 to $500,000 if expenditure in fuel and time lost in operations are accounted for. Since better tracking minimizes their number, this could potentially save the industry ~$200 million annually.

Furthermore, accurate tracking data is essential for maintaining international cooperation and compliance with space debris mitigation guidelines. Organizations like the Inter-Agency Space Debris Coordination Committee (IADC) and the United Nations Office for Outer Space Affairs (UNOOSA) rely on precise tracking information to enforce and promote best practices globally.

To sum up, improving the accuracy of space object tracking is not just a technical upgrade — it is a critical step towards ensuring the long-term sustainability and safety of space operations. Enhanced tracking capabilities will reduce collision risks, extend satellite lifespans, support effective debris removal, and save significant financial resources, ultimately safeguarding the future of space exploration and research.

Hey 👋, I’m Prabhdeep, a 17y/o AI + Space Tech enthusiast. I enjoy building meaningful projects, and aspire to bring an impact. Check out my LinkedIn, Twitter, and YouTube channel. Also feel free to subscribe here for monthly updates on the projects I build!