Harry She, LeoLabs; Chandler Phelps, LeoLabs; Owen Marshall, LeoLabs; Erin Dale, LeoLabs
Keywords: Machine learning, object typing, classification, clustering, LEO, ground-based radar network, XGBoost, characterization, Space Domain Awareness, Space Situational Awareness, LeoLabs, Radar, S-Band, UHF, Maneuverability, Artificial intelligence
Abstract:
Determining the physical characteristics and maneuverability capabilities of a satellite, as well as its orbital similarity to other spacecraft, is essential for enhancing tracking accuracy, collision avoidance, and overall space situational awareness. Existing methods for identifying an object as a Maneuverable or Non-Maneuverable Payload, Rocket Body or Debris Fragment are labor-intensive, time-consuming, and were not designed for launches with large manifests. As of February 2024, there are several hundred unknown objects in the United States public catalog. With the number of active satellites anticipated to grow from 8,000 to 58,000 by 2030, it is becoming increasingly crucial to employ rapid, accurate, and automated methods of characterization to stay abreast of the forecasted launch cadence.
LeoLabs uses a set of highly performant XGBoost classifiers trained with data collected from its global radar network to determine object type in LEO (low Earth orbit) within 72 hours of association.
Our approach involves processing raw radar data from LeoLabs’ ground-based sites for thousands of objects in the public catalog. We then extract features that accurately represent the physical characteristics and morphology of a satellite, including, but not limited to, radar cross section (RCS), the signal-to-noise ratio (SNR), measurement count, and other pertinent transformations and aggregations. We train an XGBoost model on historical data collected from all passes from LeoLabs’ radar sites for a fixed period, on a per target basis. When fed with these extracted features of a given target, the model produces an object type – Payload, Debris, or Rocket Body.
Objects classified by this model as a Payload are further processed to determine if they are Maneuverable or Non-Maneuverable. This involves a processing pipeline that predicts when an object has deviated from natural decay during its historical timeline, in combination with other derived features that score the magnitude and prevalence of such maneuvers.
Model performance is then evaluated using a hold-out test dataset of 20% of the targets from the public catalog. These models achieved a mean of ~95% accuracy against the reported values in the public catalog, as well as hand-labelled maneuverability status by expert analysts.
Furthermore, we utilize an unsupervised machine learning technique to map objects into a 2-dimensional space such that clusters that appear tend to correspond to physical characteristics and maneuverability status of the object. We have investigated using this technique to cluster newly launched objects with existing spacecraft in low Earth orbit. This algorithm is useful for determining similarity and uncovering structure in relevant high-dimensional extracted features that may be otherwise hard to dissect.
These methods can provide valuable assistance to expert human analysts as they can use such plots to better reason about the spacecraft morphology, behavior, and intent of specific categories of satellites they are knowledgeable about, allowing them to infer the characteristics of new objects by comparing the 2-dimensional location of that target to neighboring targets with known characteristics. The results of investigation have already proved fruitful for myriad practical applications, ranging from identifying if an unknown object is part of an existing family of space systems to inferring the role and mission of a given satellite. These insights enhance the accuracy of a broader, holistic object characterization process.
We propose that focusing development effort on such capabilities are fundamental in establishing a well characterized, predictable LEO space environment for the wider space community – both public and private. By extension, this also motivates data-driven space domain awareness to admonish adversarial action in space through enhanced attribution, as well as enabling progress towards formulating effective regulations that promote space sustainability and improve compliance in LEO.
Date of Conference: September 17-20, 2024
Track: Satellite Characterization