Machine Learning Approach for Accurate and Robust Satellite Tracking in Optical Space-to-Ground Communication - Time-Series Prediction for LEO Satellites

Aug 5, 2026

Upload by Ophelia

Improving orbit prediction is essential for operational reliability as Low Earth Orbit (LEO) becomes increasingly crowded with resident space objects. This work refines orbit forecasting for optical space-to-ground communication and TTC operations by introducing machine learning time-series models to capture systematic deviations from SGP4. Training data combine historical TLE-derived satellite position sequences with GNSS-derived orbital messages and laser ranging measurements. Model performance is assessed by comparing predictions against the conventional SGP4 baseline, with future evaluation planned via RIC error analysis using DLR optical ground station measurements.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Abstract

Introduction

Background: SGP4 and current limitations

Propagation degradation over time

Challenges with EKF and environmental uncertainties

Proposed solution: hybrid ML approach

Learning systematic deviations with time-series networks

Training data sources and model evaluation

Machine Learning Approach for Accurate and Robust Satellite Tracking in Optical Space-to-Ground Communication using Time-Series Prediction for LEO Satellites Maurice Uteg German Aerospace Center (DLR), Responsive Space Cluster Competence Center Eugen-Sänger-Str. 50, 29328 Faßberg, Germany; [maurice.uteg@dlr.de](maurice.uteg@dlr.de) Helmut Ribel German Aerospace Center (DLR), Responsive Space Cluster Competence Center Eugen-Sänger-Str. 50, 29328 Faßberg, Germany; [helmut.ribel@dlr.de](helmut.ribel@dlr.de) Sacha Tholl German Aerospace Center (DLR), Responsive Space Cluster Competence Center Eugen-Sänger-Str. 50, 29328 Faßberg, Germany; [sacha.tholl@dlr.de](sacha.tholl@dlr.de) Marcus T. Knopp German Aerospace Center (DLR), Responsive Space Cluster Competence Center Muenchener Str. 20, 82234 Wessling, Germany; [marcus.knopp@dlr.de](marcus.knopp@dlr.de) Abstract As space becomes increasingly congested with Resident Space Objects (RSOs) in Low Earth Orbit (LEO), improving the accuracy of orbit prediction is crucial for ensuring operational reliability, particularly for satellite tracking in optical communication and Telemetry, Tracking, and Command (TTC) operations. This work focuses on refining orbit prediction by leveraging machine learning techniques to enhance tracking capabilities. Traditional orbit approximation relies on the Simplified Perturbations model (SGP4), which calculates a satellite’s position and velocity by considering various perturbations, such as Earth’s gravitational irregularities and atmospheric drag, using an empirical model for efficient orbit determination. However, this approach is prone to errors, as it simplifies complex orbital dynamics. To address this limitation, this paper explores the potential of machine learning algorithms to analyze time-dependent data, with a particular focus on systematic deviations from SGP4 predictions that are inherently captured in historical orbit information. To achieve this, we create a set of data consisting of time-series satellite position data sets of LEO Objects from past Two-Line Elements (TLEs) as well as orbital messages derived form Global Navigation Satellite System (GNSS) observations and Laser Ranging Data. These data sets are used to train various machine learning models specialized in time series data, such as Long-Short-Term Memory (LSTM) networks to evaluate their potential for improving the robustness and accuracy of orbit forecasting. Finally, the performance of the machine learning model is evaluated by comparing its predictions with those from the traditional SGP4 model. In the future, we will be assessing prediction accuracy and analyze Radial, In-Track, and Cross-Track (RIC) errors to ensure the new model’s effectiveness using measurements from Optical Ground Stations at DLR. Introduction The rapid growth of Resident Space Objects (RSOs) in Low Earth Orbit (LEO) has transformed orbital safety into a critical challenge for space operations. With over 42,000 trackable objects currently cataloged and millions of smaller debris fragments, the risk of catastrophic collisions has increased significantly [1, 2] . Accurate orbit prediction is a key factor towards ensuring safe satellite operations. Even minor errors in predicting a satellite’s trajectory can lead to mission failures due to collisions or the mission ending early due to costly avoidance maneuvers. Following up on this, satellite collisions have occurred due to inaccuracies in prediction data and false alarms [3] . This paper addresses this challenge by proposing a hybrid approach that combines traditional orbital perturbation models with a machine learning (ML) approach to refine orbit prediction accuracy, focusing on LEO being the orbital regime characterized by high object density and collision risk. Current orbit prediction relies on the Simplified General Perturbations Model 4 (SGP4), especially for smaller Satellites like CubeSats. It is a framework that approx

FAQs

Why is accurate orbit prediction critical for optical tracking and TTC operations in LEO?

With increasing numbers of resident space objects and debris, small trajectory prediction errors can cause collisions or trigger costly avoidance maneuvers, undermining mission reliability.

What limitation of SGP4 motivates using machine learning?

SGP4 relies on simplified empirical physics, so its propagation accuracy degrades quickly and it cannot model systematic deviations caused by unmodeled effects like atmospheric density variations and spacecraft-specific characteristics.

How does the proposed work use time-series machine learning models to improve robustness and accuracy?

It trains time-series models such as LSTM networks on historical LEO position data derived from past TLEs, supplemented with GNSS orbital messages and laser ranging data, then evaluates predictions against SGP4.

Aug 5, 2026

Upload by Ophelia