Improved Orbit Prediction using Gradient Boost Regression Tree

Abstract

Orbit prediction is crucial and important for satellite tracking. To improve prediction, a person must be well equipped with knowledge of the earth’s gravitational pull, atmospheric drag, radiation pressures, basic manoeuvring objects, and other information. Thus, orbit prediction has advanced in physics-based models. Most of the time, the above-mentioned information is not publicly available. Data related to satellites is kept with the respective space organizations. Using this concept, the proposed approach employs gradient boost regression trees (GBRT) with two-line element (TLE) data and, when compared to recently developed machine learning techniques such as artificial neural networks (ANN), support vector machines (SVM), and Gaussian processes (GP), it provides improved orbit prediction accuracies in terms of position and velocity. Further, the proposed method avoids the overfitting issue and shows better approximation ability. The simulations are carried out for a total of six resident space objects in low earth orbit, medium earth orbit, and sun synchronous orbit.