Michael Lindsey,

and Franziska Weber

In recent years, the accurate prediction of satellite orbits has attracted much attention due to efforts in avoiding collisions between orbital objects. There are a number of challenging problems in orbital propagation that hinder our ability to provide an accurate forecast. Among them is the correct specification of the ionosphere-thermosphere environment, accurate drag coefficients that influence the orbital object, and precise estimation of orbital position and velocity.

The first part of the presentation will be a brief overview of the Integrated Modeling of Perturbations in Atmospheres for Conjunction Tracking (IMPACT) project. The project objective is to develop a ground- breaking new orbital dynamics framework that combines a comprehensive physics-based model of atmo- spheric drag with an accurate uncertainty quantification of orbital predictions. To achieve this goal the IMPACT team has developed and implemented a predictive physics-based ionosphere-thermosphere system, improved drag coefficients model, a novel density estimation technique based on tomography, and an orbital collision probability estimation technique based on importance sampling and Gaussian processes.

The later half of the presentation will focus in improving the estimation and prediction of orbital tra- jectories of space objects. Orbit estimation presents a particularly challenging problem since the resulting models are non-linear and have non-Gaussian distributions. A useful class of methods for improving model predictions are data assimilation techniques, which fuse model and observational data information to provide an enhanced model solution with better predictive properties. Unfortunately, traditional data assimilation techniques fail for orbital propagation given that most assume a Gaussian distribution for the model and observations. We present a series of assimilation experiments with a two-dimensional orbital propagation model to study the efficiency and applicability of three types of data assimilation methods to this problem. In particular, we consider the ensemble Kalman filter (EnKF), Monte Carlo (MC) sampling, and variational data assimilation (4D-Var). A series of assimilation experiments will be presented where a number of con- ditions are tested, including the frequency of assimilation, and number of particles/ensemble members. The experiments show how EnKF suffers from filter divergence in most cases, while both 4D-Var and MC sam- pling performance is stable and reliable. This is due mainly to the non-linearity and non-Gaussian nature of the problem.