That’s expected. Because you remove the aberrant values, the estimator only use “the best” measurements to fit the trajectory.
That is also expected. The goal of the least squares algorithm is to estimate the orbit that best fits the entire measurement arc. In contrast, orbit propagation does not optimize the trajectory from the measurement information. So the error increases with time.
Based on a Gaussian assumption, the diagonal elements of the covariance represent the square of the uncertainty of the estimated parameters. If you take the square root of these elements, you can have a representation of the uncertainty. For instance, if you estimate cartesian orbital elements, the square root of the element (0,0) give you a representation of the uncertainty of the X parameters, elements (1,1) of Y, etc.
The estimated covariance is directly related to the accuracy of your measurements (i.e, the sigma parameter you provide when initializing measurements in Orekit). If you have very accurate measurements, your covariance must be small. At the opposite, if you have inaccurate measurements, you will have a lot of uncertainties on the estimated parameters and then big covariance elements.
For mission analysis purposes, it is interesting is to see the evolution of this uncertainty. For instance, this allows to know, from an uncertainty threshold set by the mission, when the uncertainty on the orbital data becomes too big and therefore when it will be necessary to acquire new measurements with a station to improve the knowledge of the orbit.
You can find interesting topics about covariance propagation on the Orekit forum:
- topic 1
- topic 2 (on this one I recommend you to have a look on the document shared by Romain on the last message)
- topic 3
Bryan