Batch LS wouldn't give correct result for OD...(Code and data attached)

[MaximeJ]

Hi @Serrof, @spto,

@Serrof

I’ve tried to use it myself recently but without much success, as it seems to either accept all or reject all measurements

Filtering outliers strongly depends on the configuration of the OD.
If you suspect your first iterations are far away from the truth you can set the warm up to larger iterations (like 2 or 3) so that the filter activate later in the OD process (that way it won’t remove all the measurements).
If all measurements are accepted, you can try diminishing the multiplier maxSigma so that the filter gets more restrictive when it’s activated.

@spto, some thoughts to help you:

• In your example you’ve set the RADEC error (or sigma) to 0.2 radians which is about 11.46 degrees.
  This is too big in my opinion, you should try to put something closer to the real resolution of your optical observations (8 arcsec gives 3.88e-5 rad).
• If your initial orbit is a bit far from the expected solution you can try using a LevenbergMarquardtOptimizer instead of a GaussNewtonOptimizer in your BatchLSEstimator. It has a bigger convergence radius.

Also, one thing you can try is to plot the residuals at iteration 0, that means the residuals of the measurements with respect to your initial orbit and propagation (without any least-square correction).

You could extract them from a BatchLSObserver at iteration 0 of the BatchLSEstimator (see an example of an OrbitDeterminationObserver in the Java tutorials.

Or you could use a specific OrekitStepHandler.
I’ve written these two, they are in Java so you will have to translate them (the first one should be enough):

1. ResidualsHandler.java (3.8 KB)
   Single satellite version. Once it is used, call the method getEvaluations to have a Map of <ObservedMeasurement, EstimatedMeasurement>.
   ObservedMeasurement are your original measurements, EstimatedMeasurement are the theoretical evaluation of the measurements made by Orekit given your intial orbit and propagation model.
2. MultiSatResidualsHandler.java (4.1 KB)
   Multi satellite version. Works the same but with a PropagatorParallelizer instead of a classical Propagator.

Both are mere copies of Orekit’s MeasurementHandler.
The difference is that they don’t refer to any batch least-square model.

Usage of the single satellite one (still in Java).

// Suppose you have a sorted list of measurements and a fully configured propagator builder
List<ObservedMeasurement<?>> measurements;
PropagatorBuilder propagatorBuilder;

// Get start/end date
AbsoluteDate startDate = measurements.get(0).getDate();
AbsoluteDate finalDate = measurements.get(measurements.size() - 1).getDate();

// Build the propagator
Propagator propagator = propagatorBuilder.buildPropagator(propagatorBuilder.getSelectedNormalizedParameters());

// Initialize single sat step handler
ResidualsHandler residualhandler = new ResidualsHandler(measurements);

// Add the handler to the propagator
propagator.getMultiplexer().add(residualhandler);

// Propagate from first obs to last obs
propagator.propagate(startDate, finalDate);

// Get the map of results
Map<ObservedMeasurement<?>, EstimatedMeasurement<?> = residualhandler.getEvaluations();

Then, comparing the values of each EstimatedMeasurement.getEstimatedValue() with its EstimatedMeasurement.getObservedValue() will give you the residuals, that you can plot to better understand what happens.

Hope this helps,
Maxime