Willem Rood, Leiden University; Thomas Wijnen, Leiden University; Remko Stuik, Leiden University
Keywords: Automated, satellite, track, feature, detection, sky-position, extraction, determination, wide view, field of view, high cadence, Hough transform, satellite tracks, initial orbit determination, SDA,
Abstract:
Providing a complete census of satellites and orbital debris begins with the efficient detection of these objects and reliable determination of their orbits (Space Domain Awareness or SDA). The Royal Netherlands Airforce (RNLAF) has expressed the need for developing a system that is able to contribute to SDA. Wide field, high cadence astronomical surveys monitor large fractions of the sky, offering a promising platform for orbit determination purposes. As an example, the MASCARA instrument in Chile continuously monitors the local night sky using five stationary, wide field-of-view cameras with exposures taken at a ~6 seconds cadence. In these images, satellites are readily distinguishable from other objects due to their long, streak-like appearance. But in order to maximise the usefulness of this wealth of data, information regarding satellites should be extracted nearly instantaneous. We have developed a novel pipeline that can perform an almost on-the-fly automated detection of satellites streaks and extraction of positional information from the astronomical data. The main challenges we address in this paper are the processing speed (keeping up with the incoming data stream) and the accuracy of the automated extraction of sky positions of satellites.
This paper outlines how we can increase both the processing speed and the detection performance with a novel image processing algorithm that allows us to detect objects throughout the whole night, including objects in geostationary orbit. Since MASCARA uses stationary cameras, the algorithm begins by freezing the stellar motion by aligning the subsequent images to the stars, after which difference images are made to highlight features of varying brightness and suppress the apparent brightness from stars. The removal of stellar motion introduces an artificial motion for geostationary satellites, giving them a streak-like appearance rather than a point source. Several of these difference images are combined to obtain a single composite image where only the highest value of each pixel-stack and its index within said stack are stored. The resulting composite image spans a longer time-frame without overexposed stars. The result from the operations is that single-image satellite streaks are combined into so-called satellite tracks. Not only do these tracks oftentimes span across the entire image frame, but they also contain discrete time information – the index in the stack directly translates to the timestamp from the originating image. Although longer features harbour curvature due to aberrations, the increase in feature length provides easier automatic detection. Naturally, the processing time is also decreased by our algorithm as less images need to be processed in later steps. Once image stacking is complete, binary detection images are created, after which we use a probabilistic Hough transform for feature detection. Next, we try to match the detections to objects from a catalogue of satellite Two Line Elements (TLEs). These TLEs are translated to sky-positions at the timestamps of the images by the Simplified General Perturbations propagator (SGP4), and then translated to image coordinates. Specifically, the timestamps are those at the edges of the exposures (start and end) and thus also at the crossover point between subsequent exposures. This approach of detecting satellite tracks rather than single-image satellite streaks provides improved reliability for automatic object identification. The increase in feature length on the image plane gives us more confidence in distinguishing between several different candidate objects since the matching method is based on the whole feature in the image rather than singular frame positions. For the night of testdata we have used to validate our methods, we are able to detect 100 unique satellites with high confidence (>92%).
Due to the limited accuracy and validity of the TLE format we expect and observe deviations between the detected feature and the matched catalogue TLE. Also when a feature is not matched to any (known) TLE there is a need to estimate the feature’s orbital parameters for possible later use. For all valid detections we want to solve the orbital parameters based on the automated extraction of the sky-positions of the satellite feature in the images. By extracting the sky-positions of these detected objects we use two novel methods; The first interpolates the detected features frame coordinates and discrete time-data, after which we predict the coordinates of the known timestamps from the regression models. Since there is often significant levels of noise in the detected features we evaluate multiple models more robust against outliers in the data. For the second method we trace the detected feature in the discrete time domain and convolute this response with a step function. In those step responses the peaks correspond to the sky-position for the given crossover timestamp. Since we only determine sky-positions at crossover points, we do not need to perform any correction for the point spread function (PSF). To validate these methods we compare the reference positions against the estimations from the two novel methods. We have demonstrated that both methods are able to automatically extract sky-positions of satellite tracks and we are currently investigating how they can be applied to alternate observation strategies.
Currently, we are investigating further optimisation of the pipeline and are making several improvements, where the main focus is the processing speed, satellite identification and further validation of the sky-position extraction. The pipeline is also taking into account that other observatories and cameras may be used, ensuring broader usefulness for SDA. We have begun the development of several (initial) orbit determination methods, such that in the near future, we will be able to cover the complete process; from observation and track analysis, via orbit determination to observation predictions. The RNLAF has commissioned this project which is a collaboration between the RNLAF’s Space Office, the Leiden University and Delft University of Technology.
Date of Conference: September 27-20, 2022
Track: Optical Systems & Instrumentation