Automated Satellite Track Detection and Endpoint Determination in Astronomical Images

Abstract

The Royal Netherlands Airforce (RNLAF) is currently supporting the Feasibility study for Optical Tracking of Orbital Satellites (FOTOS), which aims to create a satellite observation instrument, FOTOS1, for the optical tracking of satellites, which would be a key tool towards maintaining a sustainable use of space. This thesis was tasked with producing an improved image processing pipeline with the aim of detecting more objects and accurate endpoints for orbit determination. Additionally, it was important for the pipeline to maintain the cost and time efficiency of FOTOS1. For time and cost efficiency, data was firstly reduced; the data used originated from the MASCARA instrument in Chile, which takes subsequent exposures at a fixed exposure length from dusk until dawn. The instrument consists of five cameras that virtually covers the whole local horizon. The number of frames to be processed varies depending on the time of year, but is in the range of 23, 500 to 34, 000 images. Since optical satellite tracking requires that the satellites are sunlit, we could reduce the number of images for processing by specifying an observable altitude limit of ℎ = 2, 000 kilometers. This reduced ranges from 4.35% up to 48.38% depending on the day of the year. For higher altitude limits the reduction becomes significantly smaller as the earthshadow becomes less of an effect. To increase the detection performance of the pipeline we combined 50 images to create a single detection image. Several operations were performed to highlight dynamic features among the starry background. The astrometric solution was used to align the images together and subtract subsequent images from one another. These difference images were stacked together by their maximum value to highlight longer streaking features. For the detection, the Probabilistic Hough Transform worked efficiently and returned the correct positions on the image frame. The created detection images did not lead to an increase in the number of uniquely detected objects, but it did extend the altitude range of detected objects and the number of endpoints per unique object. Since the features in the stacks are longer the detection method was able to detect objects up to GEO altitude. Also since the stacks consisted of 50 images, a single detection of a feature could contain data from multiple images and thus contain multiple endpoints. Because of this approach the number of endpoints per unique object was increased almost fourfold whilst only detecting 20% less unique objects in total. Our last task was to determine the endpoints within one track; two novel methods were produced and tested with the aim to use position and discrete time data to create an overall better representation. However, it turned out that fitting noisy data was difficult for the tested regressors (least squares, TheilSen and RANSAC). It oftentimes caused the determinations to be off by 15 or more pixelsruling those results useless for orbit determination. Both the index prediction and index tracing methods had trouble defining the shape due to noise and thus can benefit from a possible iterative approach for data selection. The double index prediction method combined with a TheilSen regressor was selected for the pipeline as it showed to be the best performing method being the most robust combination and had relatively little error compared to the other combinations. When comparing the best performing new method to the existing method, it turned out that the existing method performed better. The endpoint accuracy of the new method was centered around 2 pixels but was more distributed than the existing methods. However the quantity of endpoints was almost doubled. How these results translate into quality of orbit determination is a recommended topic for future work. During this thesis project we therefore produced a new pipeline which can be implemented for instruments with the same observation strategy as MASCARA and is compatible with different lenses and exposure times. It showed proofofconcept that stacking images and through the data reduction simulation allowed for the processing of all the images within 24 hours such that a backlog of data will be avoided.