Sam Wright, Spaceflux & University College London; Marco Rocchetto, Spaceflux; Ingo Waldmann, Spaceflux
Keywords: Optical Measurements, Deep learning, Astrometry, Data Reduction, SSA
Abstract:
Here we present the results of research and development conducted to produce an optical sensor data processing pipeline for space situational awareness (SSA) measurements. The pipeline produced is used as part of the Spaceflux telescope network. The pipeline ingests raw images collected from the telescope network and produces accurate tracking data messages (TDMs) for conveying astrometric and photometric measurements to downstream consumers.
The full pipeline comprises a number of processing techniques and two parallel approaches: a “classic” data analysis approach as well as a machine learning method to demonstrate the capabilities and benefits of a machine-learning solution. The pipeline has two parts: data reduction and data processing. The initial step calculates image statistics to perform background analysis and generate pixel brightness thresholds. The data reduction segment of the pipeline can ingest calibration frame sets and merge them to create master calibration ‘bias’, ‘dark’ and ‘flat’ frames. These are used to correct some systematic errors introduced by the sensor. The analysis portion of the pipeline can then be run on the reduced images.
When using the “classic” analysis approach, the data analysis segment locates the centroids of the bright spots by using the previously generated image statistics. Clustering algorithms are applied to the centroids and the resulting cluster geometries are used to classify the clusters as either point sources or streaks. The objects represented by the streaks and point sources depend on the telescope’s movement: in tracking rate mode (TRM) the points are satellites and the streaks are background stars, while in sidereal stare mode (SSM) this is reversed.
The pipeline then fits 2D Gaussian distributions to the identified sources to complete the characterisation of the image. By fitting these Gaussians, the pipeline can obtain brightness distributions for the sources and refine their categorisations as streaks or points; using a 2D Gaussian yields a shape which directs this refinement. The fitting process is performed iteratively using the prior streak shape to inform each successive fit; this transferability in the streak shape is due to the streak angle and length being functions of the observational setup. Once the stars in the image are identified, astrometry is performed to establish position coordinates; the astrometry routine queries the GAIA star catalogue using the known positions of visible bright stars.
We also evaluate a machine learning model approach for locating and classifying streaks and point sources. The model is a version of an object detection framework commonly used in computer vision tasks: ‘You Only Look Once’ version 5 (YOLOv5). This is trained on a subset of the simulated Shenanigans v0.3 data set generated by the SatSim package and obtained from the Unified Data Library (UDL); the subset is sampled by taking a single example image from each campaign in the simulated set to produce a total set size of 211 images, which was then split into a training set of 168 images and a validation set of 43 images. The model fitting process was accelerated by running on an Nvidia A100 GPU.
Beyond the functional requirement to produce the desired TDMs, we focused on efficiency in order to address real-time processing time performance constraints that arise when deploying at the “edge” – with the telescope. We achieved this by producing an efficient code base which runs with minimal processing time. Machine learning models can yield speed enhancements over analytically evaluated models when performing inference. Such speed enhancements can play an important role in our goal of reducing the computation time. We found a mean inference time for a test frame evaluated on a CPU using the ONNX inference-optimised framework to be 80.3 ms. The result is a data analysis pipeline with the required functionality and the capability to run within the environmental constraints.
Date of Conference: September 19-22, 2023
Track: Machine Learning for SDA Applications