Deep learning applied to satellite detection in the context of space surveillance and tracking

The recent decrease of space launch cost has led to a large number of active satellites and space debris around the Earth. Due to perturbing forces (e.g. gravitational, air drag), these objects are not bound to stable orbits, and their change in position must be tracked, and be corrected to avoid collisions. In the context of Space Surveillance and Tracking (SST), several ground-based networks of optical telescopes are routinely monitoring the sky, and updating the objects position. Most of this routine currently involves the active user involvement in satellite identification in the images, which is time consuming, while also slowing the process.

In this study, we present a method of identifying moving objects across the image (tracklets) based on deep learning (DL) methods. The images acquired from the telescope [1] are reduced, and regions of interest (ROI) centered around sources of light are extracted, and labelled into either objects of interest, or stars. Next, the set of images (N=185.000) was divided into training and validation, and a set of DL algorithms were tested for their ability to correctly classify the data. We found that the algorithms performed best when several data augmentation techniques were employed, to reach a score of 99% classification accuracy. An additional increase in score resulted when the raw pixels values (i.e. as obtained from the CCD) were used for the ROI-based model training, resulting in a final score of 99.99% accuracy.

These methods are being evaluated for the integration in an image reduction software (i.e., gendared [2]) as part of a Romanian R&D grant. It is expected that this will further increase the efficiency of the tool, enabling it to deliver accurate astrometric solutions at an increased rate, utterly needed as the near-Earth civilian infrastructure continues to grow. [1] Birlan M., et al., 2021, RoAJ, 31, 41 [2] Piso A.-M., et al., 2021, Proc. 8th European Conference on Space Debris (virtual), Darmstadt, Germany, 20-23 April 2021 Acknowledgements: This work was supported by a grant of the Romanian Ministry of Education and Research, CCCDI - UEFISCDI, project number PN-III-P2-2.1-PTE-2019- 0554, within PNCDI III.