Belén Yu Irureta-Goyena Chang, E´cole Polytechnique Fe´de´rale de Lausanne; Elisabeth Rachith, E´cole Polytechnique Fe´de´rale de Lausanne; Jean-Paul Kneib, E´cole Polytechnique Fe´de´rale de Lausanne
Keywords: NEO, machine learning, satellite, space debris, VST, Euclid
Abstract:
The present work explores the application of a convolutional neural network to detect moving objects in wide-field astronomical surveys, with a special emphasis on near-Earth objects (NEOs), but also including satellites, space debris, main-belt asteroids, and transneptunian objects (TNOs). The algorithm is developed and tested using images from OmegaCAM, the wide-field camera mounted on ESOs VLT Survey Telescope (VST). The expectation is that the developed tools will also be applicable to a range of other satellite and ground-based surveys, setting the stage for the upcoming generation of telescopes.
The proposed method combines the astronomical software Source Extractor with a machine-learning algorithm. Firstly, Source Extractor is used to detect any sources present on a given image. Secondly, the machine-learning algorithm is applied to determine whether the source is a moving object. The machine-learning algorithm features a convolutional neural network whose architecture is a modified version of the VGG-16 Very Deep Convolutional Network. This model comprises 13 convolutional layers followed by 3 fully connected layers and a rectified linear unit activation function (ReLU) that accounts for non-linearity, an essential feature in image recognition.
Both the adjustment of the hyperparameters of the model and the model training are done using a set of photometric data taken with ESO’s VST in Paranal, Chile, under the purview of Programme ID 106.216P.002. In doing this development and to make the procedure more efficient, we also implement physical constraints based on our knowledge of real observing conditions. This programme monitors gravitationally lensed quasars by taking four five-minute exposures in the visible band of the same fields each night. The camera used to take the images is OmegaCAM, a wide-field camera with a field of view of 1 x 1 square degrees. The training set is built by artificially adding tracks of different lengths, orientations and brightness to the VST images.
From its orbit in the L2 Lagrangian point, the upcoming Euclid space mission will also be able to detect moving objects provided a careful analysis of the data collected. These will be essentially main-belt asteroids and transneptunian objects, or even other interstellar objects visiting our solar system, such as 1I/2017 U1 Oumuamua. With the launch of the Euclid space mission being scheduled for 2023, the Euclid Consortium has developed the software Elvis to generate images similar to those that will be observed during the mission. These simulated images will also be used to further test the applicability of our pipeline on disparate data sets.
Given the wide field of view of the telescopes used (wider than previously used telescopes), we expect to discover objects not yet identified. A catalogue of the detected objects will be produced, analysing trends over the last 10 years, which will be particularly relevant in the case of satellites. In particular, a comprehensive accounting of satellites and space debris in astronomical images will significantly improve the busy eco-system of the Low Earth Orbit, facilitating the coordination of space traffic and reducing the hazard of debris jeopardising space missions. The data will be compared with that available in satellite databases such as Astriagraph.
Furthermore, information on the orbital elements of the detected NEOs will also serve to update the astrometry currently available in dedicated databases such as the Minor Planet Center. Since the astrometry of NEOs remains relatively constant over time, this effort to better characterise it will refine our long-term monitoring of these objects. This will improve our capacity to identify NEOs that may pose a threat to Earth, mitigating local and global damage.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications