Neural network recognizes space objects in telescope images

(ORDO NEWS) — Portuguese astrophysicists have trained artificial intelligence to automatically identify galaxies, stars and quasars in images of millions of distant stars.

The classification of space objects in astronomical images is not as simple as it might seem. Being at great distances, they look like faint, often blurry points of light.

Looking at them, it is often difficult to say what we are dealing with – a star or a galaxy, a supernova or a quasar.

At the same time, large-scale surveys of the sky, which have been carried out in recent decades, have collected colossal arrays of images that require such an analysis. It is too difficult and time consuming to carry it out by traditional methods.

Therefore, scientists from Portugal decided to automate the task and use artificial intelligence for this. Pedro Cunha and Andrew Humphrey of the Institute of Astrophysics and Space Sciences (IA) have developed the SHEEP model, which can quickly and accurately classify distant space objects in telescope images.

SHEEP is a reinforcement learning model that uses the spectra and coordinates of space objects, and also determines the photometric redshift to more accurately classify them.

The redshift, together with the coordinates, allows the system to estimate the approximate position of the source in space in order to better recognize it.

For example, if an object is in the plane of the Milky Way, it is more likely to be a star, and if it is outside that plane, it is more likely to be a distant galaxy.

To demonstrate the operation of SHEEP, the authors analyzed with its help the extensive databases of surveys SDSS (made by ground-based instruments) and WISE (made by the space telescope of the same name), containing photometric data for about three and a half million distant sources.

The accuracy of the algorithm for determining stars was 98.5 percent, galaxies – 96.7 percent, and quasars (active nuclei of young galaxies) – 99 percent.

“By allowing artificial intelligence to include data on the position of sources in space, we improved its ability to make correct decisions about the nature of these sources,” said Professor Humphrey.

Scientists hope that their system will help find interesting objects in the huge data sets that will appear in the near future – in particular, thanks to the new ESA Euclid space mission.

The device is being prepared for launch in 2023 and will be engaged in ultra-precise measurements of the redshifts of distant galaxies.


Contact us: [email protected]

Our Standards, Terms of Use: Standard Terms And Conditions.