Classification in astronomy and astrophysics

How it is used

Astronomy produces far more observations than anyone can look at. Survey telescopes photograph the whole visible sky repeatedly, and each pass flags everything that has changed in brightness or position. Classification means taking those flagged objects and deciding which known kind each one is, or simply whether it belongs to the class a particular team cares about. The examples here cover sorting alerts from sky surveys into exploding stars and other variable sources, labelling radio sources in images of the sky, separating stars by their chemical make-up or their birthplace, and reading a star's type from a single image.

Two broad approaches appear. One extracts a handful of measured quantities first, such as the shape of a brightness curve over time, and feeds those numbers to a model that learns which combinations go with which class. The other feeds the pixels of the image itself to a network that learns its own internal description of what it sees. Models are trained on examples already labelled by people, by spectroscopy, or by a computer simulation in which the true answer is known.

Why researchers use it

The reason is scale and timing. Survey telescopes raise more candidates each night than astronomers can inspect, and for some objects the useful window is short: a supernova is most informative in its first days, and a telescope must be pointed at it before that window closes. Deciding by hand is slow and the standards drift between people. A trained model applies one consistent rule to every candidate within seconds of the alert arriving, so limited follow-up time on large telescopes goes to the objects most likely to repay it.

Classification also stands in for measurements that are expensive or missing. A spectrum, which spreads a star's light out by wavelength, settles many questions but takes dedicated telescope time, so most catalogued stars lack one. The same holds for detailed chemical abundances and for knowing whether a star formed in our galaxy or was captured from another. A model trained where those answers exist can extend labels to the much larger population where they do not.

In the record

Limits

A classifier only knows the categories in its training set, and real surveys contain objects nobody labelled, which are then forced into whichever class fits worst. Labelled sets are also unrepresentative: they favour bright, nearby and already interesting objects, so performance drops on the faint majority. Where labels come from simulations, the model may learn the simulation's quirks rather than the sky's. Rare classes are rare in training too, so the model can score well overall while missing nearly all of them. A confident label is not a measurement, and mistakes propagate quietly into whatever is built on top.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Classification: the idea, across every field
• AI in astronomy and astrophysics

Every article