AI helps in Identification of Astronomical Objects

Hubble Deep Field. Courtesy: Encyclopedia Britannica

Space is vast. The number of objects in the Universe is literally astronomical. The first step toward exploring these objects is to identify them aka classify them. Objects could be stars, galaxies, quasars, supernovae etc.

Due to the sheer number and complexity of objects the task is very challenging to do manually. The further away a source of light is, the harder it is to distinguish its features and thus classify it. Astronomers have now sought the aid of AI in this task.

A machine learning pipeline called SHEEP has been developed. SHEEP first extracts the photometric redshift of celestial objects and then uses this as one of the data features fed into an ML model for training.

SHEEP combines several ML algorithms: XGBoost, LightGBM, and Catboost to obtain better classification performance. SHEEP contains two distinct classification methodologies: (i) Multi-class and (ii) one versus all with correction by a meta-learner. The dataset used by the researchers is composed of 3.5 million astronomical sources for the classification of stars, galaxies, and quasars. 

In terms of classification performance, the resulting F1-scores are as follows: 0.992 for galaxies; 0.967 for quasars; and 0.985 for stars. In terms of the F1-scores for the three classes, SHEEP is found to outperform existing approaches that use an essentially identical data set. This methodology also facilitates model and data set explainability via feature importances; it also allows the selection of sources whose uncertain classifications may make them interesting sources for follow-up observations.

Future exploration using these techniques would involve applying the SHEEP pipeline to upcoming datasets such as that produced by Euclid, which is expected to produce large volumes of data necessitating the usage of ML techniques. Including varied information such as wider wavelength coverage, morphological information, and spectral information is expected to be explored further.

For full details of this research go here.