Name: Atharva Bagul
Affiliation: Indian Institute of Science Education and Research Bhopal
Conference ID : ASI2022_403
Title : Photometric classification of compact galaxies, stars and quasars using multiple neural networks
Authors : Siddharth Chaini, Atharva Bagul, Anish Deshpande, Rishi Gondkar, M Vivek, Kaushal Sharma, Ajit Kembhavi
Abstract Type: Poster
Abstract Category : Extragalactic Astronomy
Abstract : Classification problems in astronomy are well studied, an example being the classification of point-like sources into stars, galaxies and quasars. Advanced deep learning methods work well for classifying bright sources, including galaxies close to being point-like. However, for faint and very compact sources, the performance of even these methods deteriorates. As newer surveys go deeper into the sky, they observe many compact, faint galaxies, and distinguishing them from stars and quasars becomes critical. We present MargNet - a deep learning model that caters to such star-galaxy-quasar classification for extensive surveys. We train the model using photometric features and images obtained from the SDSS DR 16 catalogue for 240,000 objects, labelled into one of three classes based on the spectroscopic classification from the SDSS DR16 catalogue. We also include a sub-sample of compact galaxies, with the ratio of half-light radius to the point spread function (PSF) FWHM less than 0.5. The smaller the ratio is, the fainter and more compact the galaxy. MargNet consists of two neural network architectures - convolutional neural networks (ConvNets) and artificial neural networks (ANNs). The ConvNet input consists of 5-passbands images, while the ANN input consists of photometric features - half-light radius, dereddened magnitude, PSF FWHM and extinction coefficient for each passband, along with the colours. We finally combine the ConvNet and the ANN through a stacking ensemble called MargNet. We test MargNet on two types - compact sources and compact sources with r-band magnitude fainter than 20 and obtain the average accuracies of 93.3% and 86.7%, respectively. We find that our implementation outperforms all existing methods in the literature for both categories. We thus successfully show that the compact criterion described earlier is a powerful approach for future surveys like the LSST.