Browsing by Author "Moya Sierralta, Cristóbal Andrés"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Depthwise convolutional neural network for multiband automatic quasars classification in ATLAS
    (2023) San Martín Jiménez, Astrid Elizabeth; Pichara Baksai, Karim Elías; Barrientos, Luis Felipe; Rojas Henríquez, Felipe Ignacio; Moya Sierralta, Cristóbal Andrés
    In recent years, the astronomical scientific community has made significant efforts to automate quasars' detection. Automatic classification of these objects is challenging since they are very distant and appear as point sources, outnumbered by other sources. Thus, performing automatic morphological classification is not straightforward; colour dimension seems better as a key concept. Previous work using machine learning tools has proposed classifiers that use features such as magnitude and colour, working only for quasar representation, which requires high-quality observational data that is not always available. Those features are computationally costly in extensive image surveys like VST ATLAS (Shanks et al. 2015). With the continuous developments in deep-learning architectures, we find a powerful tool to perform automatic classification from images, where capturing information from different bands takes relevance in this kind of approach. In this work, we developed a new quasar selection method that we hope to apply to the complete ATLAS survey in subsequent papers, where the completeness and efficiency of depthwise architecture will be compared to more standard methods such as selection on the colour-colour diagrams and machine-learning feature-based methods. This automatic quasar classification tool uses images in u, g, i, z bands available in ATLAS, heading towards new survey requirements facing the big data era. We propose a deep-learning architecture based on depthwise convolutional units that work directly with ATLAS images, reduced by the VST pipeline. Our model reaches an accuracy of 96.53 per cent with a quasar classification f1-score of 96.49 per cent, a very competitive benchmark compared to previous unscalable approaches....