Elsevier

Astronomy and Computing

Volume 42, January 2023, 100682
Astronomy and Computing

Full length article
Astronomical source detection in radio continuum maps with deep neural networks

https://doi.org/10.1016/j.ascom.2022.100682Get rights and content

Abstract

Source finding is one of the most challenging tasks in upcoming radio continuum surveys with SKA precursors, such as the Evolutionary Map of the Universe (EMU) survey of the Australian SKA Pathfinder (ASKAP) telescope. The resolution, sensitivity, and sky coverage of such surveys is unprecedented, requiring new features and improvements to be made in existing source finders. Among them, reducing the false detection rate, particularly in the Galactic plane, and the ability to associate multiple disjoint islands into physical objects. To bridge this gap, we developed a new source finder, based on the Mask R-CNN object detection framework, capable of both detecting and classifying compact, extended, spurious, and poorly imaged sources in radio continuum images. The model was trained using ASKAP EMU data, observed during the Early Science and pilot survey phase, and previous radio survey data, taken with the VLA and ATCA telescopes. On the test sample, the final model achieves an overall detection completeness above 85%, a reliability of 65%, and a classification precision/recall above 90%. Results obtained for all source classes are reported and discussed.

Introduction

Source finding is one of the most challenging tasks in upcoming radio continuum surveys, such as the Evolutionary Map of the Universe (EMU) (Norris et al., 2011) planned at the Australian SKA Pathfinder (ASKAP) telescope (Johnston et al., 2008, Hotan et al., 2021). The enhanced sensitivity, angular resolution and field of view will enable the detection of millions of sources in EMU. Such cataloguing process requires a level of automation and accuracy in knowledge extraction never reached before in source finder algorithms.

Although existing finders used in the radio community have been recently upgraded in this direction (e.g. see Hancock et al., 2018, Riggi et al., 2019, Carbone et al., 2018), and novel solutions have been developed (e.g. see Robotham et al., 2018, Hale et al., 2019, Lukas et al., 2019), many algorithmic aspects, critical for EMU but also for future SKA surveys, remain to be tackled, particularly for observations carried out in the Galactic plane. Identification of spurious sources is certainly a priority, particularly for observations with a significant diffuse background or very extended sources, where the false detection rate obtained by standard source finders can exceed the 20% level (Riggi et al., 2021a). Spurious detections in island1 catalogues (and consequently also in the corresponding fitted component catalogues) are due to the background noise and artefacts introduced in the imaging process. Among them, sidelobes around bright sources, dominate at high S/N levels, e.g. well above the standard 5σ detection threshold. Spurious detections in component catalogues are partly due to the presence of spurious islands but, mostly, to the island over-deblending of existing finders, particularly when estimating components of extended or non-gaussian islands. These spurious detections are only rarely automatically rejected in large area surveys (for example using ad-hoc selection cuts), e.g. the most widely-used approach is identifying them by visual inspection.

Source identification from multiple non-contiguous islands and classification into known classes of astrophysical objects is another poorly covered task in traditional source finders. This is particularly relevant when searching for Galactic objects in Galactic plane surveys. In these studies, the extragalactic objects constitute the most numerous background sources (90% of the catalogued sources). Although the majority of them have a single-island morphology, radio galaxies with an extended morphology (e.g. including multiple islands associated to their physical core, lobe, or jet components) can easily exceed the number of Galactic objects previously known in the considered map. For instance, in the ASKAP Scorpio survey (Riggi et al., 2021a), the number of islands associated to radio galaxies was found to be a factor 3 larger than those associated to known or candidate Galactic objects previously reported in the literature. Identification and removal of this kind of sources would therefore ease the search of unclassified Galactic objects.

Machine learning was already proven to be a valuable tool for tackling most of the aforementioned tasks. For example, Lukic et al., 2018, Lukic et al., 2020 and Wu et al. (2019) employed deep Convolutional Neural Networks (CNNs) for detecting and classifying radio galaxies in extragalactic fields. New source finders, based on deep networks, were also recently implemented and made available to the radio community. ConvoSource (Lukic et al., 2020), for instance, is a CNN-based tool for semantic segmentation of radio sources. It was trained on a dataset composed of simulated compact and extended star-forming galaxies (SFGs) and Active Galactic Nuclei (AGN) (both steep- and flat-spectrum populations), as modelled in the SKA Data Challenge I (SDC1) (Bonaldi et al., 2020). Best performances (precision = 0.73, recall = 0.83, F1-score = 0.78, all classes, SNR > 5) were obtained on SKA Band 1 simulated maps with high integration times (1000 h).

ClaRAN (Wu et al., 2019), a Faster R-CNN (Ren, 2017) based model, detects and classifies radio galaxies of different morphological classes, with overall Mean Average Precision (mAP) ranging from 0.77 to 0.84, depending on the data pre-processing used. It exploits both real radio and infrared input data, contrarily to other tools, which only use radio data.

DeepSource (Vafaei Sadr et al., 2019) is another CNN based solution that only detects point-sources, it does not classify objects, nor does it output a segmentation mask. The reported precision (recall) values on simulated datasets range from 0.45 (0.85) for an S/N of 3σ, to 0.99 (0.99) for S/N above 4σ.

Mostert et al. (2022) employed Fast R-CNN architectures to perform radio-component association from the LOFAR Twometre Sky Survey (LoTSS) data, obtaining a level of accuracy (84.0%) comparable to that typically reached in crowdsourcing analysis.

In this context, it is important to highlight that all of these solutions are not directly comparable in performance as they were trained and tested on different datasets (real or simulated, different S/N levels, and dataset sizes, etc.), targeting different types of objects, and producing different types of outputs.

In this work we present a new source finding tool, named caesar-mrcnn (Compact And Extended Source Automated Recognition with Mask R-CNN), aiming to tackle the discussed aspects in source extraction, using Mask R-CNN instance segmentation framework. caesar-mrcnn was trained to both detect and classify radio sources of different morphologies (compact or extended), imaging artefacts, and poorly imaged sources. The paper is organized as follows. In Section 2 we describe the radio observations used, and the dataset produced for training and testing scopes. In Section 3 we describe the Mask R-CNN object detection framework, and source finder implementation details. In Section 4 we present the detection and classification results obtained on test radio images. Finally, future perspectives are reported in Section 5.

Section snippets

ASKAP pilot surveys

The ASKAP EMU early science program was started in 2017, while the array commissioning was almost completed, to validate the array operations, the observation strategy, and optimize the data reduction pipeline. In this phase, several pilot surveys were carried out on target fields, also including the Galactic plane, bringing first scientific results. Among them, the EMU pilot survey (Norris et al., 2021) (area = 270 deg2, rms = 25–30 μJy beam−1, angular resolution = 12.5 × 10.9 arcsec,

Mask R-CNN: A framework for object detection and classification

Mask R-CNN is a deep learning model that has been recently proposed, combining the capability of performing object detection, classification, and instance segmentation on images. The algorithm was developed by the Facebook AI Research team in 2017 and has been used in many computer vision problems such as identifying vehicles or faces (He et al., 2017), marine mammals (Gray et al., 2019), and astronomical optical sources (Burke, 2019). Mask R-CNN is built upon, and shares a similar architecture

Model training and parameter optimization

The full image set described in Section 2 was randomly split into a train, validation, and test set, containing 60%, 10%, and 30% of the original sample size, respectively. Several training runs were then carried out on the train and validation sets to tune Mask R-CNN hyper-parameters, and understand their impact on the source detection and classification performance (see Sections 4.2 Source detection performance, 4.3 Source classification performance). Most parameters were kept to their

Summary and outlook

In this work we have presented a new source finding tool, based on the Mask R-CNN instance segmentation framework, for detecting and classifying compact, extended, flagged and spurious sources in radio continuum maps. The method was tested on a dataset of images extracted from different radio surveys, including ASKAP EMU Early Science and pilot data, and performances were studied for different parameter values. The implemented tool offers these novelty aspects when compared to traditional radio

CRediT authorship contribution statement

S. Riggi: Conceptualization, Methodology, Software, Data curation, Investigation, Formal analysis, Project administration, Supervision, Writing – original draft, Writing – review & editing, Funding acquisition. D. Magro: Software, Data curation, Investigation, Software, Writing – original draft, Writing – review & editing. R. Sortino: Software, Data curation, Investigation. A. De Marco: Data curation, Investigation, Writing – review & editing. C. Bordiu: Data curation, Methodology, Writing –

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Cristobal Bordiu reports financial support was provided by EU Framework Programme for Research and Innovation.

Acknowledgements

The Australian SKA Pathfinder is part of the Australia Telescope National Facility which is managed by CSIRO. Operation of ASKAP is funded by the Australian Government with support from the National Collaborative Research Infrastructure Strategy, Australia. Establishment of the Murchison Radio-astronomy Observatory was funded by the Australian Government and the Government of Western Australia. This work was supported by resources provided by the Pawsey Supercomputing Centre with funding from

References (46)

  • CarboneD.

    Astron. Comput.

    (2018)
  • FreiM.

    Powder Technol.

    (2021)
  • RiggiS.

    Astron. Comput.

    (2021)
  • Abadi, M., et al., 2015....
  • Astropy CollaborationA.M. et al.

    Astron. J.

    (2018)
  • Astropy CollaborationT.P. et al.

    Astron. Astrophys.

    (2013)
  • BanfieldJ.K.

    Mon. Not. R. Astron. Soc.

    (2015)
  • BeckerR.H.

    Astrophys. J.

    (1995)
  • BonaldiA.

    Mon. Not. R. Astron. Soc.

    (2020)
  • BurkeC.J.

    Mon. Not. R. Astron. Soc.

    (2019)
  • Chollet, F., et al., 2015....
  • Girshick, R., 2015. IEEE International Conference on Computer Vision. ICCV, Santiago, pp. 1440–1448....
  • GrayP.C.

    Methods Ecol. Evol.

    (2019)
  • HaleC.

    Mon. Not. R. Astron. Soc.

    (2019)
  • HancockP.

    PASA

    (2018)
  • He, K., et al., 2016. IEEE Conference on Computer Vision and Pattern Recognition. CVPR, Las Vegas, NV, pp. 770–778....
  • He, K., et al., 2017. IEEE International Conference on Computer Vision. ICCV, Venice, pp. 2980–2988....
  • HopkinsA.

    PASA

    (2015)
  • HotanA.

    PASA

    (2021)
  • IngallineraA.

    MNRAS: Lett.

    (2022)
  • JohnstonS.

    Exp. Astron.

    (2008)
  • JoyeW.A. et al.
  • Lin, T., et al., 2017. IEEE Conference on Computer Vision and Pattern Recognition. CVPR, Honolulu, HI, pp. 936–944....
  • Cited by (13)

    View all citing articles on Scopus
    View full text