=Paper=
{{Paper
|id=Vol-2540/paper55
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2540/FAIR2019_paper_60.pdf
|volume=Vol-2540
}}
==None==
https://ceur-ws.org/Vol-2540/FAIR2019_paper_60.pdf
Foreground separation of interferometric 21cm
cosmological observations using convolutional
neural networks
Lisa Dayaram1,2[0000−0001−9711−0107] , Devin Crichton1,3 , and Anban
Pillay1,2[0000−0001−7160−6972]
1
University of KwaZulu-Natal, Westville Campus,Durban,4000, South Africa
2
School of Mathematics, Statistics and Computer Science
3
Astrophysics and Cosmology Research Unit
ldayaram@gmail.com
Abstract. The Hydrogen Intensity and Real-time Analysis eXperiment
(HIRAX) is an array of radio telescopes being deployed at the Square
Kilometre Array in South Africa, which will generate terabytes of inter-
ferometric data daily. This paper investigates the use of convolutional
autoencoders to separate cosmological signal from the foreground con-
taminations in the context of 21cm observations.
HIRAX is a radio telescope array or radio interferometer that aims to calculate
the statistical distribution of matter through the different epochs of time in an
attempt to measure the expansion history of the universe. A radio interferometer
is an array of radio telescopes, used to study the naturally occurring radio light
from stars, galaxies, black holes and other astronomical objects. These telescopes
collect weak radio waves, bring it to focus, amplify it and make it available for
analysis. The images created contain foreground contaminants. The foreground
of an image refers to a bright astrophysical radio source that contaminates the
target signal. Our goal is to separate these contaminants from the image in order
to accurately trace the 21cm signal.
Foreground separation is a significant task in radio astronomy [2][13]. It is
important to reduce the risk of these contaminants in the data to avoid the
foreground uncertainties biasing the analysis. In recent years, deep learning al-
gorithms have become increasingly popular among astronomers. Modern tele-
scopes are capable of producing data faster than astronomers can process. The
rapid growth in data size and complexity demands data-driven science in as-
tronomical analysis. The efficacy of deep learning techniques [3][11] in related
domains prompted the use of Convolutional Neural Networks [5] in classifying
the foreground contaminants and the target signal.
This project aims to create a filter to separate the cosmological signal of interest,
more specifically the 21cm emission line, from the foreground contaminations
using deep learning techniques.
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0)
A key point concerning the identification of the foreground is its spectrally
smooth nature as a frequency function which, in principle, allows it to be sep-
arated from the target 21cm signal. The network used in this work is an au-
toencoder with an underlying convolutional neural network (CNN) architecture.
The convolutional layers serve to convolve the input images, effectively preserv-
ing spatial dependence whilst the hidden layers of the autoencoder act as an
effective feature detector, in this case identifying the signal data from the fore-
ground contaminants. The network was evaluated using the loss and accuracy
of predictions.
The visibility timestreams (input data) of HIRAX observations were simulated
using Draco (a pipeline for the analysis and simulation of drift scan radio data),
along with the following python packages: Driftscan, Cora, Caput. Simulated
or synthetic data plays an integral role in the interpretation of observations.
Due to the varying properties and calibrations a telescope may have it becomes
challenging to keep track of all the anomalies that arise, which adds further
constraints on the task of identifying and separating the target signal from the
foreground.
(a) Prediction (b) Signal test data
Fig. 1: Graphical results of a single channel, over frequencies ranging from 500
-700MHz over three days (72 hours), show a significant subtraction of contami-
nants. The predicted outcome (a) is compared against the signal test data (b).
Foregrounds are several orders of magnitude more intense than the 21cm
signal and are highly correlated hence we track the progress of the model by
plotting the predicted outcomes, the signal-only data and the original (contam-
inated) inputs - resulting in graphs of frequency (MHz) against time(hrs). The
colour-bars generated aids in calculating the difference in magnitudes of what
is observed from the output vs. the signal-only data and is an estimate of how
well the model performed. From the results obtained during this experiment,
the network is able to recover some of the signal and has managed to subtract a
significant amount of the foreground contaminants.
We conclude that deep learning techniques have the potential to perform well in
this domain.
References
1. Shaw, J.R. : All-Sky Interferometry with Spherical Harmonic Transit Telescopes.
The Astrophysical Journal2(781), 57–69 (2014)
2. Berger, P.,Oppermann, N., Pen, Ue-Li, Shaw, J.R.: An efficient method for removing
point sources from full-sky radio interferometric maps. Monthly Notices of the Royal
Astronomical Society4(472), 4928-4934 (2017)
3. Connor, L., van Leeuwen, J.: Applying Deep Learning to Fast Radio Burst Classi-
fication. The Astronomical Journal Society6(156), 256 (2018)
4. Chang,Tzu-Ching., Pen, Ue-Li.,Peterson, J., McDonald, P. :Baryon Acoustic Oscil-
lation Intensity Mapping as a Test of Dark Energy. Physical Review Letters9(100),
091303 (2008)
5. Diaz, J.D., Bekki, K., Forbes, D., Couch, W.J., Drinkwater, M., Deeley,
S.:Classifying the formation processes of S0 galaxies using Convolutional Neural Net-
works.Monthly Notices of the Royal Astronomical Society4(486), 4845-4862 (2019)
6. Shaw, J. R., Sigurdson, K., Sitwell, M., Stebbins, A., pen, Ue-Li. :Coaxing Cosmic
21cm Fluctuations from the Polarized Sky using m-mode Analysis.Physical Review
D8(91), 083514 (2015)
7. Gillet, N., Mesinger, A., Greig, B., Liu, A., Ucci, G. :Deep learning from 21-cm
tomography of the Cosmic Dawn and Reionization.Monthly Notices of the Royal
Astronomical Society (2019)
8. Newburgh et al. :HIRAX: A Probe of Dark Energy and Radio Tran-
sients.arXiv:1607.02059 [astro-ph], 99065X (2016)
9. Baron, D.:Machine Learning in Astronomy: a practical overview.arXiv:1904.07248
[astro-ph], (2019)
10. Ammanouil et al.:Multi-frequency image reconstruction for radio-interferometry
with self-tuned regularization parameters.arXiv:1703.03608 [cs],(2017)
11. Kerrigan et al.:Optimizing Sparse RFI Prediction using Deep Learning.Monthly
Notices of the Royal Astronomical Society, (2019)
12. Zhang et al.:Bayesian semi-blind component separation for foreground removal
in interferometric 21-cm observations.The Astrophysical Journal Supplement Se-
ries,1(222), 3 (2016)
13. Cho et al.:A Technique for foreground subtraction in redshifted 21 cm observa-
tions.The Astrophysical Journal,2(749), 164 (2012)
14. Karras et al.: Progressive growing of GANs for Improved Quality, Stability, and
Variation. arXiv:1710.10196 [cs, stat], (2017)
15. Zhang et al.: Self-Attention generative Adversarial networks. arXiv:1805.08318[cs,
stat], (2018)
16. Smith, M.J., Geach, J.E.:Generative deep fields:arbitrarily sized, random synthetic
astronomical images through deep learning. arXiv:1904.10286[astro-ph], (2019)
17. Greig, B., Mesinger, A.:21CMMC: an MCMC analysis tool enabling astrophysical
parameter studies of the cosmic 21cm signal.Monthly Notices of the Royal Astro-
nomical Society, 449(4), 4246-4263 (2015)
18. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press,
http://www.deeplearningbook.org (2016)
19. GitHub, http://cs231n.github.io/convolutional-networks/. Last accessed 11 Oct
2019
20. GitHub, https://github.com/radiocosmology/draco. Last accessed 27 Sept 2019