Super-resolving Herschel - a deep learning based deconvolution and denoising technique

Koopmans, Dennis (2023) Super-resolving Herschel - a deep learning based deconvolution and denoising technique. Master's Thesis / Essay, Astronomy.

	Text Dennis_Koopmans_Msc_Thesis.pdf Restricted to Registered users only Download (98MB)
	Text toestemming.pdf Restricted to Registered users only Download (166kB)

Abstract

The angular resolution of astronomical imaging is of key importance to its scientific exploitation, particularly for single-dish far-infrared and sub-mm surveys which typically have a factor 10 or worse resolution compared to optical surveys. Poor resolution of long-wavelength imaging data makes it highly challenging to identify multi-wavelength counterparts of detected dusty star-forming galaxies that are only detectable at the IR regime and appear to be blended. Additionally, it makes it challenging to make plans for any follow-up observations. One could observe the same field with bigger telescopes to achieve better solution but this is not always feasible. Analytical deblending approaches that use source priors detected in observations at higher resolution exist but these methods involve long computation times (∼ 60days) and are often limited to one prior per beam. This is impractical for the application on many observed fields. Additionally, with the launch of EUCLID and the James Web Space Telescope (JWST) the quantity of high resolution data only increases. In this thesis, I aim to develop a deep-learning based method to deconvolve and denoise deep Herschel imaging of the COSMOS field at 500µm, whose native resolution is 36.6”, and that can improve the angular resolution by a factor of ∼ 4.5. To achieve this I approach the problem in two ways: 1) by using a generative adversarial network (GAN) based on the Wasserstein GAN and 2) a transformer model that has never been applied in astronomy at far-infrared observations. The input maps for these models are the three native Herschel SPIRE bands at 250, 350and500µm combined with the Spitzer MIPS 24µm map. To train and test this method, I use the data generated by the Simulated InfraRed Extragalactic Sky (SIDES), Spectro-Photometric realisations of InraRed Selected Targets at all-z (SPRITZ) and SHARK simulation codes. Moreover, I use real JCMT SCUBA-2 observations at 450µm provided by the STUDIES team to evaluate the performance on observational data. I find that the transformer model is the best performing model compared to the Wasserstein GAN but also other state-of-the-art models and as a result, I use this model to super-resolve the observational 500µm images. Finally, I demonstrate how the superresolved Herschel maps can aid studies such as number counts of the FIR by providing the 500µm super-resolved number counts realised in this work. Finally, I match the detected sources of this work with other analytical methods and find 500µm sources that may constitute a completely new detection at this wavelength.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Wang, L. and Margalef Bentabol, B.
Degree programme:	Astronomy
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	11 Sep 2023 12:04
Last Modified:	14 Sep 2023 14:30
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/31377

Actions (login required)

View Item