Razim, Oleksandra (2021) Improving the reliability of photometric redshift catalogues with Self-Organizing Maps. [Tesi di dottorato]


Download (4MB) | Preview
[error in script] [error in script]
Item Type: Tesi di dottorato
Resource language: English
Title: Improving the reliability of photometric redshift catalogues with Self-Organizing Maps
Razim, Oleksandrashr.razim@gmail.com
Date: 15 February 2021
Number of Pages: 148
Institution: Università degli Studi di Napoli Federico II
Department: Fisica
Dottorato: Fisica
Ciclo di dottorato: 32
Coordinatore del Corso di dottorato:
Capozziello, Salvatorecapozziello@na.infn.it
Longo, GiuseppeUNSPECIFIED
Date: 15 February 2021
Number of Pages: 148
Keywords: Machine learning; Unsupervised machine learning; Photometric redshifts
Settori scientifico-disciplinari del MIUR: Area 02 - Scienze fisiche > FIS/05 - Astronomia e astrofisica
Date Deposited: 19 Feb 2021 09:51
Last Modified: 28 Oct 2021 11:59
URI: http://www.fedoa.unina.it/id/eprint/13280

Collection description

The already existing and upcoming massive surveys, such as KiDS, DES, Euclid and LSST, bring to reality many exiting possibilities in precision cosmology, as well as galaxy evolution and large-scale structure studies. However, to fully benefit from these surveys, we require redshifts for millions of galaxies. Currently, it is impossible to obtain these redshifts with spectroscopy only, since it would require immense observational time. For this reason, an alternative method, called photometric redshifts (photo-z) is used. This thesis is dedicated to a new data cleaning methodology, that allows to significantly improve the quality of photo-z catalogues and to guarantee the reliability of their quality metrics (i.e., to perform photo-z calibration). This methodology is based on an unsupervised Machine Learning (ML) algorithm called Self-Organizing Maps (SOM). Different components of this methodology allow to tackle several important issues. Namely, in-cell SOM anomaly detection helps to alleviate contamination of a spectral redshift catalogue with unreliable measurements and reduce the percentage of catastrophic outliers in photo-z predictions. Another approach, SOM occupation map calibration, counters the deterioration of the reliability of photo-z catalogues caused by differences between the parameter space of the train and run datasets. The methodology is tested on a deep 30-band photometric catalogue COSMOS2015. Photometric redshifts for this catalogue were obtained using a well-tested supervised ML algorithm MLPQNA. For additional comparison, SED (Spectral Energy Distribution) fitting photo-z were used. For both photo-z methods, the usage of the SOM-based data cleaning methodology reduces the percentage of catastrophic outliers by at least an order with corresponding improvements of other metrics. This result makes the SOM-based data cleaning a highly recommendable tool for preparing photo-z catalogues for the upcoming large surveys.


Downloads per month over past year

Actions (login required)

View Item View Item