arXiv2025
Accurate redshift measurements are essential for studying the evolution of quasi-stellar objects (QSOs) and their role in cosmic structure formation. While spectroscopic redshifts provide high precision, they are impractical for the vast number of sources detected in large-scale surveys. Photometric redshifts, derived from broadband fluxes, offer an efficient alternative, particularly when combined with machine learning techniques. In this work, we develop and evaluate a neural network model for predicting the redshifts of QSOs in the Dark Energy Spectroscopic Instrument (DESI) Early Data Release spectroscopic catalogue, using photometry from DESI, the Widefield Infrared Survey Explorer (WISE) and the Galactic Evolution Explorer (GALEX). We compare the performance of the neural network model against a k-Nearest Neighbours approach, these being the most accurate and least resource-intensive of the methods trialled herein, optimising model parameters and assessing accuracy with standard statistical metrics. Our results show that incorporating ultraviolet photometry from GALEX improves photometric redshift estimates, reducing scatter and catastrophic outliers compared to models trained only on near infrared and optical bands. The neural network achieves a correlation coefficient with spectroscopic redshift of $0.9187$ with normalised median absolute deviation of $0.197$, representing a significant improvement over other methods. Our work combines DESI, WISE and GALEX measurements, providing robust predictions which address the difficulties in predicting photometric redshift of QSOs over a large redshift range.