Machine Learning Approach to Integral Field Unit Spectroscopy Observations: I. HII Region Kinematics

TL;DR

This paper introduces a CNN-based method for estimating emission line parameters from SITELLE integral field spectroscopy data, achieving high accuracy and significantly faster processing, demonstrating machine learning's potential in IFU data analysis.

Contribution

The paper presents a novel CNN approach trained on synthetic data for analyzing SITELLE spectra, improving speed and accuracy in extracting dynamical parameters of HII regions.

Findings

CNN recovers velocity and broadening with <5 km/s accuracy

Method reduces computation time by over tenfold

Effective on real SITELLE observations of M33

Abstract

SITELLE is a novel integral field unit spectroscopy instrument that has an impressive spatial (11 by 11 arcmin), spectral coverage, and spectral resolution (R=1-20000). SIGNALS is anticipated to obtain deep observations (down to 3.6x10-17ergs s-1cm-2) of 40 galaxies, each needing complex and substantial time to extract spectral information. We present a method that uses Convolution Neural Networks (CNN) for estimating emission line parameters in optical spectra obtained with SITELLE as part of the SIGNALS large program. Our algorithm is trained and tested on synthetic data representing typical emission spectra for HII regions based on Mexican Million Models database(3MdB) BOND simulations. The network's activation map demonstrates its ability to extract the dynamical (broadening and velocity) parameters from a set of 5 emission lines (e.g. H{\alpha}, N[II] doublet, and S[II] doublet) in the SN3 (651-685 nm) filter of SITELLE. Once trained, the algorithm was tested on real SITELLE observations in the SIGNALS program of one of the South West fields of M33. The CNN recovers the dynamical parameters with an accuracy better than 5 km s-1 in regions with a signal-to-noise ratio greater than 15 over the H{\alpha}line. More importantly, our CNN method reduces calculation time by over an order of magnitude on the spectral cube with native spatial resolution when compared with standard fitting procedures. These results clearly illustrate the power of machine learning algorithms for the use in future IFU-based missions. Subsequent work will explore the applicability of the methodology to other spectral parameters such as the flux of key emission lines.