statmorph-lsst: Quantifying and correcting morphological biases in galaxy surveys

TL;DR

This work addresses how imaging quality biases morphologies in galaxy surveys and provides a comprehensive framework to quantify and correct these biases. By degrading a large, diverse local galaxy sample, the authors map the dependence of all statmorph and single-Sérsic metrics on resolution and depth, and derive empirical corrections via symbolic regression, including two new measures, A_X and St. They show that geometric measures are generally robust, while concentration-based bulge indicators and disturbance metrics are substantially biased by PSF, resolution, and signal-to-noise, explaining part of the apparent evolution seen in high-z studies. The study yields practical outcomes: the statmorph-lsst Python package, a 64,000-image augmentation dataset, and a set of correction functions to enable bias-aware analyses for Rubin LSST and multi-wavelength morphology studies, facilitating robust interpretation of galaxy structure across cosmic time.

Abstract

Quantitative morphology provides a key probe of galaxy evolution across cosmic time and environments. However, these metrics can be biased by changes in imaging quality - resolution and depth - either across the survey area or the sample. To prepare for the upcoming Rubin LSST data, we investigate this bias for all metrics measured by statmorph and single-component Sérsic fitting with Galfit. We find that geometrical measurements (ellipticity, axis ratio, Petrosian radius, and effective radius) are fairly robust at most depths and resolutions. Light concentration measurements ($C$, Gini, $M_{20}$) systematically decrease with resolution, leading low-mass or high-redshift bulge-dominated sources to appear indistinguishable from disks. Sérsic index $n$, while unbiased, suffers from a 20-40% uncertainty due to degeneracies in the Sérsic fit. Disturbance measurements ($A$, $A_S$, $D$) depend on signal-to-noise and are thus affected by noise and surface-brightness dimming. We quantify this dependence for each parameter, offer empirical correction functions, and show that the evolution in $C$ observed in JWST galaxies can be explained purely by observational biases. We propose two new measurements - isophotal asymmetry $A_X$ and substructure $St$ - that aim to resolve some of these biases. Finally, we provide a Python package statmorph-lsst implementing these changes and a full dataset that enables tests of custom functions (see text for links).

Figures (22)

• Figure 1: A subset of augmentations performed for an example galaxy, NGC 17, observed in F814W (I band). The $1\sigma$ surface brightness limit $\mu_0$ decreasing left to right, and the resolution $\mathcal{R}$ is degrading top to bottom. Foreground sources detected on each image are shown in white contours. As depth degrades, the large tidal tail to the southwest is lost in noise, while the internal disturbances are invisible in low-resolution imaging.
• Figure 2: The measurement error compared to the baseline for the asymmetry centre ($\mathbf{x}_0^A$, left), ellipticity ($e$, middle), and orientation ($\theta$, right), as a function of the average signal-to-noise per pixel $\langle$SNR$\rangle$ and the effective resolution $\mathcal{R}_{\rm{eff}}$. For each galaxy and each image, the error is calculated as the difference in measurement compared to the baseline. $\mathbf{x}_0^A$ is normalized by the Petrosian radius (Sec. \ref{['sec:rpet']}). The colored distribution shows the median error in each bin, while the gray contours show 1$\sigma$ scatter. Both $\mathbf{x}_0^A$ and $e$ are fairly robust to noise, with an increased uncertainty at $\langle$SNR$\rangle$<1 but no systematic offsets. They are however biased at low resolution, with typical error in $\mathbf{x}_0^A$ being $0.2\pm0.07 R_p$, and typical ellipticity error of $0.15 \pm 0.12$. $\theta$ is less reliable in the low signal-to-noise regime, with an average $20\pm35 \degree$ error when $\langle$SNR$\rangle$$\sim 1$, and an over 10$\degree$ scatter at all resolutions and depths. This error is mostly driven by circular sources, and if only inclined ($e>0.3$) sources are considered, the error in $\theta$ is below 5$\degree$. Compared to Fig. \ref{['fig:moments_grid']}, the inclusion of the PSF effect significantly improves $e$ and $\theta$ measurements.
• Figure 3: Same as Fig. \ref{['fig:moments_grid']}, for three radius metrics: $R_{p,\circ}$ (left), $R_{20}$ (middle), and Sérsic $R_{0.5}$ (right, discussed in Sec. \ref{['sec:sersic']}). The bias and scatter are measured as a fraction of the baseline, rather than an absolute offset.$R_{p,\circ}$ and $R_{20}$ are more correlated to the pixel scale $\mathcal{R}$ while $R_{0.5}^{\rm{Sersic}}$ depends on the effective resolution $\mathcal{R}_{\rm{eff}}$. The non-parametric radii are overestimated at low resolutions, up to 20% and 80% for $R_{p,\circ}$ and $R_{20}$ respectively, but $R_{0.5}^{\rm{Sersic}}$ is slightly underestimated (up to 10%) compared to the baseline. However, the uncertainty in $R_{0.5}^{\rm{Sersic}}$ is larger than that for non-parametric radii.
• Figure 4: The distribution of $R_{p,\circ}$ measurements compared to the baseline before (left) and after (right) a signal-to-noise and resolution correction from Eq. \ref{['eq:rpet']}. Before the correction, $R_{p,\circ}$ is overestimated when resolution is worse than 500 pc/px (black contours), and slightly underestimated when $\langle$SNR$\rangle$ is below 2 (white contours). After applying the correction, we can recover the baseline $R_{p,\circ}$ for a wide range of resolutions and depths.
• Figure 5: Same as Fig. \ref{['fig:moments_grid']}, for three Sérsic parameters: $n$ (left), ellipticity (middle), and orientation (right). $n$ is underestimated by up to 0.5 when there are fewer than 10 resolution elements per object, and it is uncertain, with $\pm1.3$ scatter at lowest resolution. However, $n$ is robust to noise on average, with a $\pm0.5$ scatter at lowest $\langle$SNR$\rangle$. Both $e^{\rm{Sersic}}$ and $\theta^{\rm{Sersic}}$ are, on average, much more reliable than their non-parametric counterparts (Fig. \ref{['fig:moments_grid']}), with negligible offsets from the baseline and increasing scatter in the lowest resolution and signal-to-noise regimes.