Early Results from GLASS-JWST. XVIII. A First Morphological Atlas of the 1 < z < 5 Universe in the Rest-frame Optical

On June 28–29 and 2022 November 10–11, the GLASS-JWST Early Release Science (ERS) program used the NIRISS wide-field slitless spectrograph and NIRCam imager to observe the A2744 lensing cluster. While the cluster core was observed with the grism spectrograph, NIRCam imaging was taken in parallel mode in seven bands from ∼1 to 5 μm: F090W, F115W, F150W, F200W, F277W, F356W, and F444W. The data were reduced using the NIRCam calibration files (cjwst 1014.pmap to cjwst 1019.pmap) provided by STScI. The F444W band was selected for source detection and all images were PSF-matched to this band producing a catalog reaching typical 5σ depths of ∼30.2 mag in all bands. More details regarding the observations, data reduction, and catalog generation can be found in Paris et al. (2023, Paper II) and Merlin et al. (2022, Paper II). Although the parallel fields are in the vicinity of the cluster A2744, they are sufficiently offset from the cluster core that we expect lensing magnification to be a modest effect (Medezinski et al. 2016), which does not affect morphological classification. The Gini coefficient is also robust to lensing effects.

We start with the GLASS-JWST source catalog (Paper II), which is the result of running Source Extractor (Bertin & Arnouts 1996) on the F444W imaging and selecting sources with signal-to-noise ratio (S/N) greater than 2. This yields a catalog of 6689 sources. In order to improve the quality of photometric redshifts, for the other six bands, we match the PSFs to the F444W band and perform aperture photometry in six circular apertures between 02 and 22 in diameter. This is solely for the purposes of photometry; we perform a different PSF matching to construct our color images described in Section 2.3 below. The final catalog has a 5σ magnitude limit of 29.7 in F444W.

Since we will compare our NIRCam sample with the lower-redshift universe, we select sources where morphology will be detectable and approximate rest-frame optical colors can be seen. First, as per Nanayakkara et al. (2023, Paper XVI), we obtain photometric redshifts with the EAZY (Brammer et al. 2008) photometric redshift code using eazy_v1.3.spectra.param templates provided by the software and total flux measured in a 045 aperture (i.e., 3 × FWHM; Paper II). We select all galaxies that have S/N > 5 detections in all GLASS-JWST bands and visually inspect the quality of the spectral energy distribution (SED) fits. To produce RGB images that are both close to visual true color and adopt a common convention for RGB visualization of local sources, we choose rest-frame irg bands as our RGB channels respectively. In order to produce images corresponding to these bands in the local universe, we select targets in five redshift ranges such that there is an overlap greater than 50% in three of our observed filters with rest-frame gri (from the Sloan Digital Sky Survey (SDSS); York et al. 2000) bands. These redshift ranges span 0.8 < z < 5.4 and are summarized in Table 1.

Table 1. The Redshift Ranges for Which >50% Overlap Exists between Rest-frame SDSS gri Filters and JWST NIRCam Filters

Redshift Range	Bands Used (RGB)	80% Comp.	# Sources	Morph. Frac.
		$(\mathrm{log}{M}_{* }/{M}_{\odot })$		pec/ell/disk (%)
0.8–1.1	F090W, F115W, F150W	9.4	116	37/18/46
1.4–1.7	F115W, F150W, F200W	10.1	223	52/17/32
2.4–2.6	F150W, F200W, F277W	9.4	22	45/20/35
3.4–3.7	F200W, F277W, F356W	9.0	34	62/28/9
4.4–5.4	F277W, F356W, F444W	9.5 a	39	61/30/9
Total			434

Notes. The number of sources in our final sample in each redshift range is also indicated, along with the morphological fractions as percentages for peculiar/elliptical/disky. We also show the stellar mass 80% completeness threshold, in units of $\mathrm{log}{M}_{* }/{M}_{\odot }$, estimated by comparing to the original catalog in the same redshift range.

^a In this bin we use 50% completeness due to low number statistics.

Download table as:  ASCIITypeset image

After filtering our catalog to photometric redshifts in these ranges, we assemble a total of 1318 sources for further study. Next, we discard sources that have masked-out pixels inside a 100 × 100 pixel postage stamp, and sources with noise in any of the gri-analog filters, which is ≥3 times the mean for that band. The latter filters out sources at the edge of the detector. We inspected a subset of the remaining images, and found that the following criteria produced a sample of galaxies with sufficient size and brightness to have discernible morphology. We first discard sources where the F444W segmentation map contains fewer than 400 pixels. Then, we limit our sample to sources brighter than mag <27 in F444W and <26 in the analog i band. Applying these cuts results in a sample of 434 sources. The median effective radius of discarded sources is 5.7 pixels (018). The median integrated S/N in the bands used for morphological classification was ∼60. Although only a small portion of the full GLASS catalog, this sample size is large enough to provide the first characterization of the morphologies of the 1 < z < 5 galaxy population. Finally, we obtain stellar mass estimates using the FAST++ code (Schreiber et al. 2018) as outlined in Paper XVI. We use Bruzual & Charlot (2003) stellar population models with a Chabrier (2003) initial mass function, a truncated star formation history (SFH) with a constant and an exponentially declining SFH component, and a Calzetti et al. (2000) dust law. Associated errors on stellar masses obtained using SED fitting are estimated to be ∼0.3 dex (Conroy 2013), which can be higher for fainter m_F150W > 28 sources at z > 2. Using these masses we determine the 80% stellar mass completeness threshold, listed in Table 1: for the range 4.4 < z < 5.4 the sample is 50% complete above M_* = 9.52.

We also assemble two reference samples. First, a catalog of 1000 local galaxies from SDSS DR17 (Blanton et al. 2017), randomly selected from those with spectroscopic redshifts 0.01 < z < 0.03, i-band magnitude <22, and stellar masses between 10⁸ and 10¹¹ M_⊙, which we use as a local baseline for the nonparametric morphological analysis. Second, a catalog of 369 sources within our redshift ranges taken from the ZFOURGE (Straatman et al. 2016) survey that lie in the Chandra Deep Field-South (Xue et al. 2011), along with corresponding F814W HST imaging. We use these galaxies to compare our sample to high-redshift rest-frame UV observations.

In order to produce color images, we must first match the angular resolution of different NIRCam filters. We obtain PSFs using the WebbPSF package provided by STScI. We then calculate a matched kernel for each combination of bands in Table 1, using the PhotUtils (Bradley et al. 2020) python package, such that the bluest and middle bands are PSF-matched to the reddest band. We find that for the three filters blueward of F200W, creating a matching kernel of sufficient quality to produce few visible artifacts was difficult, so in creating RGB images we have used non-PSF-matched imaging for these filters. Redward of F200W we have used PSF-matched imaging.

We construct RGB images using the filters in Table 1 according to the Lupton et al. (2004) and Trilogy (Coe et al. 2012) algorithms. To create the images in Figure 1, we used the Lupton HumVI code (Marshall et al. 2015) with the parameters Q = 1.1, α = 412, and the gri bands weighted as follows: (0.9, 1.0, 1.3). For visual inspection (see Section 2.4) we examine RGB images using four different sets of scaling parameters, with differing levels of dynamic range: the Lupton scaling, two logarithmic scalings with different noise luminosity (Trilogy), and a linear scaling.

Zoom In Zoom Out Reset image size

Eight authors (C.J., K.G., A.C., T.T., T.J., R.A., M.M., D.M.) inspected RGB images of each of the 434 selected NIRCam sources and classified sufficiently resolved sources (see Section 2.2) into four broad categories: peculiar (no clear regular shape), elliptical/S0 (spherical or oblate spheroid—greater than PSF), disk (face-on or clear, symmetric disky extended shape), and disk+bulge (clear central concentration detectable). Where the classification was too uncertain due to magnitude or resolution, the source was flagged as such. Sources were examined as RGB images with four different scalings as well as monochrome images in all seven bands using the MorphRater ¹⁹ software package designed for this purpose. Discarding sources with more than one flag left 388 sources, which forms our final sample for the analysis in this paper. In the following, each source is assigned to the category with the highest number of classifications. Across our sample, the median level of agreement was high—5/8 of the inspectors. This good agreement is likely due to the broad categories used as well as the cuts made to ensure that sources were sufficiently resolved to make a good determination.

For each of our sources we automatically calculate values of the Gini coefficient (G) and asymmetry (A) parameters from the rest-frame i-band images, following the formulation of Abraham et al. (1996b) and Abraham et al. (2007). The Gini coefficient is high in concentrated and symmetrical early-type galaxies, and is more robust against isophote thresholds than the standard concentration parameter (Abraham et al. 2007), but can also be high in asymmetrical late-type galaxies dominated by a few bright clumps. The asymmetry parameter A is determined by calculating the residual difference of a galaxy with its own rotated image, along with a noise correction (per Abraham et al. 1996a; Lotz et al. 2004). A represents the ratio of light remaining to the original object flux, thus a perfectly symmetric source would have A = 0. The axis of rotation is the brightest pixel within a 40 × 40 pixel box around the center of each source as identified by Source Extractor. These measurements, along with, e.g., clumpiness (Conselice 2003), multiplicity (Law et al. 2007), and others (e.g., Freeman et al. 2013) are well established in the literature and have been used previously to probe the relationship between optical/UV morphologies in HST (Lee et al. 2013; Mager et al. 2018) and the evolution with redshift (e.g, Whitney et al. 2021).

In Figure 1 we present a sample of sources from across the selected redshift ranges and mass bins. These sources are typical of the overall sample for those with a discernible morphology. (Several gaps in the figure were filled by extending the redshift bins by up to 0.1; these are indicated by lack of morphological designation.) As expected, at 0.8 < z < 5.4, we find a significantly higher proportion of peculiar morphologies than in the local universe, where peculiars make up less than 5% of galaxies (Murata et al. 2014). Peculiars make up 42% of the sample at z < 2, and 54% at higher redshifts. We do not see a significant increase with redshift, but this may be due to selection effects, as higher masses are overrepresented in the higher-redshift sample. The fraction of peculiar galaxies is 50% at stellar masses <10^9.5 M_⊙, and 24% at higher mass.

The presence of visible disks (both edge- and face-on) is detected in many galaxies, and we see clear spiral structure up to z = 2.4. Many of the disks show clear bulges, again up to z = 2.4 As can be seen in Figure 1, there are sources of unambiguous typical Sa/Sb morphology beyond redshift 2. Several sources additionally show distinct bars.

For reference, Figure 2 shows several examples of local disk galaxies (in gri RGB images) artificially redshifted to three of our redshift bands, as they would appear in NIRCam images. The effects of K-correction and bandpass shifting have not been considered for the purposes of this illustration; the former as it is a small effect for these bright sources (Barden et al. 2008), and the latter as our methodology relies on the good agreement between the bandpasses of the rest-frame SDSS bands and their high-z NIRCAM analogs. The scale and colors of these local galaxies can be compared to sources at similar redshifts in Figure 1. While these images are larger (here 8''; see 3'' stamps in Figure 1), many of the sources we observe at z > 2 appear morphologically similar to the local SDSS galaxies although with typically higher surface brightness.

Zoom In Zoom Out Reset image size

Overall, we see strong clues that the Hubble sequence is well developed by z = 1.5. The current generation of cosmological hydrodynamic simulations, such as TNG50 (the highest-resolution run of the IllustrisTNG suite), do predict a significant proportion of disks in star-forming galaxies at high redshifts, for instance ∼30% at z ≃ 3 and ∼50% at z ≃ 2 for M ≳ 10^9.5 M_⊙ (Pillepich et al. 2019, see their Figure 9), and moreover predict a population of well-developed spirals with bulges at z = 2 visible in rest-frame optical (although this may be sensitive to decomposition morphology; see Zana et al. 2022). As additional JWST imaging data becomes available over wider fields, we will be better able to empirically confirm (or challenge) these predictions.

In some aspects, this work is in tension with HST studies. Mortlock et al. (2013) found few disks among massive (M_* > 10¹⁰ M_⊙) galaxies above z > 2 with peculiars dominating; we find 40% disks in this mass range at z > 2. Whitney et al. (2021) find more asymmetry and higher concentration for a given mass; we find a weak trend only. Huertas-Company et al. (2015), who examine morphological evolution in CANDELS galaxies, find irregulars dominant for 2 < z < 3, with regular disks/disks and bulges dominating below z = 2; in our higher-mass bin, we find disks dominate by z = 3 and peculiars are ∼10% by z = 2. Tacchella et al. (2015) find preponderance of disk structures in 29 z ∼ 2 galaxies, with 40% sporting large (B/T > 0.3) bulges, perhaps closer to our finding but for a much smaller sample.

Otherwise, several studies predict our finding that rest-frame UV-limited observations will lead to clumpier classifications. Lee et al. (2013) perform a nonparametric analysis that quantifies this effect. Kartaltepe et al. (2015) report that galaxies observed in the UV may appear irregular when in reality they are normal disks in the optical. Mager et al. (2018) also perform CAS analysis and find galaxies appearing more symmetric (early type) at longer wavelengths for GALEX (Bianchi & GALEX Team 1999) UV observations.

Although our sample includes some merging and interacting galaxies, we have not controlled for environment in this analysis. We have not considered whether a source may be interacting with neighbors, and whether this affects the morphologies due to features such as tidal tails.

We now consider evolutionary trends with stellar mass and redshift revealed by our sample. Because we expect uncertainties of ∼0.3 dex on our estimated stellar masses (Section 2.2), we bin our sources into six stellar mass ranges of 0.5 dex width as shown in Figure 1 (from M_* < 10^8.5 M_⊙ to >10¹¹ M_⊙). We find an abundance of clumpy/peculiar galaxies across all mass bins. Bulges and disks are visible across the mass range M_* > 10^8.5 M_⊙, and a significant population of spirals with bulges is apparent at M_* > 10^9.5 M_⊙.

In Figure 3 we depict the broad morphological classification as a fraction of the sample by (photometric) redshift and mass bins. Here we divide the sample into two mass bins (limited by our sample size), greater or less than 10^9.5 M_⊙. The redshift bins correspond to the same ranges chosen for their overlap with rest-frame optical filters (Section 2.2). The blue points represent the combined contribution of "disk" and "disk+bulge" morphologies, while the "disk+bulge" fraction is also shown separately with dashed lines. We see that by z ≃ 2.4 there is a significant proportion of disky galaxies in our sample (∼20% even for the low-mass bin, and the vast majority in the high-mass end), and at higher masses the majority have a recognizable bulge. The proportion of noticeable Sc/SBcs increases below redshift 2.4 as can be seen in Figure 1. Counterintuitively, the proportion of peculiar morphologies for low-mass galaxies shows only a weak downwards evolution with decreasing redshift; previous work has suggested that by z ∼ 1 the irregular number and mass fraction is dropping rapidly (e.g., Oesch et al. 2010). A larger sample size at high redshift may resolve this apparent anomaly.

Zoom In Zoom Out Reset image size

At all redshifts and mass ranges, our sample is significantly bluer in the optical than local galaxies. The mean g − i color of the 388 sources in our morphological catalog is 0.22 (with scatter σ = 0.36), whereas the z ∼ 0 comparison sample from SDSS has a mean g − i = 0.80 (σ = 0.37). Galaxies in our highest-redshift bins (3.36–5.43) are bluest, with a mean rest-frame g − i of −0.12 (σ = 0.35). This is consistent with the large UV slopes seen at high redshifts (e.g., see Paper XVI).

The results of the Gini–asymmetry (GA) analysis are shown in Figure 4. Here we show the sample of this paper (color coded by visual classification), the SDSS local comparison sample (top panel, in gray), and our HST F814W sample (bottom panel, in orange). Histograms of the statistics are also shown for the three samples. The JWST GA diagram shows a remarkable agreement with the visual morphologies—disks and ellipticals have low A, ellipticals have high G, and peculiar galaxies have high A.

Zoom In Zoom Out Reset image size

The comparison with SDSS shows similar distributions, with the JWST sample being skewed to both higher A (0.27 ± 0.01 versus 0.19 ± 0.01 for SDSS) and higher G values (0.49 ± 0.01 versus 0.45 ± 0.01 for SDSS). We interpret this difference as arising from the overall younger age of the high-redshift population, with both coefficients being driven up the prevalence of intense knots of star formation. Naïvely, we might have expected more separation between the JWST and SDSS populations. However, as can be seen from the images in Figure 1, many galaxies in our sample have close morphological analogs in the local universe.

Compared to the HST rest-frame UV sample, our sample is significantly more symmetrical (mean A = 0.26 ± 0.01 versus HST 0.41 ± 0.01) and with lower G (mean 0.51 ± 0.01 versus HST 0.64 ± 0.01). The increased depth of the NIRCam imaging needs to be considered, as we are probing significantly lower masses at each redshift range compared to HST. In the bottom panel of Figure 4 we show the difference when the sample is mass-matched to the HST range. In the five redshift ranges, the HST mass limits are 8.8, 9.4, 9.5, 9.9, and 10.4 $\mathrm{log}M/{M}_{\odot }$ respectively. The mean Gini coefficient for the NIRCam subsample (N = 67) is 0.57 ± 0.01, with mean A lower at 0.22 ± 0.02.

The HST sample shows many galaxies with high values of G and/or A. Inspection of the F814W images show these are star-forming clumpy systems and they have blue colors. The clumpiness in the rest-UV gives rise to high G values, and they are often highly asymmetrical. In contrast the NIRCam sample is much smoother (giving rise to the lower G) and more symmetrical (lower A). We interpret this as an effect of the rest-frame optical being a much better tracer of the stellar mass, compared to the rest-UV, which traces low mass-to-light ratio star-forming complexes. In summary, as we move from the local universe to the high-redshift regime, our sources have higher Gini in the optical and less symmetry, due to their evolutionary phase. As we move from the optical to the UV, the sources have even higher Gini and asymmetry as we lose sight of extended structure and are biased toward seeing more actively star-forming regions.

Our results are in broad concordance with other recent JWST analyses. Both Nelson et al. (2022) and Fudamoto et al. (2022) find an unexpected population of very red disks at higher redshifts revealed in NIRCam data (several examples are visible in Figure 1). Ferreira et al. (2022) similarly find a high proportion of disks that dominate at z > 1.5.

Finally, we note that our results show that using rest-frame optical GA values to automate morphology classification is very promising. The values show a much clearer mapping to visual morphologies compared to a similar analysis of the rest-frame UV.