The GLASS-JWST Early Release Science Program. II. Stage I Release of NIRCam Imaging and Catalogs in the Abell 2744 Region

Image prereduction was executed using the official JWST calibration pipeline, provided by the Space Telescope Science Institute (STScI) as a Python software suite. ²¹ We adopted version 1.8.2 (Bushouse et al. 2022) of the pipeline and versions between jwst_1014.pmap and jwst_1019.pmap of the CRDS files (the only change between these versions is the astrometric calibration, which is dealt with as described below). We executed the first two stages of the pipeline (i.e., calwebb_detector1 and calwebb_image2), adopting the optimized parameters for the NIRCam imaging mode, which convert single-detector raw images into photometrically calibrated images.

Using the first pipeline stage calwebb_detector1, we processed the raw uncalibrated data (uncal.fits) in order to apply detector-level corrections performed on a group-by-group basis. These include dark subtractions, reference pixel corrections, nonlinearity corrections, and jump detection, which allows us to identify cosmic-ray (CR) events in the single groups. The last step of this pipeline stage allows us to derive the mean countrate in units of counts per second for each pixel by performing a linear fit to the data in the input image (the so-called ramp-fitting), excluding the group that is masked due to the identification of a CR jump.

The output files of the previous steps (rate.fits) are processed through the second pipeline stage calwebb_image2, which consists of additional instrument-level and observing-mode corrections and calibrations such as the geometric-distortion correction, the flat-fielding, and the photometric calibrations, which convert the data from units of countrate into surface brightness (i.e., megajanskys per steradian) and generate a fully calibrated exposure (cal.fits).

The cal.fits file also contains an rms extension, which combines the contribution of all pixel noise sources, and a DQ mask where the first bit (DO_NOT_USE) identifies pixels that should not be used during the resampling phase.

We then applied a number of custom procedures to remove instrumental defects that are not dealt with the STScI pipeline. Some of them have already been adopted in Merlin et al. (2022, hereafter M22) and are described there: we illustrate below only the main changes to the STScI pipeline in the default configuration and/or to the procedure adopted in M22.

• 1.  
  "Snowballs," i.e., circular artifacts observed in the in-flight data caused by a large CR impacts. These hits leave a bright ring-shaped defect in the image because the affected pixels are just partially identified and masked. In M22, we developed a technique to fully mask out these features, which was not necessary here. Indeed, version 1.8.1 of the JWST pipeline introduced the option to identify snowball events, expanding the typical masking area to include all the affected pixels. This new implementation provides the opportunity to correct for these artifacts directly at the ramp-fitting stage, at the cost of a higher noise on the corresponding pixels. We enabled this nondefault option and fine-tuned the corresponding parameters to completely mask all the observed snowballs and at the same time, minimize the size of high-noise areas.
• 2.  
  "NL Mask": we find groups of deviant bright pixels on the cal images taken with the NIRCam Module B LW detector, more evident on deeper pointings. They are not well corrected during the prereduction stage and are identified as "WELL_NOT_DEFINED" pixels ²² in the NonLinearity Calibration file. ²³ We recognize them by their flag in the DQ and mark them as DO_NOT_USE for the coaddition.
• 3.  
  1/f noise, which introduces random vertical and horizontal stripes in the images (see Schlawin et al. 2020). We removed this by subtracting the median value from each line/column. To remove the flux from objects as accurately as possible, we masked out all objects and the bad pixels flagged in the data quality. The masks were obtained from the segmentation maps obtained with SExtractor (version 2.25.0; Bertin & Arnouts 1996), and then they were further dilated in order to exclude the contamination from the faint outskirt of the objects, which escape detection below the SExtractor threshold. We have applied a differential procedure to dilate objects depending on their ISOAREA: the segmentation of objects with ISOAREA < 5000 pixels was dilated using a 3 × 3 convolution kernel and a dilation of 15 pixels, while for the segmentation of objects with ISOAREA ≥5000 pixels, a 9 × 9 convolution kernel and a dilation of 4 × 15 pixels was used. The procedure was executed separately for each amplifier in the SW detectors (i.e., four times for each individual image) with the exception of the denser areas corresponding to the centers of the clusters and the brightest field star, where objects are significantly larger than the amplifier width (500 pixels, corresponding to about 30'') and could not be masked efficiently. In this case, we removed the 1/f noise over the entire row. Our procedure, which was already adopted in the first release of the GLASS data (M22), is conceptually similar to the one adopted for the CEERS data (Bagley et al. 2022). We publicly release the code adopted for this step.
• 4.  
  Scattered light: we identify additive features in the F115W, F150W, and F200W images. These low-surface brightness features have already been revealed by commissioning data (see Rigby et al. 2023) and are due to scattered light entering the optical path. These anomalies have been dubbed wisps or claws, depending on their origin and morphology. Wisps have a nearly constant shape and can thus be subtracted from the images with simple templates. We removed these features by extracting their 2D profile from the available template (we do not use the entire template image to avoid subtracting its empty but noisy regions) and then normalizing the residual template to match the feature intensity in each image. Claws have first been identified and singled out in images. Their shape in each image has been reconstructed by interpolating a 2D mesh with a box size of 32 pixels and then eventually subtracted from the same image. These procedures efficiently remove most of these features, as shown in Figure 2.We found other defects in the F090W image, and to a lesser extent in the F115W, in images taken in 2022 June. These defects consist of additional scattered light in the images, resulting in artificial sources along the field of view, and they are due to a so-called wing-tilt event, i.e., a small shift of one of the wings of the primary mirror. We adopted the procedure described in M22 to identify and mask these artifacts from images. We emphasize that they do not affect the images taken in 2022 November.

Zoom In Zoom Out Reset image size

We then rescaled the single exposures to units of microjanskys per pixel using the conversion factors output by the pipeline.

We also note that our procedure to remove the 1/f noise and the background effectively removes the intracluster light (ICL) from the images. We caution the user to avoid using these images to study the ICL in detail.

The astrometric calibration was performed using SCAMP (Bertin 2006), with third-order distortion corrections. Compared to the procedure we adopted in M22, we started from the distortion coefficient computed by the STScI pipeline, stored in the cal images. We refined the astrometric solution by running scamp in cal mode, which optimizes the solution with limited variations from the starting solution. We have found this procedure both accurate and reliable, as described below. We first obtained a global astrometric solution for the F444W image, which is usually the deepest. As not enough Gaia Data Release 3 (DR3) stars can be used for every NIRCam detector, we have aligned the images to a ground-based catalog obtained in the i-band with the Magellan telescope in good seeing condition (see T22 for details) of the same region, which had been previously aligned to Gaia DR3 stars (Gaia Collaboration 2016). We then took the resulting high-resolution catalog in F444W as reference for the other JWST bands, using compact, isolated sources detected at high signal-to-noise ratio (S/N) at all wavelengths. Each NIRCam detector has been analyzed independently in order to simplify the treatment of distortions and minimize the offsets of the sources in different exposures. Finally, we used SWarp (Bertin et al. 2002) to combine the single exposures into mosaics projected onto a common aligned grid of pixels, and SExtractor to further clean the images by subtracting the residual sky background. The pixel scale of all the images was set to 0031 (the approximate native value of the short-wavelength bands) to allow for simple processing with photometric algorithms.

The final image, computed as a weighted stack of all the images from the three programs, has a size of 24,397 × 21,040 pixels, corresponding to 12.6 × 10.87 arcmin². In this frame, the area covered by the wide-band NIRCam images (F115W, F150W, F200W, F277W, F356W, and F444W) is exactly 46.5 arcmin².

Given the especially deep and sharp nature of the JWST images, where most of the faint objects have sizes below 05, the requirements on the final astrometric accuracy are extremely tight to avoid errors in the multiband photometry (where a displacement of as little as 01 can bias the color estimates). These requirements must also be met in the overlapping regions of the various surveys, which have often been observed with different detectors.

To verify the final astrometric solution, we conducted a number of validation tests, where we compare the positions of cross-matched objects in catalogs extracted from different images. For each of these catalogs, we used SExtractor in single-image mode and adopted the XWIN and YWIN estimators of the object center, which are more accurate than other choices. Given the unprecedented image quality of NIRCam, the center of extragalactic objects with complex morphology may be difficult to estimate with high accuracy, particularly when observed across a large wavelength interval. Therefore, we only compared objects with well-defined positions to minimize errors, using the ΔX, ΔY = ERRAWIN_WORLD, and ERRBWIN_WORLD estimators of the error and limited the analysis to objects with ${({\rm{\Delta }}{X}^{2}+{\rm{\Delta }}{Y}^{2})}^{1/2}\leqslant 0\buildrel{\prime\prime}\over{.} 018$. From these catalogs, we estimated both the average offset of the object centers $\overline{{\rm{\Delta }}\alpha }$ and $\overline{{\rm{\Delta }}\delta }$ and the median average deviation mad_α and mad_δ , which measure the intrinsic scatter in the alignment. In Figure 3 we report the main outcome of these tests:

• 1.  
  We first compared the positions of objects in the original Magellan i-band and the resulting F444W of the entire mosaic (left panel). We find an essentially zero offset and mad_α ≃ mad_δ ≃ 002, which is two-thirds of a pixel.
• 2.  
  We compared the F444W catalog with the catalog released by Kokorev et al. (2022) in the context of the ALMA lensing cluster survey (ALCS), containing Hubble Space Telescope (HST) and IRAC sources in the A2744 region (middle-left panel). Overall, the comparison results in a good alignment with a very small offset ($\overline{{\rm{\Delta }}\alpha }\simeq 1\mathrm{mas}$ and $\overline{{\rm{\Delta }}\delta }\simeq 2\mathrm{mas}$) and mad_α ≃mad_δ ≃ 002.
• 3.  
  We compared the F444W catalog with the AstroDeep H₁₆₀ catalog obtained on the central region of the A2744 cluster (middle panel), as obtained in the context of the Frontier Fields initiative (Merlin et al. 2016a). While the intrinsic scatter is still good (mad_α ≃ mad_δ ≃ 002), we find a systematic offset by about 1 pixel in R.A. and 2.5 pixels in decl., which is most likely due to different choices in the absolute calibration of the ACS/WFC3 data released within the Frontier Fields.
• 4.  
  We compare here the relative calibration of filters at the two extremes of the spectral range, F444W and F115W, where morphological variations and color terms may change the center position and affect the astrometric procedure (middle-right panel). We find again very good alignment with a negligible offset and small mad_α ≃mad_δ ≃ 001.
• 5.  
  Finally, we compare the astrometric solutions in the overlapping areas by independently summing the data of the three different programs and checking the accuracy in the overlapping area (right panel). Again, we find a very good alignment with a negligible offset and small mad_α ≃ mad_δ ≃ 001.

Zoom In Zoom Out Reset image size

We adopted a cross-matching radius r = 01 in most of the validation tests, except in the comparison with the AstroDeep catalog, for which we adopted a much larger radius (r = 04), given the larger offset at the level of 0075. We also note that the scatter in δ seems consistently lower than that in α, but we failed to identify a clear origin for this effect, which does not impact the global accuracy. We therefore conclude that the astrometric procedure is accurate and adequate for the goals of this Stage I release. In the future, we plan to explore further and validate other options for astrometric registration and also release images with a smaller pixel scale to better exploit the unprecedented image quality of the JWST data. However, we note that the GLASS-JWST data have a very limited dithering pattern (which was driven by spectroscopic requirements) and so may benefit only marginally from moving to smaller pixels.

Since the detection band has the coarsest resolution, we performed a PSF match for all the other NIRCam images for color fidelity. We created convolution kernels using the WebbPSF models publicly provided by STScI, ²⁴ combining them with a Wiener filtering algorithm based on the one described in Boucaud et al. (2016). To smooth the images, we used a customized version of the convolution module in t-phot (Merlin et al. 2015, 2016b), which uses FFTW3 libraries. However, we note that this approach cannot fully correct for the inhomogeneities of the PSF: the calibration upon which WebbPSF is calibrated is inevitably initial, and the JWST PSF depends on time and position (Nardiello et al. 2022), and our data set is the inhomogeneous combination of data obtained at different times and with different PAs, so that the PSF definitely changes over the field. For this version of the catalog, we used the UNCOVER PSF models (epoch: 2022/11/07, PA: 41.2 deg) as average PSFs, and we plan to improve our PSF estimation in future versions of the catalog that will be released in Stage II.

Similarly, concerning the HST images, we note that all of them have too few stars to obtain a robust estimate of the PSF directly from the images, so we adopt in all cases existing HST PSFs that were taken from CANDELS. This approximation may introduce small biases in the final catalog. ACS images have been PSF-matched to F444W, while for the WFC3 F105W, F125W, F140W, and F160W images, which have a PSF larger than that of F444W, we have done the inverse, that is, we smoothed the F444W image and the WFC3 F105W, F125W, and F140W to the F160W and followed a slightly different procedure that we describe below.

The total flux is measured with a-phot on the detection image F444W by means of a Kron elliptical aperture (Kron 1980). As we have shown in M22, simulations suggest that Kron fluxes measured with a-phot are are somewhat less affected by systematic errors, while being slightly more noisy.

Then, we used a-phot to measure the fluxes at the positions of the detected sources on the PSF-matched images, masking neighboring objects using the SExtractor segmentation map. Given the wide range of magnitudes and sizes of the target galaxies, we measured the flux in a range of apertures: the segmentation area (the images being on the same grid and PSF-matched) and five circular apertures with diameters that are integer multiples (2, 3, 8, and 16 times) of the FWHM in the F444W band, which correspond to 028, 042, 112, and 224 diameters. For the four WFC3 images (which have a PSF larger than that of F444W), we first filtered the F444W to their FWHM and then measured the colors between the filtered F444W and the WFC3 images. To minimize biases when these colors are combined with those of the other bands, we use in this case apertures the same multiples of the WFC3 PSF adopted for the other bands. We remark that this procedure is only approximate and delivers a first-order correction of the systematic effects due to different PSFs. In a future release, we plan to adopt more sophisticated approaches to optimize the photometry, including but not limited to the improvement of the PSF estimate and applying t-phot on WFC3 images that have a larger PSF.

Total fluxes are obtained in the other bands by normalizing the colors in a given aperture to the F444W total flux, i.e., by computing f_m,total = f_m,aper/f_F444W,aper × f_F444W,total, as described in M22.

We release the five catalogs described above (one computed on segmentation, and four computed on the different apertures), and we leave it to the user to choose the most suitable catalog for a given science application. In general, small-aperture catalogs are more appropriate for faint sources as they match their small sizes and minimize the effect of contamination from nearby sources, which can be high for blended objects. Larger apertures may be more appropriate for brighter sources and especially cluster members.

We have performed a few validation tests to primarily verify the flux calibration, which has been the subject of many revisions in these first months, and to a lesser extent, the procedure adopted to derive the photometric catalog.

The overlap between the southern quadrants in GLASS1 and GLASS2 offers a nice opportunity to test the NIRCam flux calibration. Indeed, the two GLASS observations have been observed in two epochs (2022 June and November) with a PA difference of nearly 150°. As a result, the southern quadrants of GLASS1 and GLASS2 largely overlap, but have been observed with modules B and A, respectively. We therefore obtained stacked images of the two epochs separately, built a photometric catalog with the same recipes, and checked the magnitude difference between objects observed with different detectors. The result of this exercise, done on all bands, is reported in Figure 6. We note that in the short bands, the two modules are made of four detectors, each with an independent calibration, which we plot all together in Figure 6. The comparison, that is, limited to objects observed with high S/N > 25, shows that the average magnitude difference between the two modules is in general quite small, and is below 0.05 mag in all cases (see Figure 6 and its captions for details). This confirms that the flux calibration between the different modules is reasonably stable at this stage.

Zoom In Zoom Out Reset image size

As a further check to validate the photometric pipeline, we compared the colors of the sources in the region of the A2744 cluster with those measured for the Frontier Fields survey (Lotz et al. 2014) within the AstroDeep project (Merlin et al. 2016a; Castellano et al. 2016). We chose to compare the colors of sources to avoid possible systematics deriving from the total flux estimates obtained on two different bands in the two catalogs; some residual offsets and trends can be due to the different segmented areas. The cross match was performed by selecting objects having m_H160 < 24, assuming a positional accuracy Δr < 04 to account for the possible mismatch in absolute astrometry between the two. This comparison is shown in Figure 7, where we show isophotal colors computed on PSF-matched images as a function of the H160 magnitude of the Astrodeep catalog, also providing running and global median offsets. We note that in Merlin et al. (2016a), we explicitly modeled and subtracted the ICL and the brightest cluster sources. This was especially needed to derive accurate colors on Spitzer images, which have a much poorer resolution. In this case, our procedure to remove the 1/f noise, and the background effectively removes a significant fraction of the ICL, to the point that objects falling in these areas do not show any systematic shift of the objects in Figure 7. Therefore, the removal of the brightest cluster members is postponed to the release of Stage II. We also note that for the same reason, our images are not adequate for studying the ICL emission. The comparison shows that when the same approach is used to estimate colors, i.e., isophotal magnitudes are adopted, the agreement between the two catalogs is good (although the shallowness and low resolution of the IRAC bands makes the comparison less accurate).

Zoom In Zoom Out Reset image size

From this comparison, we conclude that—quite reassuringly—the overall photometric chain is consistent between the well-established Frontier Fields data and these new data. At the same time, we remark that the optimal choice of the aperture depends on the size and type of object under study. Small apertures tend to have higher S/N and should be preferred for faint sources. For the brightest sources, larger apertures should be preferred. It is also possible to estimate rough color gradients by comparing the various apertures that we release. We also tested that applying the same technique without PSF matching introduces an offset of ∼0.2 mag in the final colors, which would clearly affect the derived photometric redshifts and spectral engery distribution fitting results.

Finally, in an effort to cross-validate our results prior to release, in the leadup to this paper, we compared our catalogs to those under development by the UNCOVER team (Weaver et al. 2023, in preparation) based on the same raw data sets. The image-processing and photometric procedures adopted by the two teams have significant differences. The main differences are (i) the image coaddition (the UNCOVER team adopts Grizli from Brammer 2019, while we use a custom pipeline that uses scamp and swarp); (ii) the object detection (UNCOVER uses an optimally stacked F277W+F356W+F444W image after removing the ICL, while we use F444W); and (iii) techniques and tools for PSF matching and photometry. For these reasons, we found some differences between the catalogs, especially for faint sources at the detection limit, as expected. However, our comparison of working versions of the catalogs produced by the two teams shows a good agreement in the colors and magnitudes of the vast majority of objects overall, as shown in Figure 8 in the Appendix.