CEERS Epoch 1 NIRCam Imaging: Reduction Methods and Simulations Enabling Early JWST Science Results

"Snowballs" are large cosmic-ray events that can affect hundreds of pixels and have a circular morphology on the NIRCam detectors. We see an average of ∼25–30 snowballs per detector in a ∼900 s exposure. The fluence of counts in the center of the snowballs is very high and can sometimes saturate the detector. The cosmic-ray flagging step of jwst often successfully identifies and flags the central cores of these large cosmic rays, leaving the more diffuse wings of the snowballs present in the count-rate maps output by Stage 1. The top left panel of Figure 2 shows an example of a count-rate map containing multiple snowballs that were only partially removed by the pipeline.

Zoom In Zoom Out Reset image size

We identify snowballs as large contiguous sets of pixels in the GROUPDQ arrays that have been flagged as jumps (JUMP_DET). We found it helpful to divide the snowballs into two tiers: large ones, which require a rather large mask to address the extended wings, and smaller ones, which do not require masking as much additional area. In order to separate snowballs from smaller cosmic-ray impacts, we median-filter the GROUPDQ arrays with separate two-dimensional top-hat kernels with radii of 7 and 15 pixels to identify the smaller and larger snowballs, respectively. This filtering recovers extended groups of pixels that have an approximately circular footprint. We also include in the snowball mask any saturated pixels that are within these big groups. We then grow the resulting snowball footprints via binary dilation and a two-stage top-hat growing kernel with radii of 7 and 35 pixels. We add this updated snowball mask to the GROUPDQ array and run the ramp-fitting step from jwst Stage 1. The flux in the affected pixels is determined by the remainder of the ramp, using the slope of the ramp excluding the newly identified cosmic-ray jump. The top middle panel of Figure 2 shows the DQ array for the count-rate map in the left panel, with the original footprints of the snowballs identified with red contours. The top right panel shows the output count-rate map that results from performing this snowball correction. In the bottom row, we show the portion of the mosaic that corresponds to the region displayed in the top row to demonstrate the effectiveness of this correction.

The snowball correction is performed with our Python script snowball_wrapper.py, which runs Stage 1 of the pipeline, saves the ramps from the first run of the ramp-fitting step, identifies the snowballs and grows their footprint, flags them as cosmic rays, and then runs the ramp-fitting step again to create a count-rate map that excludes the flagged portions of the ramps.

We note that persistence from the saturated cores of very bright snowballs can show up in subsequent exposures. These are faint enough not to be rejected when the dithered images are combined. They can show up as single-band detections that look a lot like faint emission-line galaxies. For pure emission-line sources in WFSS exposures, it is important to inspect the individual exposures and their DQ arrays.

We next subtract the "wisp" features from the F150W and F200W images. Wisps are created from stray light reflected off the secondary mirror supports. They are visible on detectors A3, A4, B3, and B4 and are most prominent in F150W and F200W. The strength of the wisp features depends on the source of the reflected light, and so the wisps can have a variable brightness from exposure to exposure. The top left panel of Figure 3 shows a version of the F200W mosaic for CEERS NIRCam pointing 1 from which we have not subtracted the wisps as an example of the large-scale structure of this feature. They extend across two detectors in each module.

Zoom In Zoom Out Reset image size

The NIRCam team has provided wisp templates constructed from images obtained by several NIRCam commissioning and ERS programs observed early in Cycle 1. At the time of this writing, two sets of templates have been released, in 2022 July and August. Both sets of templates are available on the JWST User Documentation page on Claws and Wisps. ³⁶ We use the updated wisp templates released on 2022 August 26, which are the result of reprocessing the original observations with updated processing steps and reference files, including a correction for the large-scale variations in the ground flats consistent with that introduced by pmap 0956.

We scale the templates to account for the variable brightness of the wisp feature and subtract them from the images in the following way. We first apply the flat field to each image to match the flat-fielded templates and to measure the wisp feature in the flattened images. For each image, we perform a very "cold" source detection (i.e., a very high detection threshold of 5.5σ) with Photutils (Bradley et al. 2020) to identify and mask out large, bright sources in the image while avoiding the wisp feature itself. Next, we smooth the wisp template with a Gaussian kernel with σ = 2 pixels. For an array of coefficients a, we find the value that minimizes ${\sigma }_{\mathrm{MAD}}^{2}(I-{aW})$, where σ_MAD is the median absolute deviation, I is the flat-fielded masked image, and W is the smoothed wisp template. We find that in almost all cases we need to scale the wisp template by a factor <1. We scale the original (unsmoothed) wisp template by a and subtract it from the original (un-flat-fielded and unmasked) image. We illustrate this process in Figure 3.

While this scaling does well at removing the large-scale feature, our wisp subtraction is still preliminary. The current wisp templates were produced with an updated but still early NIRCam reduction. Additionally, in early Cycle 1 observations, the dithers were often small compared to the size of the sources observed. As a result, some ghost images of bright sources are present in some of the wisp templates and therefore become regions of oversubtraction (by ∼3%–4%) in the CEERS images. We also note that we see some evidence for the presence of weak wisps in F115W, but as there are no templates available for this filter, we leave it uncorrected. The wisp templates will be improved throughout Cycle 1 with additional observations and updates to the NIRCam reduction and reference files.

Our final custom step on the count-rate images is to measure and remove 1/f noise, which is correlated noise introduced in the images when the detectors are read out (Schlawin et al. 2020). The noise presents as horizontal and vertical striping patterns that vary from row to row and column to column. Ideally, this noise would be fit and removed during the reference pixel fitting or the up-the-ramp fitting of jwst, an update that may be included in future versions of the Calibration Pipeline. In the meantime, we perform our own correction for 1/f noise in each individual count-rate map as follows.

As an additive effect, the noise pattern should be removed before applying the multiplicative flat-field correction, yet the striping patterns are best measured on flat images. We therefore apply the flat field to the count-rate maps for pattern measurement, but we subtract it from the original count-rate images. We next mask all bad pixels (with a DQ flag value >0) and source flux that is identified with Photutils using a tiered source detection method. This method consists of convolving the image with progressively smaller kernels and running a segmentation-style source detection at each step. In this way we are able to detect both large, extended sources and small, compact sources with detection parameters optimized for each source size. We use four tiers, with Gaussian kernels of σ = 25, 15, 5, and 2 pixels (on the original 0031 and 0063 pixel⁻¹ scales for the SW and LW channels, respectively). These values were chosen after experimenting with several filter kernels to aggressively mask as much source flux as possible. Next, we calculate a pedestal value for the sky background by fitting a Gaussian to the distribution of pixel values in the masked image.

We measure the striping pattern using a sigma-clipped (2σ) median along first rows and then columns. For the row correction (horizontal striping), we measure and remove the pattern individually for each of the four amplifiers. In some images, especially in the SW filters, the difference from amplifier to amplifier can vary by ∼3%–5%, and so this amplifier-dependent correction works better than using the median from the full row. However, in some cases, especially around bright or extended sources, a large number of pixels in a given amp-row are masked, leaving too few to calculate a robust median. In these cases, we use the median of the entire row for the affected amp-row. The threshold used to determine the cutoff in number of masked pixels per amp-row varies from image to image to allow for the best pattern subtraction. We illustrate this amp-row and column subtraction in Figure 4.

Zoom In Zoom Out Reset image size

This amp-by-amp approach works well for NIRCam imaging, especially for fields like the EGS that are relatively sparsely populated (compared to a globular cluster, star field, or regions with high nebulosity, for example, where a majority of pixels in individual amplifier rows will be masked). However, it is not a good approach for NIRCam WFSS observations, where the length of a trace in the row grism is approximately the same size as an amplifier for some filters.

The observations of CEERS MIRI pointing 2 on 2022 June 21 encountered a problem writing data to disk, and one of the F2100W (with NIRCam F200W+F444W in parallel) image sets was repeated on June 28. See G. Yang et al. (2023, in preparation) for a full discussion of the effect on the MIRI images. The NIRCam images from June 21 were not affected, and the repeated observations offer additional depth in F200W and F444W, though at a slightly different position angle. However, the repeat observations obtained on June 28 appear to suffer from significant persistence on SW detector A3.

We searched MAST for all NIRCam observations obtained in the 24 hr before the NIRCam2 repeat imaging. Program 1022 (PI: J Stansberry) observed Jupiter ∼6 hr before the CEERS imaging. This commissioning program aimed to test the Fine Guidance Sensor (FGS) guiding near a bright planet and to characterize the scattered light in each instrument. There were two observations obtained with Jupiter on SW detector A1 (F212N) and LW detector ALONG (F322W2). In both cases, the same illumination pattern is present on SW detector A3 as that we see in the CEERS images. We show the calibrated (processed through pipeline Stages 1 and 2, flat-fielded and in units of MJy sr⁻¹) A3 detector images from program 1022 in the top two panels of Figure 5. The pattern is fainter in the second image, obtained 20.3 minutes after the first, indicating that the pattern is fading with time as expected for persistence. The feature on the detector A3 images may be scattered light from Jupiter imaged off-detector, fading persistence that occurred while slewing, or fading persistence from an earlier observation. The program observed before 1022, and 8–10 hr before CEERS, is proprietary—GO 2473, PI: L Albert. Program 2473 targeted two brown dwarfs, WISE J053516.80–750024.9 and WISE J073444.02–715744.0, with NIRCam module B. The feature in the CEERS images is on the far side of the A3 detector from module B, and so we do not believe that the persistence was caused by observations from this program. We note that neither the observations of Jupiter on detector A1 nor any of the LW images left persistence in the CEERS images.

Zoom In Zoom Out Reset image size

We create a mask using the two A3 images from program 1022, identifying affected pixels as those with a flux >20 MJy sr⁻¹ in the first exposure and >10 MJy sr⁻¹ in the second exposure. These thresholds were found through trial and error to identify the persistent pattern and avoid flagging noise fluctuations. The bottom right panel of Figure 5 shows the mask we created in this way, with flagged pixels in white. However, we found that this mask was not sufficient around the scattered light pattern. The wings of the pattern were still present in our F200W mosaics, and so we manually masked out a region around the affected area, shown as the red polygon in Figure 5. This method represents a preliminary approach to handling the persistence in this field. In future data reductions, we will create a more careful mask incorporating the CEERS images to identify the fading pattern. Here we have also created a mask for detector A3 only. In the future we will explore persistence on all detectors.

We apply the mask to the snowball-corrected count-rate maps by setting the image pixels to zero in the science extensions and setting their DQ values to D0_NOT_USE. We then perform the corrections for wisps and 1/f noise as described above. We create three sets of mosaics in F200W and F444W for NIRCam2: one containing just the original imaging, one with just the repeated imaging, and one including both sets of imaging.

We perform an astrometric calibration using a modified version of the jwst TweakReg routine, with the modifications primarily aimed at exposing more fitting parameters and being able to input our own catalogs. We note that similar modifications have now been fully implemented in the official STScI pipeline. The TweakReg routine calculates the transformation needed to align images to an absolute WCS frame. It does this by comparing the source positions detected in each input image with their matches in a reference catalog. The early versions of TweakReg performed source detection largely internally, with few tunable parameters available to the user, and offered only a Gaia catalog as the absolute reference. With our modified version of this step, we create a SourceExtractor (Bertin & Arnouts 1996) catalog for each individual input image, providing improved source detection, deblending, and centroiding over the internal TweakReg source identification. We also use a custom reference catalog derived from an HST F160W 003 pixel⁻¹ mosaic ³⁸ in the EGS field with astrometry tied to Gaia-EDR3 (see Koekemoer et al. 2011, for further details on the methods for image processing and mosaic creation). In both cases we use the SourceExtractor windowed coordinates for improved centroiding, which especially improves positions for the compact sources that we primarily use for our alignment. We again note that these options are now available in the pipeline version of TweakReg, and so our modified version is no longer needed.

We calculate both a relative and absolute astrometric correction. The images are first aligned relative to each other, where shifts in x and y are determined between images of the same detector. This step accounts for uncertainties in the guide star alignment and pointing accuracy during dithering. The rms of this relative astrometry is ∼3–6 mas per source. The images are then aligned to the HST F160W reference catalog, allowing for xy shifts and rotations. In the LW images we also allow for a scaling factor to account for any additional distortion across the larger detectors, though we note that the calculated deviations from a unity scaling factor are small (∼1 × 10⁻⁵). The rms of this absolute alignment (WFC3-to-NIRCam) is ∼12–15 mas and is driven by the larger F160W PSF. The rms of the alignment between NIRCam images in different filters (NIRCam-to-NIRCam) is ∼5–10 mas, with SW images at the smaller end and LW images at the larger end of that range. Figure 6 illustrates the quality of the astrometry in F200W and F277W and shows both the HST-to-NIRCam and NIRCam-to-NIRCam alignments.

Zoom In Zoom Out Reset image size

In summary, for all four NIRCam fields (1, 2, 3, and 6), for all filters and all detectors in both modules, the rms astrometric alignment quality is generally ∼5–10 mas per source between NIRCam filters (SW—LW) and ∼12–15 mas per source relative to the absolute frame defined by the HST F160W mosaic. There is, however, one exception to this: for NIRCam pointing 3, specifically for a $\sim 1^{\prime}$ region in the quadrant of the mosaic mostly covered by the B2 detector, the astrometric offsets relative to the HST F160W mosaic are larger, ∼005 (or about one-fifth of the 023 HST F160W mosaic PSF in this mosaic ³⁹ ). These offsets are also present in this region of the LW BLONG module, suggesting that it is not an alignment problem with a specific SW detector. After detailed examination of the relevant data sets, a possibility is that the HST F160W exposures for the particular visit in that part of the mosaic were affected by a low-level guide star tracking issue, at the sub-PSF level (∼1/5 of a PSF), which can occasionally occur and is not always revealed by the guide star telemetry keywords in the image headers. Since these residuals are small compared to the size of the F160W PSF, they are unlikely to significantly affect the overall HST-NIRCam photometry since our apertures are generally substantially larger than the HST PSF. We therefore followed a different process for aligning this NIRCam pointing. We first aligned F277W to F160W, excluding all sources in the upper left quadrant of detector BLONG from the fit. We then used F277W rather than F160W as the reference catalog for all other filters in this pointing. The rms of the alignment is still ∼3–6 mas (relative) and ∼5–10 mas (absolute NIRCam-to-F277W).

Before combining individual exposures into mosaics, we perform three additional corrections to the calibrated images. First, we subtract a pedestal value in MJy sr⁻¹ from each exposure. This step is necessary because the SkyMatch routine of jwst does not successfully match the background across all detectors, likely because the small dithers result in little to no overlap between detectors. In order to calculate the background, we mask bad pixels and source flux using the tiered masks created for each image in Section 3.1.3. We then fit a Gaussian to the distribution of unmasked, sigma-clipped pixel fluxes and take the position of the peak as the image pedestal. This single value is subtracted from each image, and the headers are updated to reflect this value. Second, we calculate the sky variance in each background-subtracted image. To do this, we mask sources in four tiers, as described in Section 3.3.3. We then block-sum the image in units of 7 × 7 pixels and use astropy's biweight_midvariance to make a robust estimate of the variance in blocks that have had no pixels masked. We estimate the equivalent per-pixel variance by dividing by the number of pixels per block (49). We then scale the read-noise variance array (VAR_RNOISE) to reproduce this value. Because VAR_RNOISE is used to compute the inverse variance that is used as weighting during the drizzle process, this ensures that the resulting error arrays do a good job of predicting the rms sky fluctuations (when estimated over large enough scales to avoid the pixel-to-pixel correlations introduced by drizzling).

Third, we make one final correction to the variance arrays to ensure that bad pixels are properly flagged. With our simulated mosaics (see Appendix D.4), we found that some known bad pixels had values of exactly zero in the variance arrays. When the dithered images were coadded, these areas in the output error array had relatively low rms compared with the average. In these areas, the input error array with the missing data or bad pixel did not contribute to the rms of the affected pixel. As a result, there were "holes" in the output error arrays in sets of three, matching our three dithers. Spurious source detections were more prevalent in these areas because the rms used for detection was low. For each individual image, we therefore replace any pixels that are exactly zero in the three variance arrays with values of infinity (numpy.inf). This correction ensures that the affected pixels are correctly down-weighted in the drizzling.

We create individual mosaics for each pointing using the Stage 3 routine Resample, which uses the drizzle algorithm with an inverse variance map weighting (Casertano et al. 2000; Fruchter & Hook 2002) to combine images into a single distortion-free image. The output mosaics are drizzled onto a common WCS with the same tangent point as the HST mosaics we have created in this field (see footnote 39). All HST and NIRCam mosaics are therefore pixel-aligned across all filters. The mosaics have pixel scales of 003 pixel⁻¹, and we chose not to shrink the input NIRCam pixels when drizzling (i.e., pixfrac = 1). This larger pixfrac is preferred for CEERS because the majority of the mosaic is covered by at most three exposures. The output pixel scale is only slightly smaller than the original scale in the SW channel (0031 pixel⁻¹) and just over two times smaller than that of the LW channel (0063). We discuss tests to evaluate the optimal drizzle parameters in Appendix D.4. The choice of pixfrac leads to increased pixel correlations in the output image, which we discuss in Section 4 and have not yet addressed in our mosaics.

These individual mosaics are included in our data release, and the files and extensions are described in Appendix A. We also created a set of combined, Epoch 1 mosaics that include all four pointings. Constructing these Epoch 1 mosaics requires considerable memory (∼600 GB, depending on the filter), and so we use the Frontera computing system at the Texas Advanced Computing Center (TACC). Frontera provides a set of large memory nodes that offer 2.1 TB of memory each with the option to link multiple nodes. 

Finally, we estimate and subtract any remaining background in the mosaics using a custom Python script that efficiently masks source flux before fitting the unmasked pixels with a two-dimensional model. The steps are outlined in Figure 7. The first step involves taking out large-scale fluctuations, masking bright pixels in the residual, and then running a large ring-median filter across that image to create the first good estimate of the background. The inner ring radius is 24, so this step preserves the wings of all but the largest galaxies in the image. This step creates background that is flat enough on large scales to permit us to mask sources above a fixed threshold without any perceptible gradients in source size and density due to large-scale changes in the background. We mask four tiers of sources, with tier number 1 intended to mask the most extended galaxies, moving progressively to the smallest galaxies at tier number 4. This is achieved by convolving the image with four different widths of Gaussians and masking areas with a minimum number N of connected pixels above a fixed threshold of 1.5σ. These masks are then grown by dilating them with a circular top-hat kernel with different kernel radii for each tier. We construct a source mask for each filter, including the pixel-aligned images in six available HST filters (F606W, F814W, F105W, F125W, F140W, and F160W). We merge all 13 separate masks into one, which is used to mask source flux in all filters. In this way we can exclude flux from sources that may be just below the detection threshold in some filters but not others.

Zoom In Zoom Out Reset image size

We then measure the background in the unmasked regions using the Photutils Background2D class. We use the biweight_location estimator to robustly calculate the average background in sigma-clipped boxes of 10 × 10 pixels in a grid across the image. The resulting low-resolution gridded background model is then median-filtered over 5 × 5 adjacent boxes, and the BkgZoomInterpolator algorithm is used to interpolate the filtered array and construct a smooth background model.

A "before and after" example of background removal is shown in Figure 8. Any estimation of background requires decisions about what is "source" and what is "background." The procedure outlined above does a very good job of removing residual wisps and other artifacts in the NIRCam images, while preserving the wings of most of the galaxies of interest. However, it does suppress the wings of bright galaxies (intentionally) to enable the detection of faint neighbors.

Zoom In Zoom Out Reset image size

To assess the background subtraction, we examined the stacked flux in pixels as a function of distance from the edge of the four separate tiers of masks. In most bands, for most tiers, there is very little "spillage" of galaxy flux into the pixels that were used to estimate the background. This spillage is impossible to avoid entirely. It is at a low enough level that we do not think the wings of galaxies that are above our detection threshold are significantly biasing the large-scale background estimation. In the left panel of Figure 9 we show an example for F200W in one of our fields.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

On large scales, one would like to see the rms of the background approach the rms expected from a completely flat image affected only by the counting statistics of the incoming signal. The right panel of Figure 9 shows the per-pixel rms that is inferred from measuring the rms in source-free (i.e., unmasked) regions of an F200W image, when block-summed over successively larger block sizes and then rescaled back to the per-pixel rms. On small scales, the rms is suppressed in the combined images because the drizzling procedure introduces a correlation between pixels. On medium scales, the residual wings of galaxies—which are impossible to remove entirely—slightly boost the rms. The rms approaches the ideal on larger scales, albeit with scatter in this measurement because there are very few blocks of 60 × 60 pixels and larger that are entirely source-free.

In Table 3, we present the 5σ limiting magnitudes in each filter and pointing. These depths are measured on our background-subtracted mosaics in circular apertures with radius r = 01 and scaled to an empirical estimate of the noise in the image as described by Finkelstein et al. (2022b). Briefly, apertures ranging in size from r = 005 to r = 15 are placed across each image, avoiding source flux and bad pixels. A robust estimate of the 1σ noise is calculated in each aperture, and following the examples of Papovich et al. (2016), Bagley et al. (2022), and Finkelstein et al. (2022a; see also Labbé et al. 2003; Blanc et al. 2008; Whitaker et al. 2011), a second-order polynomial is fit to the noise as a function of aperture size. We use this function to estimate the noise in an r = 01 aperture. Using stacked PSFs in each filter, we then measure the fraction of the total flux that is enclosed in an aperture of this size and correct the noise estimate by this amount. See Finkelstein et al. (2022b) for a full description of each step in this process. The 5σ depths for each pointing and filter are listed in Table 3. For comparison, we also list estimates from v2.0 of the JWST Exposure Time Calculator (ETC), calculated for point sources in an aperture with radius r = 01. The measured CEERS depths are consistent with the ETC estimates, ranging from ∼0.2% fainter to ∼0.3% brighter than the expectations. A color image of the Epoch 1 mosaic including all seven filters is displayed in Figure 10.

In this section, we explore the flux calibration of the CEERS mosaics, with the goal of identifying any residual corrections that may be needed to synchronize the zero-points in each detector. Ideally, this analysis would be performed by comparing the measured fluxes of the same set of sources when they are observed on different NIRCam detectors. However, due to the nature of the CEERS observing strategy, the NIRCam dithers are not large enough to observe a given source across multiple detectors, and so we do not have a direct test of detector-to-detector zero-points. We therefore explore the relative photometric calibration between detectors with two methods, which we describe below.

For the first method, we measured aperture photometry on NIRCam, HST/WFC3, and Spitzer/IRAC data for galaxies with m < 24 using Kron (1980) apertures. We used CANDELS HST mosaics (Grogin et al. 2011; Koekemoer et al. 2011) for the WFC3 filters and the Spitzer-Cosmic Assembly Deep Near-infrared Extragalactic Legacy Survey (S-CANDELS; PI G. Fazio) mosaics ⁴⁰ (Ashby et al. 2015) for the IRAC imaging. As described by Ashby et al. (2015), the IRAC mosaics were created using Corrected Basic Calibrated Data exposures processed with IRAC Pipeline versions S18.7.0, S18.18.0, and S19.1.0 and include the array position-dependent flux and pixel solid angle corrections. We first matched the NIRCam F115W PSF to that of WFC3 F125W, then used SourceExtractor in dual-image mode to measure source fluxes with identical apertures in the two filters, and finally compared these measurements detector by detector. We performed this analysis on the NIRCam mosaics, where each detector can be identified separately because the CEERS dithers are not sufficient to cover the detector gaps (see Figure 1). We repeated this same process comparing NIRCam F150W to WFC3 F160W, NIRCam F356W to IRAC 3.6 μm, and NIRCam F444W to IRAC 4.5 μm. We note that the passbands of these pairs of filters are not exactly the same, preventing us from using this method to obtain an absolute photometric calibration. Moreover, the derivation of absolute zero-points is challenging with extended sources owing to uncertainties in aperture corrections, making it difficult to separate the color-dependent correction for missing source wings from the flux calibration. Additionally, uncertainties in the PSFs used for image convolution and uncertainties in the local background determination in crowded regions contribute to making it challenging to use this comparison for an absolute calibration. Yet this approach of comparing NIRCam, WFC3, and IRAC fluxes in PSF-matched images (which we refer to below as "Method 1") should be robust for estimating relative calibrations between detectors provided that enough sources are used.

We use a second method for evaluating the relative flux calibration of the F200W, F277W, and F410M filters, for which there are no direct imaging data with comparable depth and spatial resolution available. For this "Method 2," we used the CANDELS photometric catalog (Stefanon et al. 2017) to obtain synthetic magnitudes in the NIRCam bands by convolving the filter transmission curves with the stellar population synthesis (SPS) models best fitting the spectral energy distributions of bright (F160W < 24 mag) galaxies. We then compared these synthetic magnitudes to Kron (1980) aperture photometry in the NIRCam images. As a test of the performance of "Method 2," we used it to evaluate the NIRCam photometry in F115W, F125W, F356W, and F444W and compared the results to those obtained via "Method 1." For these four filters, we obtained consistent results with the two methods, compatible within 1%–2%, but with a 1.5× larger scatter for "Method 2." We conclude that "Method 2" provides a robust way to evaluate the relative flux calibration for F200W, F277W, and F410M.

Figure 11 shows our results comparing the detector-to-detector photometric calibrations for CEERS data reduced with CRDS pmap 0959 (red circles) and 0989 (blue circles). The pmap 0959 results are from an earlier reduction of the CEERS data, while the pmap 0989 results are those presented in this paper. In each case, we use "Method 1" with F115W, F150W, F356W, and F444W and "Method 2" with F200W, F277W, and F410M. We calculate the relative offsets of all SW detectors with respect to B1 and both LW detectors with respect to BLONG. We note that we have also compared our results for pmap 0959 (red circles) with those of Boyer et al. (2022), who used the globular cluster M92 to determine residual photometric offsets for this pmap. The relative detector-to-detector photometric offsets we measure using galaxies are consistent to within 1%–2% with those from Boyer et al. (2022) obtained with stars. Although stars are better suited for flux calibration, this comparison with Boyer et al. (2022) validates our approach for using galaxies to evaluate the flux calibration in CEERS imaging.

Zoom In Zoom Out Reset image size

As can be seen in Figure 11, the improved flux calibration released with pmap 0989 (blue circles) reduces the detector-to-detector offsets that existed with the flux calibration of pmap 0959 (red circles). However, with pmap 0989, we still detect offsets between detectors at the <5% level, with a median of 2.5%. We do not apply these small offsets to our images at this time, but we warn the reader that the relative and absolute calibration of NIRCam images is still uncertain at the ∼2%–4% level.

Finkelstein et al. (2022b) have also explored the accuracy of the photometric calibration by exploring what average offsets in flux calibration would be needed to minimize the colors between observations and best-fitting spectral templates for a sample of several hundred spectroscopically confirmed galaxies. They also find that the calibration is good to 1%–2% in most bands, and no worse than ∼5% (in the bluest bands). Again, we note that these corrections were derived based on galaxies and thus depend on aperture corrections. Future photometric calibration updates based on stellar observations will reduce the amplitude of these corrections.