S-CANDELS: THE SPITZER-COSMIC ASSEMBLY NEAR-INFRARED DEEP EXTRAGALACTIC SURVEY. SURVEY DESIGN, PHOTOMETRY, AND DEEP IRAC SOURCE COUNTS

Because the scientific emphasis of S-CANDELS is on detecting and characterizing galaxies at very high redshifts, it is vital that the S-CANDELS fields be placed where sensitive photometry is available in multiple bands other than the IRAC 3.6 and 4.5 μm filters. Near-infrared (NIR) and visible imaging deep enough to match the IRAC observations reported here are of special importance. Accordingly, we chose to locate S-CANDELS inside the wider and shallower fields already covered by SEDS, in regions that enjoy deep optical and NIR imaging from HST/CANDELS. These S-CANDELS fields are thus the Extended GOODS-south (aka the GEMS field, hereafter ECDFS; Rix et al. 2004; Castellano et al. 2010), the Extended GOODS-north (HDFN; Giavalisco et al. 2004; Wang et al. 2010; Hathi et al. 2012; Lin et al. 2012), the UKIDSS UDS (aka the Subaru/XMM Deep Field, Ouchi et al. 2001; Lawrence et al. 2007), a narrow field within the EGS (Davis et al. 2007; Bielby et al. 2012), and a strip within the UltraVista deep survey of the larger COSMOS field (Scoville et al. 2007b; McCracken et al. 2012). These five S-CANDELS fields are distributed in ecliptic longitude and declination to permit ground-based followup from both hemispheres.

The depths, areas, and sensitivity of earlier IRAC coverage of the five S-CANDELS fields up to and including the SEDS campaigns are described by Ashby et al. (2013a). The S-CANDELS observations were of a similar character, but had a different etendue. Figure 2 shows the cumulative depth versus area plots for S-CANDELS, which had a design depth of 50 hr.

Zoom In Zoom Out Reset image size

The S-CANDELS observing strategy was designed to maximize the area covered to full depth within the CANDELS ${H}_{160}$ area. Each field was visited twice⁷ with six months separating the two visits. Table 1 lists the epochs for each field. All of the IRAC full-depth coverage is within the SEDS area (Ashby et al. 2013a), and almost all is within the area covered by HST for CANDELS. (See Figures 3–12.)

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Each of the two observation epochs accumulated 19 hr integration time per pointing, or less when a field had pre-existing coverage other than SEDS. For efficiency, each position in the field was usually observed for 2 frames of 100 s each before moving the telescope to the next position. The medium Reuleaux 36 dither pattern was used throughout, except for the EGS field, which used an 18-point dither pattern.⁸ The AORs thus sampled each sky position at many positions on the arrays. Each AOR (pair of linked AORs for the EGS) covered the full intended field of view for one wavelength, but the 3.6 and 4.5 μm coverage did not overlap for UDS, HDFN, or CDFS. For COSMOS and EGS, the overlap was only partial. However, the IRAC fields of view switch places every six months, so the area observed at 3.6 μm in one epoch was observed at 4.5 μm in the alternate epoch and vice versa to achieve complete coverage of the intended area at both wavelengths.

We used the same procedures to reduce the S-CANDELS data as were applied earlier to the SEDS observations described by Ashby et al. (2013a). In the following we therefore describe only the most important aspects of the S-CANDELS reductions.

All suitable data were combined into full-mission mosaics that include coextensive imaging from SEDS and other projects from both the cryogenic and warm missions (see Table 1 for the complete lists) into full-mission mosaics covering the CANDELS fields. Processing was based on IRAC Corrected Basic Calibrated Data exposures generated by the pipeline versions indicated in Table 1. The different pipeline versions differ only in matters of minor artifact correction, not in overall calibration⁹ of the 3.6 or 4.5 μm exposures.

Before mosaicking, all the IRAC exposures were corrected for long-term residual images and for column pulldown. The mosaics were constructed with IRACproc (Schuster et al. 2006) in the same way as was done for SEDS, but over narrower fields. In the ECDFS and the HDFN, which have very large datasets, computer memory constraints prevented us from making the mosaics in a single IRACproc run. For these fields we mosaicked subsets of the exposures and subsequently mean-averaged the results into a single mosaic. As with SEDS, all the S-CANDELS mosaics were pixellated to 06 and were aligned to the tangent-plane projections used by the CANDELS team (Table 5 of Koekemoer et al. 2011). Figure 3 through 12 show the final IRAC mosaics for all five fields.

The final mosaics, coverage maps, model images, and residual images are all available from the Spitzer Exploration Science Programs website.¹⁰

Source confusion is pronounced even for the 12 hr SEDS mosaics; the problem impacts the deeper S-CANDELS mosaics even more strongly. We therefore used StarFinder (ver.1.6f; Diolaiti et al. 2000) to identify sources because StarFinder is optimized for identification and photometry of heavily blended sources in crowded fields (e.g., globular cluster stars). As was done for SEDS, the S-CANDELS catalogs were constructed in two steps. First, StarFinder was used to identify and locate sources (even faint, blended ones). Second, a custom code was used to correct biases in the StarFinder photometry.

The source-identification step was performed on the full-depth S-CANDELS mosaics. StarFinder implements an algorithm based on iterated fitting and subtracting of a template point-spread function (PSF) image. Because the S-CANDELS mosaics are small and heavily confused, we were unable to identify enough isolated, sufficiently bright point sources to construct useful PSFs template images from the S-CANDELS mosaics themselves. Instead, we used the PSF template images constructed earlier for SEDS. Because the fields were observed at similar roll angles for both SEDS and S-CANDELS, this should not introduce significant error. StarFinder is capable of repeating the source-identification algorithm using the residuals it generates from a first pass through the mosaics. This allows the code to refine (reduce) its estimate of the background noise in the absence of the brightest objects. We therefore configured StarFinder to process each field three times, as was done for SEDS, with identical parameter settings. In particular, the software was not allowed to deblend sources closer together than 09, roughly half the FWHM of the PSF in the two warm IRAC passbands. Based on our inspections of the final (third-pass) residual images (Figure 13 shows an example), we judged this approach successful. The residual images are manifestly free of large-scale background artifacts, and the faint sources (which are all effectively point sources) are well-fitted by our approach.

As was done for SEDS, the aperture magnitudes for each source were measured after re-inserting its fitted PSF at its fitted position in the StarFinder residual image and then measuring background-subtracted fluxes interior to diameters of 24, 36, 48, 60, 72, and 120. Thus the S-CANDELS aperture magnitudes are relatively free of contamination by nearby neighbors (because to a good approximation they have been subtracted off) of both the photometered source and the nearby background. The S-CANDELS catalogs contain both the PSF-fitted magnitudes based on the iterative procedure described above and the aperture magnitudes. Because we used the same PSF templates and photometric apertures as SEDS, we also used the same SEDS photometric corrections to correct the original, PSF-fitted magnitudes to total magnitudes. All S-CANDELS catalogs include these aperture corrections, which are given by Ashby et al. (2013; their Table 2). To avoid spurious sources, only objects detected in both bands were included in the catalogs.

The completeness and reliability of S-CANDELS were assessed on a field-by-field basis with the standard Monte Carlo approach, identical to that used for SEDS (Ashby et al. 2013a). SEDS established, by matching CANDELS F160W sources to IRAC-detected objects in COSMOS, that all IRAC sources fainter than 23 AB mag are point sources at IRAC resolution. This is true even for a majority of IRAC sources brighter than 23 AB mag. We therefore used only point sources in our completeness and depth simulations. We simultaneously inserted simulated sources in both the 3.6 and 4.5 μm mosaics at identical locations. For simplicity, the simulated sources were created with [3.6]−[4.5] = 0, i.e., no attempt was made to insert sources with a range of colors. After the S-CANDELS mosaics had been modified by inserting simulated sources, source identification and photometry was performed in exactly the same way as for the unmodified mosaics. The completeness and magnitude bias were assessed by comparing the results to the a priori known input sources over the range of magnitudes seen for the real sources. The results are given in Table 2 and shown in Figure 14.

Zoom In Zoom Out Reset image size

S-CANDELS completeness is identical to that of SEDS for sources brighter than about 24.5 AB mag. For sources fainter than 24.5 mag, however, S-CANDELS is significantly more complete than SEDS, recovering a larger but flux-dependent fraction of the simulated sources. The improvement relative to SEDS ranges up to a factor of several, depending on the specific field and source magnitude. Taken at face value, S-CANDELS reaches 50% completeness at roughly [3.6] = [4.5] = 25 mag in all fields except the ECDFS, where (because of the additional coverage from the ERS and IUDF programs), the 50% completeness threshold is reached at 25.3 mag. Users of S-CANDELS data are cautioned that these are generalizations; the depths are variable across the S-CANDELS fields, and the completeness at any one location is a strong function of both the local source density and the total exposure time.

As with SEDS, the S-CANDELS estimates of photometric error and bias were also based solely on the artificial source simulations, in order to account for the impact of source confusion in these very deep mosaics. The photometric uncertainties and biases are given respectively in Tables 3 and 4 for each of the S-CANDELS fields, and are shown in Figure 15. The S-CANDELS photometric uncertainties are very close to those measured in the shallower SEDS mosaics (Figure 27 of Ashby et al. 2013a). This is discussed in Section 4.2.

Zoom In Zoom Out Reset image size

As with SEDS, the measurement bias is relatively small for sources brighter than the 50% completeness limit but grows rapidly at progressively fainter magnitudes (Table 4). This appears to confirm an interpretation in which faint sources are increasingly difficult to deblend from their neighbors. The contamination of the photometric apertures by imperfectly subtracted brighter neighbors affects the photometry even though the measurements were made in source-subtracted residual images. The S-CANDELS catalogs have been corrected to remove the resulting average magnitude bias.

We verified our astrometry by comparing the positions of extracted sources to their counterparts in the references used by the CANDELS team. We also compared to the bright sources in the Two Micron All Sky Survey (Skrutskie et al. 2006). The results are shown in Table 5. The S-CANDELS astrometry is consistent with previous work in the five CANDELS fields. The scatter measured for the positions of sources on the IRAC and non-IRAC catalogs is of order 02, which is also consistent with analogous measurements carried out in other fields, e.g., SEDS, the SSDF (Ashby et al. 2013b), and SDWFS (Ashby et al. 2009).

We verified our photometry by comparing to the measurements obtained earlier in the shallower SEDS campaign, which were themselves already validated against S-COSMOS (Sanders et al. 2007), SpUDS (version DR2), the EGS (Barmby et al. 2008), and SIMPLE (Damen et al. 2011).

To verify the S-CANDELS flux calibration, we matched the S-CANDELS catalogs to those from SEDS. In all cases, the matching was done using a $0\buildrel{\prime\prime}\over{.} 5$ search radius, i.e., roughly twice the S-CANDELS $1\sigma$ astrometric uncertainty and one-third the IRAC PSFs' FWHMs. Only SEDS sources brighter than 26 AB mag (the SEDS 3σ detection limit) were used for the comparison. Sources close to saturation (<15.4 mag in 200 s exposures) were excluded. The results are shown in Figure 16.

Zoom In Zoom Out Reset image size

For sources brighter than 25 mag in all fields, S-CANDELS photometry agrees very well with SEDS. There are a few exceptions. In HDFN three of seven sources in the [3.6] = (16.0,16.5) bin differ by ∼0.2 mag from SEDS, and two of five sources in the brightest 4.5 μm ECDFS bin are discrepant at a similar level. All of the discrepant sources are bright point sources (Milky Way stars). The S-CANDELS photometry in the complementary IRAC band for these sources agrees with that from SEDS (Figure 16). Variability is therefore unlikely to be the issue. All discrepant sources lie in parts of the mosaics that combine SEDS and S-CANDELS exposures, and moreover the S-CANDELS residual images for these sources are markedly different than those from SEDS. This suggests that although the PSF fitting technique worked for the vast majority of S-CANDELS sources, it failed for these few objects, for reasons particular to the details of their immediate surroundings and the mechanics of StarFinder.

The photometry for faint sources follows a more consistent pattern. Over a wide magnitude range in both SCANDELS bands the agreement between SCANDELS and SEDS is excellent. In all five fields, however, as the SEDS 26 mag sensitivity limit is approached, a bias becomes apparent in the sense that SEDS sources are systematically brighter than their S-CANDELS counterparts. The bias is lowest overall in the HDFN (∼0.1 mag), and highest in the UDS (∼0.5 mag).

To better understand the reason for the faint-source bias, we inspected both the SEDS and S-CANDELS data (mosaics and catalogs) at the locations of the most problematic sources, i.e., those with discrepancies greater than 0.5 mag. Apart from a tendency—by no means universal—to lie in the outskirts of bright sources, the discrepant sources present no obvious common trait in the residual images. They do not lie in regions with obvious background artifacts. Indeed, the S-CANDELS and SEDS photometry of neighbors within $7\prime\prime$ of discrepant sources agree within the uncertainties, with very few exceptions. The discrepancies are therefore not attributable to issues with the background modeling. The vast majority of discrepant sources also have the same number of neighbors within $7\prime\prime$ in both SEDS and S-CANDELS. The problem therefore does not generally arise from the StarFinder deblending procedure; the same numbers of sources lie in the peripheries of the discrepant sources in both SEDS and S-CANDELS. Finally, we compared the coordinate offsets of matched SEDS and S-CANDELS sources. We found no evidence to suggest that the most discrepant sources were significantly spatially offset in SEDS and S-CANDELS, relative to sources with consistent photometry. Inappropriate placement of the StarFinder PSF centroids and apertures is therefore not likely to be the problem.

Having ruled out issues with offset coordinates, poor background estimation, and deblending of different numbers of neighbors, we tentatively attribute the faint-source bias to flux boosting by very-low-level cosmic rays that are not efficiently rejected at SEDS depths. With the factor-of-four greater number of exposures available to S-CANDELS, there is statistical power to reject faint outliers that cannot be ruled out at SEDS depths. This hypothesis is consistent with the fact that the bias is seen to be most pronounced in the faintest two SEDS magnitude bins, and is of roughly the same size as the SEDS uncertainties themselves. The S-CANDELS magnitudes show similar bias but only for sources roughly 0.5 mag fainter than for SEDS, so we cannot rule out an analogous effect in the faintest S-CANDELS bins (cf. Table 4, Ashby et al. 2013a, Table 5).

S-CANDELS detects roughly 135,000 sources in the combined 0.16 deg² area covered by the five fields in the survey. Figures 17 and 18 and Table 6 present the resulting differential source counts along with Milky Way star counts estimated from the Arendt et al. (1998) model for the S-CANDELS lines of sight.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The S-CANDELS counts rely on completeness corrections that are based on simulated sources with zero color, i.e., $[3.6]-[4.5]=0$. At faint levels they could therefore in principle suffer from subtle systematic effects, because real sources span a range of colors (Figure 19). Our simulations do not fully account for faint, blue 3.6 μm sources, which would tend to elude detection in the 4.5 μm band. Faint, red 4.5 μm sources would be under-counted for the same reason. However, these systematic effects are unlikely to severely bias the S-CANDELS counts. The real IRAC color distribution peaks at $[3.6]-[4.5]=0$, and the vast majority (∼80%) have colors within 0.4 mag of the peak (Figure 19), even down to faint levels. Moreover, the area-weighted mean S-CANDELS counts (Figures 17(f) and 18(f)) show very close agreement with SEDS over the full range of comparison. For $[3.6]=[4.5]\gt 20$ mag, the counts show no significant deviations from those found for SEDS, suggesting the SEDS completeness corrections were accurate. S-CANDELS uses the same techniques, so its completeness corrections should be similarly robust.

Zoom In Zoom Out Reset image size

At levels brighter than roughly 18 AB mag in both S-CANDELS bands and in every field, the IRAC counts are consistent with the star count models. SEDS contains relatively few galaxies brighter than 18 mag. The vast majority of sources fainter than 18 mag, however, are galaxies: the contributions of Milky Way stars to the faint counts are negligible.

Helgason et al. (2012) modeled galaxy counts using an ensemble of galaxy luminosity functions assembled from deep multiband observations. They then used this ensemble to predict faint galaxy counts in several passbands, including 3.6 and 4.5 μm. They used existing counts to constrain the faint-end slopes of their luminosity functions. Specifically, only a limited range of faint-end slopes, corresponding to a range of acceptable values for the parameter α, was found to be consistent with existing counts. That range extends from their so-called high-faint-end, with $\alpha =-1.2$, to the low-faint-end ($\alpha =-0.8$). They considered also a "default" model that averages these two cases. The IRAC counts closely follow the "default" model all the way down to [3.6] = [4.5] = 26 mag (Figures 17(f) and 18(f)). At fainter levels, the counts depart upward in the direction of the high-faint-end scenario. This may not be real, because it occurs at magnitudes where the completeness correction is largest, magnifying any small systematic errors that might be present in the counts. It is also consistent with the possibility that faint sources undergo flux boosting, as described in the preceding section. What can be said with confidence is that the Helgason et al. (2012) models work very well down to very faint levels. More sensitive observations that can overcome the source confusion seen in the IRAC mosaics (e.g., imaging with the James Webb Space Telescope or the Wide-field Imaging Surveyor for High Redshift; WISH) will be necessary to confirm this picture for the faintest IRAC-detected sources.

For sources brighter than 24.5 mag, the S-CANDELS source detection fraction is not significantly better than that of SEDS despite a factor-of-four improvement in overall integration time. Sources at 24.5 mag or brighter lie well above even the SEDS detection threshold. Sensitivity alone is therefore not limiting the bright-source detection by IRAC in this regime. Moreover, the empirically determined S-CANDELS photometric uncertainties for bright sources are very similar to those for SEDS. We suggest that source confusion is the dominant contributor to the photometric uncertainties for magnitudes $\lt 24.5$ mag. Deeper IRAC observations alone will not improve the detection fraction or the photometric uncertainties for such bright sources.

For fainter sources, the picture is more nuanced. Source confusion is undoubtedly a factor, as evidenced by the similarity of SEDS and S-CANDELS uncertainties down to the limits of the surveys. However, for the deeper S-CANDELS, the detection fraction at $\gt 24.5$ mag is up to factors of several larger. Inspection of the respective catalogs revealed that the majority of the faint S-CANDELS sources not detected by SEDS lie in relatively source-free portions of the fields. It is in precisely these places that the improvement in sensitivity can be effective at identifying faint objects by decreasing the background shot noise.

One way to better understand the impact of source confusion on deep IRAC imaging is to quantify the available source-free area that will yield additional IRAC detections when imaged more deeply. A conservative estimate of this area is that in which detected sources contribute less surface brightness than the surface brightness noise level σ. We estimated σ for both SEDS and S-CANDELS using the residual images, i.e., after removing detected sources, and allowing for the effect of correlated noise.¹¹ By this definition, the source-free areas in 12 hr SEDS integrations are ∼40% and ∼50% in the 3.6 and 4.5 μm bands, respectively. The fractions that remain free in the 50 hr S-CANDELS mosaics are smaller, ∼30% and ∼40% at 3.6 and 4.5 μm, respectively. In other words, of order half the SEDS area and one-third the S-CANDELS area is effectively clear. Within those areas, integrating longer to reduce the shot noise can improve the detection statistics, and we see the results in the increased S-CANDELS completeness (relative to SEDS) for sources fainter than 24.5 mag.

In summary, source confusion does play a role for faint sources, but the much deeper S-CANDELS program nonetheless detects a significantly greater fraction of such sources. Somewhat counter-intuitively, the faint IRAC sources are not as rigidly limited by source confusion as the bright ones.

The IRAC colors of the sources give clues to their redshifts and luminosities. For example, Sorba & Sawicki (2010) and Barro et al. (2011) showed that the $[3.6]-[4.5]$ color is a useful photometric redshift indicator, especially for separating galaxies at $z\gtrsim 1.3$ from those at $z\lesssim 1.5$. Ashby et al. (2013a) showed (their Figure 31) that the observed color distribution is in fact bimodal for $[3.6]\lt 23.5$ and that the red peak grows relative to the blue one as fainter sources are considered. Figure 19 shows the same trend for the fainter sources observed in S-CANDELS. Smaller color uncertainties than in SEDS give a hint of bimodality for $23.5\lt [3.6]\leqslant 24.5$, but fainter sources show a single peak. Sources with blue colors are still present, but their proportion is significantly smaller than at brighter magnitudes. The effect of increasing uncertainty in the photometry of the faintest sources (Figure 15), and likewise in their colors, is visible as a broadening of the wings of their histogram. The faintest sources plotted, at $[3.6]\approx 25$, correspond to ∼5 mag fainter than ${L}^{*}$ at $z=1.2$ or ∼3 mag fainter than ${L}^{*}$ at $z=2.9$ (based on luminosity functions given by Helgason et al. 2012). At $z\gt 3$ the galaxy space density decreases approximately exponentially, and such galaxies will constitute only a small fraction of the sample. Therefore most of the faint sources are likely to be galaxies a few magnitudes fainter than ${L}^{*}$ at $z=1.2$–3.

Space-based surveys such as S-CANDELS, hold the potential to identify the source giving rise to the Cosmic Infrared Background (CIB). That part of the CIB that arises in discrete sources can be robustly estimated by identifying and photometering those sources and subsequently computing their contribution to the CIB in toto. Indeed, this was one of the original motivations for both SEDS and S-CANDELS. The outcome of the SEDS measurement was that IRAC-detected sources account for only about half of the DIRBE CIB estimates. More specifically, SEDS 3.6 and 4.5 μm sources account respectively for 5.6 ± 1.0 and 4.4 ± 0.8 nW m⁻² sr⁻¹ down to 26 mag (see Ashby et al. for details of the measurement). This is less than half the estimate from DIRBE (13.3 ± 2.8 nW m⁻² sr⁻¹; Levenson et al. 2007), although one should bear in mind that the DIRBE estimate depends on modeling and subtracting a large and inherently uncertain zodiacal foreground that is very bright compared to the CIB surface brightness.

With its greater completeness, S-CANDELS confirms the initial SEDS measurements (Figures 17 and 18) of the resolved fraction of the CIB. However, the increase in the resolved CIB from S-CANDELS is small: just 0.08 ± 0.03 and 0.09 ± 0.03 nW m⁻² sr⁻¹ at 3.6 and 4.5 μm, respectively. Thus even with the fourfold increase in overall integration time, the marginal increase in resolved CIB light from S-CANDELS is much less than the uncertainty in the original SEDS measurement. The revised estimate for the total contribution of resolved sources to the CIB in the IRAC bands is 5.7 ± 1.0 and 4.5 ± 0.8 nW m⁻² sr⁻¹.