Near-infrared Survey and Photometric Redshifts in the Extended GOODS-North Field

This work presents new J-band and H-band observations (PI: L. Lin) in the extended GOODS-N region (Figure 1) together with K_s-band data obtained through Hawaiian (PI: L. Cowie) and Canadian (PI: L. Simard) programs using WIRCam on the CFHT (Wang et al. 2010). The pixel size of WIRCam is 03, and the transmission curves of the three NIR filters are shown in Figure 2. All observations were carried out during 2006–2015 (see Table 1), and the images cover the extended GOODS-N region with area of 28' × 28'. As observations have accumulated during this long-term project, several studies utilizing these data have already been carried out (e.g., Younger et al. 2007; Keenan et al. 2010; Cassata et al. 2011; Murphy et al. 2011; Shim et al. 2011; Guo et al. 2012; Lee et al. 2012; Lin et al. 2012; Penner et al. 2012).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The observation setups in J and H were similar to those used in K_s (Wang et al. 2010). Within each observing block (OB), small dithering patterns on the order of tens of arcseconds were used to cover both vertical and horizontal gaps between chips. In each dither position, two to four subframes were taken with a typical exposure time of 60 s for J and 15 s for H. In addition, we offset the centers of dithering by a few arcminutes after each OB to reduce any systematics associated with different chips. The typical seeing for the J, H, and K_s images was between 07 and 085 (FWHM). The total integration times in J, H, and K_s are 47, 24, and 52 hr, respectively.

The data were preprocessed using the SIMPLE Imaging and Mosaicking PipeLinE (Wang 2010). The procedure included flat-fielding, distortion correction, sky subtraction, cross-talk removal, and photometry calibration against the Two Micron All Sky Survey (2MASS; Skrutskie et al. 2006) point-source catalog, as described in detail by Wang et al. (2010) and Lin et al. (2012). The preprocessed frames were further astrometrically calibrated against the Sloan Digital Sky Survey Data Release 6 (DR6) catalog and internal photometry calibrated using the AstrOmatic software Scamp (Bertin 2006). Finally, all the reduced frames were median stacked using Swarp¹⁹ (Bertin et al. 2002). For the final NIR images, the 5σ limiting magnitudes (2'' diameter aperture) reach J = 24.7, H = 24.2, and Ks = 24.4.

Taking advantage of our deep homogeneous NIR images in the wide field, we assembled publicly released data covering wavelengths from ultraviolet (UV) to midinfrared (MIR) to compute photometric redshifts (photo-zs). The data we collected are as follows:

• 1.  
  UV: The far-UV (FUV) and the near-UV (NUV) data were obtained from the Galaxy Evolution Explorer (GALEX). We matched our NIR detections (i.e., z-, J-, H-, or K_s-detected sources, see Section 3.1) with the GALEX General Release 6/7 (GR6/7). Although the GALEX image has a point-spread function (PSF) of ∼5'', we adopted a search radius of 1'' to decrease the number of sources with contaminated photometry caused by blending. The UV data are used only for the X-ray source counterparts (Section 4.2) and therefore will not affect sources that are not X-ray counterparts.
• 2.  
  Visible: Capak et al. (2004) provided U, B, V, R, I, and z images from the Suprime-Cam²⁰ on the Subaru telescope with 5σ limiting magnitudes ranging from 25 to 27 mag. Additionally, we added y-band data taken from the Subaru/Suprime-Cam instrument (Ouchi et al. 2009).
• 3.  
  MIR: We used the IRAC 3.6 and 4.5 μm data from SEDS (Ashby et al. 2013). IRAC 5.8 and 8.0 μm data are much shallower than the other two bands, and few objects were detected at these wavelengths. For the SED fitting process, we used only λ ≤ 5 μm, as suggested in the LePhare manual (page 26), because the stellar light is dominant at these wavelengths.

Table 2 provides detailed information on the photometric data used to calculate the photometric redshifts. In all, 14 bands were used for X-ray counterparts and 12 bands for all other sources.

X-ray observation is an efficient method to detect AGNs. Studies based on the 4 Ms Chandra Deep Field-South (4Ms-CDFS; Xue et al. 2011) and 2 Ms Chandra Deep Field-North Survey (2Ms-CDFN; Alexander et al. 2003; Xue et al. 2016) show that more than 80% of X-ray sources are AGNs. As described by Salvato et al. (2009, 2011), these powerful objects have more complex SEDs than normal galaxies and therefore require special treatment for photo-z estimation. Xue et al. (2016) provided an updated 2Ms-CDFN X-ray catalog, which contains 683 X-ray sources including 196 additional X-ray sources compared to the former catalog from Alexander et al. (2003). We used this X-ray catalog to match with our catalog and identify objects to treat as AGNs in the photo-z estimates.

Spectroscopic redshifts (spec-zs) were collected from a number of works, mainly those of Barger et al. (2008), Cowie et al. (2004), Wang et al. (2010), Cooper et al. (2011), and Xue et al. (2016). We matched our NIR detections (see Section 3.1) with spectroscopic data within a maximum separation of 1''. In total, about 3600 sources (∼4%) have spec-z information, and most of them are in the central GOODS-N region. For more reliable estimates of the photo-z quality, we used only 3459 good-quality spec-zs as indicated by the quality flag claimed in the literature.

To combine photometric data from different surveys, we have to make sure that all images are referred to the same astrometric reference frame. Comparing the coordinates in Capak's visible images with our NIR images, the median values of systematic offsets are ∼03–04 without any correction. To obtain consistent astrometry, we aligned all the visible images (i.e., U, B, V, R, I, z, and y) with the WIRCam K_s-band mosaic. After this calibration, the median values of the systematic offsets between visible and NIR images range from 013 to 016. This small offset allows us to do the source detection and flux extraction for all the images on the same astrometric reference frame.

Photometry was extracted based on a detection image created by stacking the non-homogenized NIR (i.e., J, H, and K_s) and z-band images. These wavelengths are sensitive to UV-luminous sources at high redshift (Laigle et al. 2016). The stacked image was created with the CHI2 mode of swarp (Bertin et al. 2002). In this mode, the output image is the square root of the reduced χ² of pixel values in input images at a given position.²¹ This so-called χ² image was used as the detection image in the dual-image mode of SExtractor to define photometric apertures for each single-band image. In total, 93,598 sources were detected.

For photometry extraction, we used astrometry aligned to the WIRCam K_s-band image. However, in the final released catalog, we provide absolute astrometry aligned with the Advanced Camera for Surveys (ACS) catalog from Giavalisco (2012). Figure 3 compares our zJHK_s detections with the ACS catalog to show the astrometric offsets. The median values of positional offset in right ascension (R.A.) and declination (Decl.) are 0085 and 0004, respectively.

Zoom In Zoom Out Reset image size

Our 10 ground-based images (U, B, V, R, I, z, y, J, H, and K_s) have FWHMs of the PSF ranging from 08 to 13. This will give inaccurate fluxes if the same fixed-size aperture is adopted for all the bands because the ratio of aperture flux density to total flux density will depend on PSF size. In order to obtain accurate total flux densities and colors, which can be used to fit the SEDs and compute photo-zs, we need to take the varying PSFs into account. We adopted the approach of Capak et al. (2007), degrading the better-seeing images to the largest PSF (that of the U-band image) of 13 using a Gaussian kernel appropriate to each image. To be specific, we used the PSF characterization (calculating FWHM on images) and equalization (smoothing with a Gaussian kernel) tools in the Subaru SuprimeCAM reduction package (sdfred) in iraf.²²

Because the PSF sizes of GALEX and IRAC are very large, to avoid blending, we did not include UV and IRAC data in the PSF homogenization. Instead we obtained the UV total flux densities directly through the GALEX online query system "CasJobs."²³ The IRAC total flux densities were extracted with the improved IRACLEAN (Hsieh et al. 2012; Laigle et al. 2016), which is designed especially for IRAC.

3.3.1. Absolute Photometry of NIR Images

The absolute photometry of our WIRCam J, H, and K_s images was calibrated with 2MASS (i.e., The Two Micron All Sky Survey) point sources in the observed field. Figure 4 shows the magnitude differences between the 2MASS and the calibrated WIRCam magnitudes. The scatter becomes larger at mag <14.5 and >16.5. The magnitude differences are mostly constant between 14.5 mag (the brightest WIRCam sources) and 16.5 mag (where 2MASS S/N deteriorates). Therefore, we chose the median value in the range 14.5 < mag < 16.5 to set the calibration for each WIRCam NIR image. In the J band, there are some sources with large scatter at J < 15.5. Most of these are saturated stars, and their WIRCam photometry is not reliable. However, these sources do not affect our calibrations. As mentioned by Wang et al. (2010), the magnitude uncertainties in Figure 4 are dominated by 2MASS. Therefore our NIR flux calibration is limited by the 2MASS systematic uncertainty.

Zoom In Zoom Out Reset image size

3.3.2. U to K_s Flux Extraction

As mentioned above, U to K_s photometry was performed with SExtractor in dual-image mode. We used the zJHK_s χ² image as the detection image and measured aperture photometry on each PSF-homogenized image. Fixed-aperture flux densities (i.e., FLUX_APER) have higher signal-to-noise ratios than the automatic-aperture flux densities (i.e., FLUX_AUTO) (Kron 1980). Therefore, fixed-aperture photometry gives "cleaner" colors for SED fitting and more accurate photo-zs. Tests showed that 2'' FLUX_APER gave better photo-z quality (i.e., better accuracy and fewer outliers) than either FLUX_AUTO or 3'' FLUX_APER. We therefore used 2''-aperture photometry for photo-z computation. However, both 2'' FLUX_APER and FLUX_AUTO are included in the released catalog.

The SExtractor parameters we used are mainly adopted from Wang et al. (2010) and Hsieh et al. (2012), who observed with the same instrument (i.e., CFHT/WIRCam) as we did. The parameter CLEAN_PARAM in particular is critical for source detection. We examined values of CLEAN_PARAM from 0.1 to 1.0 and found that the number of detections increased from ∼90,000 to  ∼ 130,000. Setting CLEAN_PARAM = 0.1 avoids a large number of false detections as demonstrated by Yang et al. (2014). We also visually inspected random sources on the images and confirmed that false detections are greatly reduced with a lower value of CLEAN_PARAM. Table 3 lists the main SExtractor parameters used to obtain the photometry.

Table 3.  SExtractor Parameters

Parameter	Value
DETECT_MINAREA	2
DETECT_THRESH	1.2
ANALYSIS_THRESH	1.2
FILTER	Y
FILTER_NAME	Gauss_1.5_3x3.conv
DEBLEND_NTHRESH	64
DEBLEND_MINCONT	0.00001
CLEAN	Y
CLEAN_PARAM	0.1
SEEING_FWHM	0.8
BACK_SIZE	24
BACK_FILTERSIZE	3
BACK_TYPE	AUTO
BACKPHOTO_TYPE	LOCAL
BACKPHOTO_THICK	40
WEIGHT_TYPE	MAP_WEIGHT

Download table as:  ASCIITypeset image

Because GALEX and IRAC photometry is provided as total flux density, to combine them with visible and NIR photometry, we need to convert visible and NIR aperture flux densities to total flux densities for each band using the equation:

$$\begin{eqnarray}&&{f}_{\mathrm{tot}}={f}_{\mathrm{aper}}\times \mathrm{median}\left(\displaystyle \frac{{f}_{\mathrm{auto}}}{{f}_{\mathrm{aper}}}\right)\times 1.06.\end{eqnarray} \tag{ 1 }$$

First we converted the fixed-aperture flux densities (f_aper) to automatic-aperture flux densities (f_auto) by multiplying the median ratio between FLUX_AUTO and FLUX_APER. FLUX_APER misses about 5%–10% of the flux density, and therefore we multiplied by the factor 1.06, the same value adopted by Yang et al. (2014).²⁴

3.3.3. Photometric Uncertainties

IRAC images have PSFs with FWHM around 18 (IRAC Instrument Handbook²⁶ ), larger than the visible and NIR bands. Therefore they require different procedures for deblending objects and determining the fluxes accurately. Several methods have been developed for performing IRAC flux measurements. In this work, we used the algorithm IRACLEAN (Hsieh et al. 2012). The method was further improved (Laigle et al. 2016) to obtain more accurate fluxes for blended objects with large flux differences and separations less than one FWHM. Unlike other methods that first deconvolve high-resolution images with high-resolution PSFs and then convolve with the low-resolution PSFs, the IRACLEAN approach merely deconvolves the IRAC low-resolution images with the IRAC PSFs and adopts source positions and surface brightnesses as priors to derive the IRAC fluxes. Without requiring identical morphologies for each object (as other methods do), this method minimizes the effect of morphological k-correction (see Hsieh et al. 2012 and Laigle et al. 2016 for details).

In the IRACLEAN procedure, we used the high-resolution WIRCam zJHK_s χ² image as a prior for the low-resolution IRAC photometry. The IRAC flux densities were registered to the associated detections in the stacked-zJHK_s segmentation map generated by SExtractor. The uncertainty was estimated from fluctuations in the local area for each object in the residual map, taking the nonindependence of mosaic pixels into account.

We used simulations, performed separately for each WIRCam filter, to characterize the completeness functions of the WIRCam observations. Artificial stars of different magnitudes were generated and added to the reduced WIRCam images. Then we applied the same source extraction technique to detect and extract photometry. These artificial stars were recovered or missed depending on their flux, local noise, crowding, etc. Thus the completeness of photometry in each filter was evaluated as a function of magnitude based on the fraction of recovered artificial stars input in the reduced image (Figure 5). The limiting magnitudes corresponding to 95% completeness are 24.4, 23.7, and 23.7 in the J, H, and K_s bands, respectively.

Zoom In Zoom Out Reset image size

In fact, most of our detections are extended sources. To be more realistic, for J and H bands, we compared the number density of our detections with the deeper HST/WFC3 observation from the 3D-HST catalog (Skelton et al. 2014) to deduce the completeness (Figure 6). We assumed that the 3D-HST catalog is complete at magnitudes less than 26 in J and H bands, and our catalog reaches 99% completeness in J and H bands at 24.5 and 24.2 mag, respectively. Indeed the number densities of our work at some points are higher than the 3D-HST catalog. This could be due to the different filter bandpasses. For K_s bands, however, no public image is deep enough for us to do this measurement in this field. Therefore we extrapolated from the K_s-band number density to estimate 99% completeness at 23.9 mag.

Zoom In Zoom Out Reset image size

4.1.1. Galaxy Templates

4.1.2. Photo-z Results

The photo-z result using 12 bands (i.e., U, B, V, R, I, z, y, J, H, K_s, 3.6, and 4.5 μm) is shown in Figure 8. We achieved an overall accuracy σ_NMAD = 0.036 with an outlier fraction η = 7.3% from the comparison between photo-z and spec-z. For the bright (R < 23) sources, we obtained σ_NMAD = 0.025 and η = 1.58%. For low-redshift (z < 1.0) sources, we obtained σ_NMAD = 0.034 and η = 4.0% (Table 4).

Zoom In Zoom Out Reset image size

Table 4.  Photo-z Quality for Non-X-Ray and X-Ray Sources

	Non-X-Ray Sources				X-Ray Sources
	N	$\overline{{b}_{z}}$	σNMAD	η(%)	N	$\overline{{b}_{z}}$	σNMAD	η(%)
Total	2861	−0.013	0.036	7.34	352	−0.003	0.040	10.51
R < 23	947	−0.011	0.025	1.58	229	0.001	0.036	6.99
R > 23	1914	−0.014	0.044	10.19	123	−0.012	0.054	17.07
z < 1.0	1893	−0.013	0.034	4.07	224	0.001	0.035	5.36
z > 1.0	967	−0.012	0.042	13.75	128	−0.011	0.054	19.53

Download table as:  ASCIITypeset image

There are several possible reasons for photo-z outliers. The first is uncertain photometry, including faint or blended objects. (See two examples in Figure 9). Second, outliers could be due to the absence of representative templates in the library. For instance, Onodera et al. (2012) found that photo-zs in the Ilbert et al. (2009) sample are often underestimated for quiescent galaxies at 1.5 < z < 2. Third, misinterpreting a spectral break as Lyman break or Balmer break could lead to a catastrophic failure of photo-z (Dahlen et al. 2010). Fourth, the broadband data points may miss or misfit emission lines in the SEDs. In this case, as demonstrated by Ilbert et al. (2009) and Salvato et al. (2009), medium- or narrow-band data can help to pinpoint the emission line features and significantly improve the photo-z quality. Due to the lack of medium- or narrow-band data in this work, it is not surprising that we are not able to achieve the same photo-z qualities as presented in other deep fields such as COSMOS and CDFS. The upcoming Survey for High-z Absorption Red and Dead Sources (SHARDS; Pérez-González and Cava 2013) will help to improve the photo-z quality in the GOODS-N region.

Zoom In Zoom Out Reset image size

Because the majority of X-ray detections are associated with AGNs, we cannot use pure galaxy templates to fit their SEDs. A separate set of AGN–galaxy hybrid templates is required.

4.2.1. Cross-matching

The first step in computing photo-zs for X-ray sources was to cross-match our zJHK_s detections to the 2Ms-CDFN X-ray catalog (Xue et al. 2016). The X-ray source positions used K_s-band images from Wang et al. (2010) as the reference frame, which is the same as ours. The median offset between the X-ray positions and our zJHK_s positions is 02. This value is smaller than the median value of X-ray positional uncertainty (06). Therefore, we adopted a simple search radius of 1'' to match our zJHK_s detections with X-ray sources. 602 X-ray sources have zJHK_s detections within this search radius. Xue et al. (2016) used likelihood-ratio matching to identify the visible/near-infrared/midinfrared/radio (ONIR) counterparts. We compared their positions to those of our zJHK_s counterparts. When more than one zJHK_s detection was found within the search radius, we chose the one closest to the ONIR counterpart. zJHK_s counterparts within 1'' radius are labeled "xflag = 1" in the catalog.

For the remaining 81 X-ray sources lacking counterparts within the 1'' search radius, we checked their multiband cutouts visually. 56 of them have zJHK_s detections nearby that are consistent with the ONIR counterparts from Xue et al. (2016), as shown in Figure 10. In our released catalog, we treated these 56 sources as the zJHK_s counterparts of the X-ray sources and provided photo-z measurement of them. The photo-zs of these 56 sources should be used with caution because of the large separation as shown in Figure 10. These sources are labeled "xflag = 2" in the catalog.

Zoom In Zoom Out Reset image size

In total, 25 out of 683 X-ray sources have no zJHK_s counterparts. With visual inspection, we group these 25 sources into three cases:

• 1.  
  Case 1: 8 sources that are only detected by IRAC (Figure 11(a)).
• 2.  
  Case 2: 13 sources that have no zJHK_s or IRAC detections around the X-ray detection (Figure 11(b)).
• 3.  
  Case 3: 4 sources that have highly extended zJHK_s sources nearby. These sources have a large offset to the X-ray source position and therefore are not likely counterparts (Figure 11(c)).

Zoom In Zoom Out Reset image size

In Case 1 above, the IRAC sources can be considered the correct identifications for the X-ray sources. However, because Case 1 has only IRAC photometry, we cannot compute photo-zs for these eight sources. In Case 2, 11 of them have either no zJHK_s or ONIR counterparts around the X-ray detection, whereas the remaining two X-ray sources have ONIR counterparts in bluer bands but no zJHK_s detections. For Case 3, the extended zJHK_s sources are likely foreground galaxies that hide the true counterpart. We have excluded objects in all three cases above from the subsequent analysis for the X-ray sources.

4.2.2. AGN–Galaxy Hybrid Training

4.2.3. Photo-z Results

The UV bump from an AGN accretion disk is a significant spectral feature for distinguishing an AGN from a nonactive galaxy. Therefore, UV data are particularly important for obtaining accurate photo-zs for AGNs, as demonstrated by Hsu et al. (2014). 222 out of 683 X-ray sources are detected in either FUV or NUV. For these sources, we included UV data in addition to the same 12 bands used for the non-X-ray objects when computing photo-zs. As shown in Table 4 and Figure 12, we achieved an overall σ_NMAD for X-ray counterparts almost as good as that for non-X-ray sources. The outlier fraction was higher, especially for fainter sources and those at z > 1. To demonstrate the benefit of UV data, we compared photo-z results with and without UV data for 174 X-ray sources that have spec-z. Figure 13 shows that σ_NMAD is almost unchanged, but the outlier fraction η decreases from 9.8% without UV data to 6.3% when UV data are added.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The causes of the catastrophic failures of photo-zs for the X-ray sources are more complicated than for the non-X-ray sources. In addition to the reasons discussed in Section 4.1.2, the outliers could also be due to variability (Salvato et al. 2009) and to the uncertain contributions from AGNs.

In our released catalog, for both X-ray and non-X-ray sources, z_p is defined by the peak value of the redshift probability distribution function p(z). Figure 14 shows three examples of our results. In some cases (e.g., the object [HLD 2018]=47637 in Figure 14), there are two or more peaks in the probability distribution function. In such situations, we defined photo-z as the highest peak of p(z).

Zoom In Zoom Out Reset image size

Yang et al. (2014) used the code EAzY (Brammer et al. 2008) to estimate photo-zs for sources in the field. The main differences between our work and theirs are as follows: (1) The templates used for SED fitting. Yang et al. (2014) used only eight galaxy templates for non-X-ray sources and added three QSO templates for X-ray sources. We used 31 galaxy templates for non-X-ray sources and 78 galaxy or galaxy–AGN hybrid templates for X-ray sources. (2) The integration time of the NIR images. The NIR data we used are 0.5–1.0 mag deeper. (3) The X-ray catalog. Yang et al. (2014) used the Alexander et al. (2003) X-ray catalog, whereas we used the one from Xue et al. (2016). (4) UV data. We added UV data to compute photo-zs for the X-ray sources. For 2842 non-X-ray sources with spec-zs in common between our work and Yang et al. (2014), the two studies achieved comparable quality (σ_NMAD = 0.036, η = 7.1%). At z < 1, both works have similar error distributions as shown in Figure 15. At z > 1, our work is more peaked near $\tfrac{{\rm{\Delta }}z}{1+{z}_{s}}=0$ and has smaller scatter (σ_NMAD = 0.042), whereas Yang et al. (2014) tended to underestimate photo-zs and had slightly larger scatter (σ_NMAD = 0.048).

Zoom In Zoom Out Reset image size

For the X-ray sources, Xue et al. (2016) adopted two main photo-z catalogs (Skelton et al. 2014; Yang et al. 2014). Table 5 lists the comparisons for the 269 sources in common among all three works. Our photo-z accuracy is similar to that of Yang et al. (2014). The accuracy of the 3D-HST catalog (Skelton et al. 2014) is better than the other two studies for both X-ray and non-X-ray sources because Skelton et al. (2014) used the extremely deep visible and NIR data from HST. However, Skelton et al. (2014) fitted the X-ray sources with normal galaxy templates, and that gave a larger outlier fraction (11.9%) compared to ours (8.6%). Figure 16 compares our $\tfrac{{\rm{\Delta }}z}{1+{z}_{s}}$ distribution with the two other studies. Although Yang et al. achieved a slightly lower outlier fraction than we did, their $\tfrac{{\rm{\Delta }}z}{1+{z}_{s}}$ distribution for X-ray sources has a negative offset (see the upper panel in Figure 16), which was also mentioned by Xue et al. (2016). This is probably caused by systematic errors. The offsets of $\tfrac{{\rm{\Delta }}z}{1+{z}_{s}}$ are quantified by $\overline{{b}_{z}}$ in Table 5.

Zoom In Zoom Out Reset image size

Table 5.  Comparison of Our X-Ray-source photo-z Quality with Recent Work

		This Work			Yang et al. (2014)			Skelton et al. (2014)
	N	$\overline{{b}_{z}}$	σNMAD	η(%)	$\overline{{b}_{z}}$	σNMAD	η(%)	$\overline{{b}_{z}}$	σNMAD	η(%)
Total	269	−0.003	0.039	8.55	−0.014	0.040	7.43	0.001	0.032	11.90
R < 23	170	0.000	0.036	4.12	−0.016	0.035	2.94	0.002	0.030	7.06
R > 23	99	−0.011	0.055	16.16	−0.009	0.053	15.15	−0.002	0.042	20.20
z < 1.0	169	−0.001	0.037	3.55	−0.015	0.037	4.73	0.005	0.030	5.33
z > 1.0	100	−0.007	0.053	17.00	−0.012	0.048	12.00	−0.008	0.043	23.00

Download table as:  ASCIITypeset image