THE BLANCO COSMOLOGY SURVEY: DATA ACQUISITION, PROCESSING, CALIBRATION, QUALITY DIAGNOSTICS, AND DATA RELEASE

A common feature of multi-CCD cameras, such as the Mosaic2 imager, is cross talk among the signals from otherwise independent amplifiers or CCDs. This leads to a CCD image containing not only the flux distribution that it collected from the sky, but also a low-amplitude version of the sky flux distributions that appear in other CCDs. The cross-talk correction equation is described by

$$\begin{equation} I_{i} = \sum _{j=1}^N \alpha _{ij} I_\mathrm{raw}^{j}, \end{equation} \tag{ 1 }$$

where I_i denotes the cross-talk-corrected image pixel value in the ith CCD, α_ij denotes the cross-talk coefficients, and I^j_raw is the raw image pixel value. We used cross-talk coefficients provided by NOAO through the survey. As part of the cross-talk correction stage, the raw image (which contains 16 extensions) is split into one single-extension file per CCD. The processing and calibration of CCD mosaics can proceed independently for each CCD after the cross-talk correction, and therefore we split the images to enable efficient staging of the data to the compute resources.

Detrending is the process that removes the instrumental signatures from the images. Detrending, in this context, includes overscan correction, bias subtraction, flat fielding, pixel-scale correction, and fringe and illumination correction. Both the overscan correction and bias correction are required to remove the bias level present in the CCD and any residual, recurrent structure in the DC bias. Overscan correction is done for all raw science and calibration images. We subtract the median pixel value in the overscan region in each row for both the amplifiers in each CCD from the raw image pixel values after the cross-talk correction stage.

The median bias frame is created using nightly bias frames taken during the late afternoon, and subtracted from the nightly data. The flat-field correction is typically derived from dome flats taken for each observing band. The input dome flat images are overscan corrected, bias corrected, and then scaled to a common mode and then median combined. The resulting flat-field correction is scaled by the inverse of the image mode, creating a correction with a mean value of about unity. For the bias correction and the flat correction, the variation among the input images is used to create an inverse variance weight map that is stored as a second extension in the correction images. The creation of correction images also requires a bad pixel map, which is an image where pixels with poor response or with high dark current are masked and excluded from the images. These bad pixel maps are created initially using bias correction and flat-field correction images to identify the troublesome pixels.

The bias and flat-field corrections are then applied to the science images to remove pixel-to-pixel sensitivity variations. These corrections are only applied to those science pixels that are not masked. In this process, each science image receives an associated inverse variance weight map that encodes the Poisson noise levels and Gaussian propagated noise from each correction step on a per pixel basis. In addition, each science image has an associated bad pixel map (short integer) where a bit is assigned to each type of masking (i.e., pixels masked from the original bad pixel map, or masked due to saturation, cosmic ray, etc.). In our data model, the science image has three extensions: image, weight map, bad pixel map. Each measured flux at the pixel level comes along with its statistical weight and a history of any masking that has been done on that pixel.

For the Mosaic2 imager which has significant focal plane distortion, the pixel scale varies significantly over the field, leading to a significant trend in delivered pixel brightness as a function of position even with a flat input sky. For such detectors flattening the sky introduces a photometric non-flatness to the focal plane. Typically this pixel-scale variation is corrected during the process of remapping to a portion of a tangent plane, but in our case we prefer to do the single-epoch cataloging on images that do not suffer from correlated noise. Therefore, we apply a pixel-scale correction to account for variation of pixel response as a function of x and y position for each CCD. We first created master template images to determine photometric flatness corrections using astrometrically refined images from the Mosaic2 camera that we use to calculate the solid angle of each pixel. The correction image is then normalized by the median value, providing a flat-field-like correction image that can be used to bring all pixels to a uniform flux sensitivity. To avoid reintroducing trends in the sky with this correction, we apply this correction only to the values of each pixel after subtraction of the modal sky value. Effectively, this correction scales only source flux while maintaining a flat sky.

Illumination and fringe corrections are derived from fully processed science observations in a particular band. These can be from a single night or shared across nights. Usually if there was only one exposure from a given band in a night, we use science observations from neighboring nights to create the illumination and fringe correction images. Illumination corrections are done for all images, but fringe corrections on the Mosaic2 camera are needed only for i and z bands. To create these correction images, we first create sky flat templates. This requires a process of stacking all the detrended images in a band–CCD combination after first flagging all pixels contaminated by source flux. Source-contaminated pixels are determined by applying a simple threshold above background with a variable grow radius so that all neighboring pixels of a pixel determined to contain source flux are also masked. Modal sky values are then calculated for each image using pixels that are not flagged for any reason (object pixels, hot column, saturated, interpolated, etc.). The reduced images are then scaled to a common modal sky value, median combined, and then rescaled to a unit modal value.

This science sky image then contains a combination of any illumination and fringe signatures that are common to the input images. To create the illumination correction, we adaptively smooth the science sky images with a kernel that is large in the center and grows smaller near the edges. This effectively averages out the effects of any fringing, leaving an illumination correction image behind. The fringe correction is then produced by first differencing the science sky image and the illumination correction image, leaving behind an image of the small-scale structure (i.e., fringe signature) that is common to all the science images. This fringe correction image is then scaled by the modal value of the science flat image to produce a fractional fringe correction image.

The illumination correction image is applied like a flat-field correction to all previously corrected images, thereby removing any trends that are introduced by the differences in illumination of the dome flats and the flat sky. The fringe correction is applied by first scaling the correction image by the modal value of the sky in the science image and then subtracting it. The results of these two corrections are visually very impressive. The fringe effects in i and z bands are nicely removed in almost all cases. We have found some problem images where the fringe correction leaves clearly visible fringe signatures, and these are cases where only a few frames in i or z were taken on a particular night, and the use of images from neighboring nights to create the corrections was not adequate.

We expect that the residual scatter we measure could be further reduced using a star flat technique to better characterize the non-uniformities in the pupil ghost. Nevertheless, the delivered data quality from our current flattening prescription produces data that meet our data quality requirements. We note that the same prescription has been used previously to meet the data quality requirements of the SuperMACHO experiment in the processing of Mosaic2 data.

At the end of this series of image detrending steps which includes overscan, bias, flat field, pixel-scale, and illumination and fringe corrections, the pipeline creates eight images (one for each CCD) for every science exposure. These single-epoch image FITS files are called red images, and they contain three extensions: the main image, a bad pixel mask (BPM), and an inverse variance weight image. The BPM contains a short integer image where any unusable pixels have non-zero values (coded according to the source of the problem). The weight map is an inverse variance image map that tracks the noise on the pixel scale and where the weight is set to zero for all masked pixels.

Besides pointing errors, wide-field imagers exhibit instrumental distortions that generally deviate significantly from those of a pure tangential projection. In addition, the vertical gradient of atmospheric refractivity creates a small image flattening of the order of a few hundredths of a percent (corresponding to a few pixels on a large mosaic), with direction and amplitude depending on the direction of the pointing. These three contributions are modeled in the SCAMP (Bertin 2006) package that we use for astrometric calibration. SCAMP uses the TPV distortion model,¹⁸ which maps detector coordinates to true tangent plane coordinates using a polynomial expansion.

SCAMP is normally meant to be run on a large set of SExtractor catalogs extracted from overlapping exposures together with a reference catalog, in order to derive a global solution. However, since our pipeline operates on an image-by-image basis, we proceed in two steps: we first run SCAMP once on a small subset of catalogs extracted from BCS mosaic images to derive an accurate polynomial model of the distortions where the distortions in the R.A./decl. tangential plane are expressed as a third degree polynomial function of the CCD x/y position. This Mosaic2 distortion map modeled using a third-order polynomial per CCD for a BCS exposure is shown in Figure 5. The astrometric solution computed in this first step of calibration is based on a set of overlapping catalogs from dithered exposures which provides tighter constraints on nonlinear distortion terms (than catalogs taken individually). Using this model, we create a distortion catalog that encodes the fixed distortion pattern of the detector. We then run SCAMP on catalogs from each individual exposure (i.e., the union of the catalogs from each of the eight single-epoch detrended images), allowing only linear terms (two for small position offsets and four for the linear distortion matrix) describing the whole focal plane to vary from exposure to exposure. The solutions for the World Coordinate System (WCS) including the TPV model parameters are then inserted back into image headers. This approach capitalizes on the expected constancy of the instrumental distortions over time.

Zoom In Zoom Out Reset image size

We use the USNO-B1 (Monet et al. 2003) catalog as the astrometric reference. For astrometric refinement, the cataloging is done using SExtractor, and using WINdowed barycenters to estimate the positions of sources.

The astrometric accuracy is quite good, as can be demonstrated with the BCS co-adds. First, the accuracy is at the level of a fraction of a PSF or else significant PSF distortions would appear in the co-adds, and this is not the case. Second, we can measure the absolute accuracy relative to the calibrating catalog USNOB by probing for systematic offsets in R.A. or decl. between our object catalogs and those from the calibration source. Figure 6 shows the distribution of median offsets within all the co-add tiles for both R.A. and decl. The mean of the histograms is 00104 in R.A. and 00084 in decl., and the corresponding rms scatter is 47 mas and 45 mas, respectively. The USNOB catalog itself has an absolute accuracy with characteristic uncertainty of 200 mas (Monet et al. 2003), which then clearly dominates the astrometric uncertainty of our final catalogs.

Zoom In Zoom Out Reset image size

To catalog all objects from single-epoch images, we run SExtractor using PSF modeling and model-fitting photometry. A PSF model is derived for each CCD image using the PSFEx package (Bertin 2011). PSF variations within each CCD are modeled as an Nth degree polynomial expansion in CCD coordinates. For our application, we adopt a 26 × 26 pixel kernel and follow variations to third order. An example of variation of the FWHM of the PSF model across a single-epoch image is shown in Figure 7. The FWHM varies at the 10% level across this CCD due to both instrumental and integrated atmospheric effects.

Zoom In Zoom Out Reset image size

A new version of SExtractor (version 2.14.2) uses this PSF model to carry out PSF-corrected model-fitting photometry over each image. The code proceeds by fitting a PSF model and a galaxy model to every source in the image. The two-dimensional modeling uses a weighted χ² that captures the goodness of fit between the observed flux distribution and the model and iterates to a minimum. The resulting model parameters are stored and “asymptotic” magnitude estimates are extracted by integrating over these models. This code has been extensively tested within the DESDM program on simulated images, but the BCS data provide the first large-scale real world test. For the BCS application, we adopt a Sérsic profile galaxy model that has an ellipticity and orientation. This model fitting is computationally intensive and slows the “lightning-fast” SExtractor down to a rate on the order of 10 objects s⁻¹ on a single core. The SExtractor config file detection parameters are shown in Table 1.

The advantages of model-fitting photometry on single-epoch images that have not been remapped are manifold. First, pixel-to-pixel noise correlations are not present in the data and do not have to be corrected for in estimating measurement uncertainties. Second, unbiased PSF and galaxy model-fitting photometry is available across the image, allowing one to go beyond an approximate aperture correction to aperture magnitudes often used to extract galaxy and stellar photometry. Third, there are morphological parameters that can be extracted after directly accounting for the local PSF, which allows for improvements in star–galaxy classification and the extraction of PSF-corrected galaxy shear. A more detailed description of these new SExtractor capabilities along with the results from an extensive testing program within DESDM will appear elsewhere (E. Bertin et al., in preparation).

From the WCS parameters which are computed for every reduced image, one can approximate the footprint of the CCD on the sky using frame boundaries in R.A. and decl. For the BCS survey, we have a pre-defined grid of 36′ × 36′ tangent plane tiles covering the observed fields. Based on this, for every red image which is astrometrically calibrated, we determine which tiles it overlaps. We then use SWarp (Bertin et al. 2002) to produce background-subtracted remapped images that conform to sections of these tangent plane tiles. A particular red image can be remapped to up to four different remap images in this process. Pixels are resampled using Lanczós-3 interpolation.

Remapping also produces a pixel weight map and we also remap the bad pixel map (using nearest neighbor remapping). In this process of remapping, zero weight pixels in the reduced images generically impact multiple pixels in the remap image given the size of the interpolation kernel. These remaps are then stored for later photometric calibration and co-addition. This on-the-fly remapping need not be done, because at a later stage of co-addition one could in principle return to the red images, but given the PSF homogenization we do prior to co-addition we have found it convenient to do the remapping as we are processing the nightly data sets.

Our initial strategy for photometric calibration involved traditional photometric calibration using the standard fields observed on photometric nights along with the image overlaps to create a common zero point across all of our tiles. In fact, within DESDM we have developed a so-called Photometric Standards Module (PSM; Tucker et al. 2007) that we use to fit for nightly photometric solutions, and then we apply those solutions to all science images and associated catalogs from that night. For BCS this involves determining the zero points of all images on photometric nights through calibration to identified non-variable standard stars from the SDSS Stripe 82 field (Smith et al. 2002).

This procedure was used for processing and calibration of the BCS data processed in Spring 2008. But closer analysis of these data showed that we were not able to control photometric zero points to the required level to allow for cluster photometric redshifts over the full survey area. We therefore abandoned this method for BCS in favor of relative photometric calibration using common stars in overlapping red images followed by absolute calibration using the stellar locus (described in more detail in Section 3.2.5). One problem we faced is that so-called photometric nights exhibited non-photometric behavior in the standard field observations. There was no reliable photometric monitor camera at CTIO during our survey, and so observers simply used the time honored tradition of watching for clouds to make the call on a night being photometric. Because of our strategy for standard star observations (beginning, middle, end of night), even those nights that exhibit consistent photometric solutions need not have been photometric over the full nights. Therefore, we felt it safer to assume that no night was truly photometric and to calibrate the data using an entirely different approach.

The results from the PSM module for those nights exhibiting good photometric solutions are still useful. They have allowed us to monitor changes in the detectors and to measure the color terms in transforming our photometry onto the SDSS system. We provide a brief description of this procedure, although no science results in this paper are based on PSM-related direct photometric calibration. We expect to apply this method for absolute photometric calibration of DES data where we will indeed have an IR photometric monitoring camera on the mountain. The PSM solves for the following equation:

$$\begin{equation} m_\mathrm{inst} - m_\mathrm{std} = a_n + b_n \times ({\rm stdColor} - {\rm stdcolor}_0) + kX, \end{equation} \tag{ 2 }$$

where a_n is the photometric zero point for all eight CCDs, b_n is the color term, stdColor is the fiducial color around which we define our standard solutions, which is g − r for g and r bands, and r − i for i and z bands, stdcolor₀ is a constant equal to g − r = 0.53 for g and r bands and r − i = 0.09 for i and z bands, k is the first-order extinction coefficient, and X is the airmass. The PSM module solves for a_n, b_n, and k for each photometric night. Using these values for the PSM a, b, and k, one can also estimate the expected zero point for every exposure. We calculate it as follows:

$$\begin{equation} {\rm ZP} = -a + 2.5 \log \mathrm{(exptime)} -kX. \end{equation} \tag{ 3 }$$

We applied the PSM on about 30 BCS nights which were classified as photometric. We also checked for trends in variation of color terms as a function of CCD number. Only the i-band color term shows some variation, and this approximate constancy of color terms greatly simplifies the co-addition of the data, because we do not have to track which CCDs have contributed to each pixel on the sky. The color terms we have used for photometric calibration are −0.1221, −0.0123, −0.1907, and 0.0226 in griz, respectively. We also examine the band-dependent extinction coefficient (k) calculated using data from the photometric nights. For the ensemble of about 30 photometric solutions in each band, we find the median griz extinction coefficients at CTIO over the life of the survey to be 0.181, 0.104, 0.087, and 0.067 mag airmass⁻¹, respectively.

This completes the description of all the steps of the nightly processing or single-epoch processing that we do for BCS.

During single-epoch processing we extract instrumental magnitudes. To produce science ready catalogs, we must calculate the zero point for every image and re-calibrate the magnitudes. The photometric calibration is done in two steps. The first step is a relative zero-point calibration that uses the same object in overlapping exposures, and the second is an absolute calibration using the stellar locus.

The relative calibration is done tile by tile rather than simultaneously across the full survey. We use two different pieces of information to calculate the relative zero points. The primary constraint comes from the average magnitude differences from pairs of red images with overlapping stars. The stars are selected based on the SExtractor flags and spread_model (discussed later in Section 3.2.4) values. In cases where there are not enough overlapping stars, we use the average CCD to CCD zero-point differences derived from photometric nights. In previous versions of the reduction, we also used direct zero points derived from photometric nights (see Section 3.1.6) and relative sky brightnesses on pairs of CCDs. As previously mentioned, the direct photometric zero-point information is contaminated at some level. The sky brightness constraints also seem to be problematic for BCS, because only g- and r-band data were taken on dark nights with no moon present which can introduce a gradient across the camera. To avoid a degradation of the calibration, we used neither the sky brightness constraints nor the direct photometric zero points.

We determine the zero points for all images in a tile by doing a least-squares solution using the inputs described above. For this least-squares solution there are N input images, each with an unknown zero point in the vector z. We arbitrarily fix the zero point for one image and calibrate the remaining images relative to it. We have M different constraints in the constraint vector c. The matrix A is N × M and denotes the images involved for each constraint. The resulting system of equations is described by Az = c, where we use singular value decomposition to solve for the vector z. This gives the relative zero points needed to co-add the data for a particular tile.

Combining images with variable seeing generically leads to a PSF that varies discontinuously over the co-added image. This affects star–galaxy separation and contributes to variation across the image in the completeness at a given photometric depth. The PSF accuracy could be quite poor in regions where there are abrupt changes to the PSF which would translate into biases in the photometry that would be difficult to track. The main steps involved in the process of PSF homogenization include: (1) modeling the PSF using PSFEx for all remap images contributing to a co-add tile, (2) choosing the parameters of the target PSF, (3) using PSFEx to generate the homogenization kernel, and (4) carrying out the convolution to homogenize all the remap images to a common PSF.

To reduce PSF variation, we processed our images to bring them to a common PSF within an image and from image to image within a co-add tile. To do this we apply position-dependent convolution kernels that are determined using power spectrum weighting functions that adjust the relative contributions of large-scale and small-scale power within an image in such a way as to bring the PSFs within and among the image samples into agreement. The target PSF is defined to be a circular Moffat function with the FWHM set to be the same as the median value of all input PSFs:

$$\begin{equation} \chi ^2 = \vert \Psi - \sum _{l}Y_l(x_i)\kappa _l\ast \Psi _\mathrm{median} \vert ^2, \end{equation} \tag{ 4 }$$

where Y_l are the elements of a polynomial basis in x−y. The target PSF is defined to be a circular Moffat function with the FWHM set to be the median FWHM of the input images. We imposed a cut on input image PSF FWHM < 1.6 arcsec. This selects only images with relatively good seeing. Images from each band are homogenized separately. The FWHM of the target PSF for all BCS tiles is shown in the bottom panel of Figure 8.

Zoom In Zoom Out Reset image size

Another price of homogenization is that noise is correlated on the scale of the PSF. While the noise is already correlated to some degree through the remapping interpolation kernel, PSF homogenization characteristically affects larger angular scales than does the remapping kernel. This leads to biases in photometric and morphological uncertainties, and can also affect initial object detection process in SExtractor. To address this within DES, we account for the noise correlations on two critical scales by producing two different weight maps. The first weight map is used to track the pixel-scale noise, and the second weight map is used to correct for the correlated noise on the scale of the PSF. The pixel-scale weight map is used by SExtractor in determining photometric and morphological uncertainties. The PSF scale weight map is used by SExtractor in the detection process. Extensive tests within DESDM have shown this approach to be adequate to produce unbiased photometric and morphological uncertainties and to enable unbiased detection of objects within co-adds built from homogenized images. These results will be presented in detail elsewhere. For the BCS processing, we used only a single pixel-scale weight map, tuned to return the correct measurement uncertainties within SExtractor.

We use SWarp to combine the PSF-homogenized images to build the co-add tile. Inputs include the relative flux scales derived from the calibration described in Section 3.2.1. We combine the homogenized remap images using the associated weight maps and BPM for each image. The values of the flux-scaled, resampled pixels for each image are then median combined to create the output image. This allows us to be more robust to transient features such as cosmic rays in the i and z bands where there are three overlapping images. Also, objects with saturated pixels in all single-epoch images will contain pixels that are marked as saturated in the co-add images as well. This ensures accurate flagging of objects with untrustworthy photometry during the co-add cataloging stage. The resulting output co-add image's size is 8192 × 8192 pixels or approximately 0.6 × 0.6 deg.

Figure 9 shows a map of the FWHM as a function of position over one homogenized co-add image. Variations are at the level of ∼1% over the co-add, as compared to the ∼10% variations that are typical for Mosaic2 across a single CCD (see Figure 7). The constancy of PSF as a function of a position ensures that the PSF model can be modeled accurately and that the PSF-corrected model-fitting photometry is unbiased. The PSF homogenization process also circularizes the PSF. Figure 10 shows the distribution of ellipticities for the Mosaic2 single-epoch images and co-added images (color coded by band). The single-epoch ellipticity varies up to 0.1 with a modal value around 0.02. By contrast, the ellipticity distribution of the BCS co-adds is peaked at a fraction of a percent with a median value of 0.001.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To catalog the objects from co-added images, we run SExtractor in dual-image mode with a common detection image across all bands. For BCS, we use the i-band image as the detection image, because it has three overlapping images so the cosmic-ray removal is good, and it is by design the deepest of the bands. We then run SExtractor using model-fitting photometry using this detection image and co-added image in each band. This ensures that a common set of objects are cataloged across all bands. In both single-epoch and co-addition cataloging, the detection criterion was that a minimum of 5 adjacent pixels had to have flux levels about 1.5σ above background noise. The full SExtractor detection parameters used for both co-added and single-epoch images are shown in Table 1. In all we catalog about 800 columns across four bands. However, for the public data release, we have released 60 columns from SExtractor per object. This full list can be found in Table 3. Most of the parameters are described in the SExtractor manual online. There are a few additional parameters which are not yet released in the public version of SExtractor. These include model magnitudes and a new star–galaxy classifier called spread_model, which is a normalized simplified linear discriminant between the best-fitting local PSF model (ϕ) and a slightly more extended model (G) made from the same PSF convolved with a circular exponential disk model with scale length = FWHM/16 (where FWHM is the full width at half-maximum of the PSF model). It can be defined by the following equation:

$$\begin{equation} {\tt spread\_model} = \frac{{\bf \phi ^T x}}{{\bf \phi ^T\phi }} - \frac{{\bf G^T x}}{{\bf G^T\phi }}, \end{equation} \tag{ 5 }$$

where x is the image vector centered on the source. The distribution of spread_model for BCS catalogs is discussed in Section 3.2.7. More details of spread_model will be described elsewhere (E. Bertin et al., in preparation).

Once all objects from the co-add are cataloged (in instrumental magnitudes), we proceed to obtain the absolute photometric calibration using Stellar Locus Regression (High et al. 2009). The principle behind this is that the regularity of the stellar main sequence leads to a pre-determined line in color–color space called the stellar locus. This stellar locus is observed to be invariant over the sky, at least for fields that lie outside the galactic plane. The constancy of the stellar locus has been used as a cross-check of the photometric calibration within the SDSS survey (Ivezić et al. 2007).

Absolute photometric calibration is done after the end of co-addition. We select star-like objects using a cut on the SExtractor spread_model parameter and magnitude error. We then match the observed stars to Two Micron All Sky Survey (2MASS) stars from the NOMAD catalog, which is a combination of USNOB and 2MASS data sets, and which have JHK magnitudes (Skrutskie et al. 2006). Color offsets are varied until the observed locus matches the known locus. Because the 2MASS magnitudes are calibrated with a zero-point accuracy at the ∼2% level, one can bootstrap the calibration to the other bands. The known locus is derived using the high-quality “superclean” SDSS–2MASS matched catalog from Covey et al. (2007). It consists of ∼300,000 high-quality stars with data in ugrizJHK. A median locus is calculated for each possible color combination in bins of g − i.

The fit is done in two stages. First, a three parameter fit is done to the g − r, r − i, i − z colors. Another fit is done using g − r and r − J where the shift in the g − r color is fixed from the first fit. The first fit provides an accurate calibration of the colors and the second fit fixes the absolute scale. The fit is done this way because only a fraction of the stars that overlap with 2MASS are saturated in all bands. We perform the stellar locus calibration for model and 3 arcsec aperture magnitudes, separately. The model magnitude calibration is then used to calibrate the other magnitudes in the catalog (except for the 3 arcsec magnitudes). The calibration of the 3 arcsec aperture magnitudes determines a PSF-dependent aperture correction for the mag_aper_3 magnitudes only. We have found these small aperture magnitudes to provide higher signal-to-noise colors for faint galaxies in comparison to mag_model and mag_auto.

An example of the stellar locus fits for one co-add tile is shown in Figure 11. Red points are the observed colors of the stars, and the blue line is the median SDSS–2MASS locus. The orthogonal scatter about the stellar locus for all three color combinations for BCS tiles is shown in Figure 12. The rms orthogonal scatter about the stellar locus in (g − r, r − i), (r − i, i − z), and (g − r, r − j) is 0.059, 0.061, and 0.075, respectively. Given the scatter and the number of stars available for calibration, we can determine the zero points in our bands with sub-percent accuracy.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To validate our photometric calibration algorithms, we applied exactly the same procedure to the full SDSS–2MASS catalog in Covey et al. (2007). This catalog includes noiser objects than the catalog we used to derive the median stellar locus. We selected four areas (between R.A. of 120° and 350°) and divided each into 1° × 1° patches. We match the objects to obtain 2MASS magnitudes and then apply the same calibration procedure as we did for the BCS catalogs. The rms scatter distributions for all three color combinations can be found in Figure 12 (bottom panel). The corresponding scatter for SDSS in (g − r, r − i), (r − i, i − z), and (g − r, r − j) is about 0.041, 0.035, and 0.05, respectively, and is about 1.5 times smaller than for the BCS catalogs. This is clear evidence for higher scatter in our stellar photometry as compared to the SDSS photometry. Assuming this additional source of scatter adds in quadrature with the SDSS observed scatter, we estimate the extra noise in BCS color combinations compared to SDSS is 0.039, 0.054, and 0.048 in (g − r, r − i), (r − i, i − z), and (g − r, r − j), respectively. Because these noise sources are getting contributions from each color, we can estimate that the noise floors are δ(g − r) ∼ 0.027, δ(r − i) ∼ 0.038, and δ(i − z) ∼ 0.038. These then imply noise floors in the stellar photometry within griz bands of approximately 1.9%, 2.3%, 2.7%, and 2.7%, respectively. This is then in good agreement with the typical repeatability scatter seen in these bands (see Figure 14) when one considers that g and r bands each have two overlapping exposures and i and z bands each have three.

Our current catalogs contain two star–galaxy classification parameters provided by SExtractor: class_star, which has been extensively studied and spread_model, which has been newly developed as part of the DESDM development program. In order to test their performance and range of magnitudes up to which these measures can be reliably used, we plot the behavior of these two classifiers in the i band as a function of $mag\_model$ in Figure 13. class_star lies in the range from 0 to 1. At bright magnitudes, one can see two sequences in class_star for galaxies and stars near 0 and 1, respectively. The two sequences begin merging as bright as i = 20 and are significantly merged beyond i = 22. As described in Section 3.2.4, spread_model uses the local PSF model to quantify the differences between PSF-like objects and resolved objects. In the spread_model panel, it is clear that there is a strong stellar sequence around the value 0.0, and that galaxies exhibit more positive values. The narrow stellar sequence and the broad galaxy sequence begin merging at i = 22 in the BCS, but there is significant separation in the two distributions of points down to i = 23. Along with spread_model comes a measurement uncertainty, and so it is, for example, possible to define a sample of objects that lie off the stellar sequence in a statically significant way. For the BCS data, a good cut to separate stars would be ${\tt spread\_model}<0.003$. Detailed studies of this new classification tool have been carried out within the DESDM project and will be carried out elsewhere.

Zoom In Zoom Out Reset image size

During the processing within the DESDM system, a variety of quality checks are carried out. These include, for example, thresholding checks on the fraction of flagged pixels within an image and the χ² and number of stars used in the astrometric fit of each exposure. In addition, the system is set up to report on the similarity between correction images (bias, flat, illum, and fringe) against stored templates that have been fully vetted. During the BCS processing this last facility was not used.

Our experience has been that problems at any level of processing are most likely to show up in the stages of relative and absolute photometric calibration. Therefore, for the BCS processing done here we capture a range of photometric quality tests including the number of stars used in the stellar locus calibration and the rms scatter about the true stellar locus of the calibrated data (see Figure 11). In addition, we examine the photometric repeatability for common objects within overlapping images contributing to each tile. In Figure 14, we show an example for the g band in tile BCS0549-5043. This shows the magnitude difference between pairs of overlapping objects versus the average magnitude. The scatter here includes both statistical and systematic contributions, and the envelope of scatter grows toward faint magnitudes, as expected. Outlier rejection is done on the point distribution, and all 3σ outliers are filtered out and colored red. In the top panel, we plot the mean and rms as well as the outlier fraction of these repeatability distributions as a function of magnitude. The mean and rms numbers are listed in millimagnitudes. Also, the statistical uncertainties of the model magnitudes are used to estimate the systematic magnitude error contribution to the rms. On the bright end where the statistical noise is very small, the systematic contribution to the rms is close to the total, which is 10 mmag in this case. As one moves toward the faint end, the statistical contribution increases and the estimated systematic contribution plays only a small role in explaining the scatter. This is just how we expect the photometry to behave.

Zoom In Zoom Out Reset image size

Repeatability plots indicate systematic contributions to the photometric errors at the 10–20 mmag levels for typical g- and r-band tiles. For i- and z-band tiles, the systematic noise is closer to 30–40 mmag. For all the BCS tiles, we have examined these repeatability and stellar locus plots to probe whether the scatter is in acceptable ranges. In cases where tiles did not meet these quality control tests, we worked on the relative and absolute photometric calibration to improve the data. In addition to these photometry tests, we examined the sky distribution of cataloged objects within each tile. In cases where large numbers of faint “junk” objects were found, we attempted to remove them by adjusting the cataloging. At present all our tiles meet these quality tests except for a handful of tiles that are marked in red in Figure 1. This includes four tiles in the 5 hr field and six tiles in the 23 hr field, corresponding to ∼4% of the 80 deg² region. Ideally, we would reimage these regions to obtain better data.

For every BCS night, the detrending pipeline creates three main types of science image files which we denote as raw, red, and remap. The co-add pipeline produces four co-add images per tile for each of the four bands. Once we have calibrated co-add catalogs for all the processed tiles, we run a post-processing program to remove duplicate objects near the edges of the tiles. This is necessary because there is a 2 arcmin overlap between neighboring tiles. The program selects sources that appear in neighboring tiles that lie within 0.9 arcsec radius and for each pair it keeps the object that lies farthest from the edge of its tile. In this way a single, science ready catalog is prepared for each field. The 23 hr field catalog contains 1,877,088 objects, and the 5 hr field contains 2,952,282 objects with i model magnitude <23.5. In the next section, we review additional tests of the data quality.