THE SEVENTH DATA RELEASE OF THE SLOAN DIGITAL SKY SURVEY

The SDSS imaging survey was primarily designed to give a single pass across the sky, thus in the CAS, each photometric measurement is flagged either Primary or Secondary. Primary objects designate a unique set of detections (i.e., without duplicates) using the geometric boundaries of survey stripes.¹⁰⁶ The set of Secondary objects includes repeat observations of the same object in overlapping strips and stripes. Primary objects are associated with a run and field which is the primary source of imaging data at that position. In DR7, the union of the Legacy and SEGUE footprints serves as the Primary footprint; a quantity inLegacy in the fieldQA table in CAS indicates those objects which lie within the original Legacy Northern Galactic Cap Survey ellipse, as defined in York et al. (2000). Legacy imaging can also be distinguished by the stripe number for each run; Stripes 9–44, 76, 82, and 86 are in the Legacy Survey, all others are SEGUE stripes or other miscellaneous pieces of sky (Figure 1).

While resolving the sky into a seamless Primary region of unique detections of objects is ideal for many science queries, it is sometimes convenient to query data by run without regard to the way the survey resolves overlaps and imposes the boundaries of the edge of the survey. These boundaries are restricted to matched pairs of North and South strips in the main DR7 CAS. Therefore in many runs, several fields at the beginning or end which do not have a match in the corresponding other strip are not included in the main CAS. Thus, we have now made available a separate runs database within the CAS, which includes all fields in all runs, and which allows one to query objects by which run they are imaged in.

The runs database contains 530 complete runs from SDSS-I and SDSS-II, where Primary is set strictly based on geometric limits within each scan, regardless of overlapping runs or stripes. The runs database also contains several scans outside the regular DR7 Legacy or SEGUE footprints. For example, Stripe 205 is covered by runs 4334, 4516, 6751, and 6794, and follows the Sagittarius Stream, which is in three pieces, the first running from (α,  δ) = (240°, − 15°) to (200°, + 10°), the second centered at (135°, + 35°), and the last (overlapping several other runs) which ends at (45°, + 10°).

The SDSS stripe along the celestial equator in the Southern Galactic Cap ("Stripe 82") was imaged multiple times in the Fall months. This was first carried out to allow the data to be stacked to reach fainter magnitudes, and through Fall 2004, these data were taken only under optimal seeing, sky brightness, and photometric conditions (i.e., the conditions required for imaging in the Legacy Survey; York et al. 2000). There were 84 such runs made public in previous data releases. In Fall 2005, 2006, and 2007, 219 additional imaging runs were taken on Stripe 82 as part of the SDSS supernova survey (Frieman et al. 2008), often under less optimal conditions: poor seeing, bright moonlight, and/or nonphotometric conditions. These data have been photometrically calibrated following the prescription of Bramich et al. (2008), whereby the photometry of bright stars is tied to that of photometric data on a field-by-field basis (see Ivezić et al. 2007 for a similar approach). Bramich et al. solved for photometric offsets both parallel and perpendicular to the scan direction in data from a given CCD; we found that the term perpendicular to the scan direction added little, and we did not include it here. As Bramich et al. (2008) show, the resulting photometric calibration is good to 0.02 mag at the bright end in up to 1 mag of atmospheric extinction. Of course, under nonoptimal conditions, these data will not necessarily reach as deep as normal survey images.

SDSS judges photometricity of a given night by monitoring fluctuations in the night sky measured by a wide area infrared camera (the "cloud camera") sensitive at 10 μm, where clouds are emissive (Hogg et al. 2001). If the sky fluctuations are small and constant, then the night is photometric. Clouds cause the fluctuations to increase. Plots of cloud cover and seeing for most nights on which Stripe 82 was observed are available as part of the DR7 web documentation listing all Stripe 82 scans. In addition, for those runs which the cloud camera indicated as nonphotometric, we examined the fluctuations in the zero point for each CCD in the camera as a function of time using the photometric calibration procedure of Bramich et al. (2008). These zero-point values are available in the CAS; rms variations of more than 0.1 mag are an indication of considerable variable cloud cover, and a value of more than 1 mag suggests that the approximate calibration procedure of Bramich et al. (2008) breaks down, and the resulting photometry should be regarded with caution. All 303 runs covering Stripe 82 are made available as part of the Stripe82 database, which is structured like the runs database.

We have carried out a co-addition of the repeat imaging scans on Stripe 82 taken through Fall 2005 under the best conditions (see below). The co-addition includes a total of 122 runs, covering any given piece of the >250 deg² area between 20 and 40 times (Figure 3), and the results are made available in the Stripe82 database as well. The co-addition runs are designated 100006 (South strip) and 200006 (North strip), respectively in the DAS, and 106 and 206 in the CAS.

The co-addition is described in detail in J. Annis et al. (2009, in preparation); see also Jiang et al. (2008). From the list of runs on Stripe 82 taken through the Fall 2005 season, all fields with seeing in the r band worse than 2'' FWHM, r-band sky brightness brighter than 19.5 mag in 1 square arcsecond, or whose photometric correction à la Bramich et al. (2008; see above) was greater than 0.2 mag were excised; this rejected 32% of the available data. The individual runs were remapped onto a uniform astrometric coordinate system. Interpolated pixels in each individual run (e.g., for bad columns, bleed trails, and cosmic rays) were masked in the co-addition process. The sky was subtracted from each frame, and the images co-added with weights for each frame proportional to the transparency and inversely proportional to the square of sky noise and seeing on each frame. Strongly discrepant pixels were clipped in the co-addition. The effective seeing FWHM is ∼12 (for the southern strip of the stripe) and ∼13 (for the northern strip).

The resulting co-added images were run through the SDSS photometric pipeline, yielding the catalog made available in the Stripe82 database. Rather than deriving the point-spread function (PSF) from scratch, we synthesized the PSF at each point in the sky by taking the suitably weighted sum of the PSFs output by the SDSS photometric pipeline from each of the individual runs.

Color–color diagrams of stars and counts of stars and galaxies as a function of magnitude demonstrate that the photometry reaches roughly 2 mag fainter than single SDSS scans, similar to what is expected given the number of runs in the co-add. We have found that star–galaxy separation is improved over that in the single scans, in that the cut can be made closer to the stellar locus. In the main survey, objects with m_PSF − m_model>0.145 are flagged as galaxies in a given band. However, the stellar peak in the PSF − model magnitude difference distribution in the co-add is much narrower, allowing objects with m_PSF − m_model>0.03 in r to be flagged as galaxies.

The co-addition does not properly propagate information on saturated pixels in individual runs, and therefore the photometry of objects brighter than roughly r = 15.5 is suspect. Unfortunately, there is no processing flag that one can use to identify such data; we recommend a simple magnitude cut.

The SDSS photometry is quoted in terms of asinh magnitudes, as described by Lupton et al. (1999), whereby the logarithmic magnitude scale transitions to a linear scale in flux density f at low S/N:

$$\begin{equation} m = -\frac{2.5}{\ln 10} \left[{\rm asinh}\left(\frac{f/f_0}{2\,b}\right) + \ln (b)\right]. \end{equation} \tag{ 1 }$$

The magnitude at which this transition occurs is set by the quantity b, which is roughly the fractional noise in the sky in a PSF aperture in 1'' seeing (EDR paper). Here f₀ = 3631 Jy, the zero point of the AB flux scale. The quantity b for the co-addition is given in Table 2, along with the asinh magnitude associated with a zero-flux object. Compare with the equivalent numbers for the main survey, given in Table 21 of the EDR paper. Table 2 also lists the flux corresponding to 10f₀b, above which the asinh magnitude and the traditional logarithmic magnitude differ by less than 1% in flux.

As with the main survey, it is important to use the various processing flags output by the photometric pipeline (e.g., as recommended by Richards et al. 2002) to reject spurious objects, and to select objects with reliable photometry.

As was noted in the DR6 paper, the SDSS imaging pipeline (photo) was designed to analyze data at high Galactic latitudes, and is not optimized to handle very crowded fields. The Legacy Survey is restricted to high latitudes, and photo performs adequately throughout the Legacy footprint. However, at lower latitudes, when the density of stars brighter than r = 21 grows above 5000 deg⁻², the pipeline is known to fail, as it is unable to find sufficiently isolated stars to measure an accurate PSF, and the deblender does poorly with overly crowded images. Many of the SEGUE scans probe these low latitudes (Figure 1), and we therefore adapted an alternative stellar photometry code called PSPhot developed by the Pan-STARRS team (Kaiser et al. 2002; Magnier 2006) to be used for these runs. In brief, we first run this code, and then run photo using the list of objects detected by PSPhot as input to help photo's object finder in crowded regions. This approach thus provides two sets of photometry at low latitudes.

Like, e.g., DAOPHOT (Stetson 1987), PSPhot begins with the assumption that every object is unresolved, and therefore does a better job than photo in crowded stellar regions. It uses an analytical model based on Gaussians to describe the basic PSF shape, with parameters which may vary across the field of the image to follow the PSF variations. It also uses a pixel-based representation of the residuals between the PSF objects and the analytical model, which is also allowed to vary across each field. Candidate PSF stars are selected from the collection of bright objects in the frame by searching for a tight clump in the distribution of second moments. After rejecting outliers, the PSF fit parameters are used to constrain the spatial variations in the PSF model.

Unlike photo, PSPhot processes each frame separately (without any requirement of continuity of PSF estimation across frame boundaries), and each filter separately (without any requirement that the lists of objects between the separate filters agree). The pipeline outputs positions and PSF magnitudes (and errors) for each detected object; the results are found in the PsObjAll table in the CAS. The resulting photometry is then matched between filters using a 1'' matching radius. While the estimated PSF errors output by photo include a term from the uncertainty in the PSF fitting, this component is not included in the errors reported by PSPhot.

We then run photo, asking it to carry out photometry at the position of each object detected by PSPhot, in addition to the positions of objects photo itself detects. This allows photo to do a much better job of distinguishing individual objects in crowded regions. In addition, the pipeline is fine tuned to less aggressively look for overlap between adjacent objects, and not to give up as soon as it does at high latitude when faced with deblending large numbers of objects. We describe below how the photometry directly out of PSPhot, and that from photo, compare.

The SDSS PSF photometry had an offset applied to it to make it agree with aperture photometry of bright stars within a radius of 743; this large-aperture photometry was in fact what was used by ubercalibration to tie all the photometry together (Padmanabhan et al. 2008). In crowded regions, finding sufficiently isolated stars to measure aperture photometry becomes difficult. PSPhot photometry was forced to agree with these large-aperture magnitudes for bright stars; this was done in practice by determining, for each CCD in the imaging camera for each run, the average aperture correction needed to put the two on the same system, using stars at Galactic latitude |b|>15°, where crowding effects are less severe.

If any part of a SEGUE imaging run extended to |b| < 25°, the entire run was processed through the photo and PSPhot code. This sample includes most (but not all) of the SEGUE imaging runs. These PSPhot+photo processed runs, designated with rerun = 648 in the DR7 CAS and DAS, are declared the Best reduction of these SEGUE runs. There is also an inferior Target version of these SEGUE runs which was used to design SEGUE spectroscopic plates; it is based on photo alone, as the PSPhot code was unavailable at the time the plates were designed. The Target reductions have rerun = 40 and are segregated to the SEGUETARGDR7 database.

This processing revealed a problem with photo. In crowded regions, one cannot find sufficiently isolated stars to measure counts through such a large aperture, and in practice, the code corrected PSF magnitudes to an aperture photometry radius of 300 instead, whenever any part of a given run dipped below |b| = 8°. Thus, the aperture correction was underestimated, typically by 0.03–0.06 mag, depending on the seeing. This was not a problem for any of the Legacy imaging scans, but is very much an issue for the SEGUE runs. Fortunately, there is a strong correlation, in a given detector, between the aperture correction from a 300 aperture to a 743 aperture (as measured on high-latitude fields), and the seeing. We therefore applied this correction after the fact to the photo PSF, de Vaucouleurs, exponential, and model magnitudes for all SEGUE runs affected by this problem. This was carried out before ubercalibration, so these runs are photometrically calibrated in a consistent way.

The quality of the photometry produced by PSPhot  and by photo with the PSPhot-detected objects as input, was evaluated by comparing the magnitudes computed by the two methods. Within each field, we calculated the median of the difference of PSF magnitudes for stars with 14 < u,  g,  r,  i,  z < 20. This median difference had an rms of 0.014 mag. Fields with a difference greater than 0.08 mag are suspect, and further investigation is needed to determine which of the two pipelines might be at fault. We followed McGehee et al. (2005) to measure reddening-free colors of the same stars that track the stellar locus:

$$\begin{eqnarray} {Q}_{{gri}}&=& (g - r) - {E}_{{gri}}\, (r - i), \\ {Q}_{{riz}}&=& (r - i) - {E}_{{riz}}\, (i - z), \end{eqnarray} \tag{ 2 }$$

where E_gri = 1.582 and E_riz = 0.987. These are normalized to equal zero at high Galactic latitude (note that these colors do not include the u band).

Median Q_gri and Q_riz colors in each field were computed for objects identified as stars in each field, and satisfying magnitude and color cuts as follows: 14 < (u,  g,  r,  i,  z) < 20, 0.5 < (u − g) < 1.9, 0.0 < (g − r) < 1.2, −0.2 < (r − i) < 0.8, and −0.2 < (i − z) < 0.6. The Q-parameters were found to be lower by up to 0.1 mag at low Galactic latitudes; to remove this effect, we fitted a model of a constant plus Lorentzian to the median Q values as a function of Galactic latitude, and subtracted it. The distributions of the Q_gri and Q_riz values for both photo and PSPhot are compared as density plots in Figure 4. From Equation 2, photometric errors in a single filter manifest themselves differently: δg as a shift in Q_gri, δr as a line with slope dQ_riz/dQ_gri = −1/(1 + E_gri) = −0.35, δi as a line with slope dQ_riz/dQ_gri = −(1 + E_riz)/E_gri = −1.07, and δz as a shift in Q_riz.

Zoom In Zoom Out Reset image size

The photo data in a given field were flagged as bad when either |Q_gri| or |Q_riz|>0.12 mag (≳5σ) as measured from photo magnitudes, and similarly for the PSPhot outputs. Of course, a field could be flagged as bad in both sets of outputs. By this criterion, about 2% of the fields processed with PSPhot were flagged bad based on the photo outputs, and 3.6% were bad based on PSPhot photometry. The vast majority of the flagged fields are within 15° of the Galactic plane, and essentially all the fields in which the median difference between photo and PSPhot photometry was greater than 0.08 mag in a given band were flagged as bad by the Q criteria. This flag and the Q_gri and Q_riz quantities themselves can be found in the fieldQA table in the CAS.

Although more fields are flagged based on the PSPhot outputs, the PSPhot scatter in Figure 4 is tighter at both high and low Galactic latitudes than for photo. The PSPhot stellar photometry is therefore preferred for studies of the stellar locus (we have not fully assessed its robustness to outliers), but comes with the caveat that fields flagged bad should be identified in the fieldQA table and be culled.

An alternative check of SDSS photometry in dense stellar fields was carried out by An et al. (2008), who reduced the SDSS imaging data for crowded open and globular cluster fields using the DAOPHOT/ALLFRAME suite of programs (Stetson 1987, 1994). At a stellar density of ∼400 stars deg⁻² with r < 20, they found ∼2% rms variations in the difference between photo and DAOPHOT magnitudes in the scanning direction in all five bandpasses (see their Figure 3). The systematic structures are likely due to imperfect modeling of the PSFs in photo, given that DAOPHOT magnitudes exhibit no such large variations with respect to aperture photometry. In other words, the PSF variations were too rapid for the photo pipeline to follow over a timescale covered approximately by one field (≈10' or ≈54 s in time).

An et al. (2008) further examined the accuracy of photo magnitudes in semicrowded fields using three open clusters in their sample. Stellar densities in these fields were as much as ∼10 times higher than those in the high Galactic latitude fields, but photo recovered ∼80%–90% of stellar objects in the DAOPHOT/ALLFRAME catalog. An et al. (2008) found that these fields have only marginally stronger spatial variations in photo magnitudes than those at lower stellar densities.

Section 4.6 of the EDR paper describes a series of flags available in the database to assess the quality of each field in the imaging data; this includes information on whether the data in a given field meet survey requirements on seeing and sky brightness. We have added additional criteria to assess the quality of each field. The CAS table called fieldQA includes a flag called ProblemChar associated with each field, which is set when:

• 1.  
  The median of the telescope focus over three frames moved more than 60 μm, indicating a problem with the automated focus of the telescope (Gunn et al. 2006; Problemchar = "f").
• 2.  
  The rotator angle moved more than 25'' between adjacent fields (corresponding to a 055 image shift at the edge of the camera) (ProblemChar="r").
• 3.  
  The astrometric solution shifted by more than 4 pixels (16) from a smooth interpolation between adjacent fields (ProblemChar="a").
• 4.  
  Miscellaneous other problems, including voltage problems in the camera, and lights left on in the telescope; this was triggered in only two imaging runs (ProblemChar="s").

We flag all fields with these problems in any of the five bandpasses. Because the imaging observations are done in drift-scan mode (Gunn et al. 1998), different areas of the sky are observed simultaneously in each bandpass and referenced to the field number of the r-band observation. Thus in the case of focus problems, we mark the 11 fields preceding, and the three fields following the field in question in all camera columns in the run as bad. For the rotator and astrometric shift problems, we similarly mark the nine preceding and the one following field as bad. Only 0.3% of all fields in the CAS are marked with one of these problems (the majority of which are due to focus problems); these flags should be consulted when examining the reliability of the photometry in a given area of sky.

Early SDSS imaging runs were astrometrically calibrated against Tycho-2 (Høg et al. 2000), which yielded statistical errors per coordinate for bright stars (r < 20) of approximately 75 mas and systematic errors of 20–30 mas. Later runs were calibrated against preliminary versions of the USNO CCD Astrograph Catalog (UCAC; Zacharias et al. 2000), which yielded improved statistical errors per coordinate of approximately 45 mas, with systematic errors of 20–30 mas (Pier et al. 2003). Proper motions were not available for the preliminary versions of UCAC. Since the typical epoch difference between the SDSS and UCAC observations is a few years and the typical proper motion of UCAC stars is a few mas year⁻¹, this introduces an additional roughly 10 mas of systematic error in the positions due to the uncorrected proper motions of the calibrating stars.

All of DR7 has been recalibrated astrometrically against the Second Data Release of UCAC (UCAC2; Zacharias et al. 2004). While the systematic errors for UCAC2 are not yet well characterized, they are thought to be less than 20 mas (N. Zacharias 2008, private communication). UCAC2 also includes proper motions for stars with δ < +41°. For stars at higher declination, proper motions from the SDSS+USNO-B catalog (Munn et al. 2004) have been merged with the UCAC2 positions. With these improvements, all DR7 astrometry has statistical errors per coordinate for bright stars of approximately 45 mas, with systematic errors of less than 20 mas. The mean differences per run between the old and new calibrations is a function of position on the sky, with typical absolute mean differences of 0–40 mas. The rms differences are of order 10–40 mas for runs previously reduced against UCAC, and 40–80 mas for runs previously reduced against Tycho-2, consistent with what we would expect given the errors in the reductions.

Note that the formal SDSS names of objects in the CAS are of the form SDSS Jhhmmss.ss±ddmmss.s. Because of the subtle changes in the astrometry, these names will be slightly different for many objects between DR6 and DR7. The user should be aware of this in comparing objects between DR6 and DR7.

The CAS includes proper motions for objects derived by combining SDSS astrometry with USNO-B positions, recalibrated against SDSS (Munn et al. 2004). These are given in the ProperMotions table in the CAS.¹⁰⁷ An error was discovered in the proper motion code in Data Releases 3 through 6, which causes smoothly varying systematic errors, in the proper motion in right ascension only, of typically 1–2 mas year⁻¹ (see Munn et al. 2007 for a full description of the problem and its effects). This error has been corrected in DR7, thus any use of proper motions should use the DR7 CAS.

Several of the SEGUE target selection algorithms evolved during the course of SDSS-II. The most significant changes occurred to the K-giant algorithm, as it was realized that good color-based luminosity separation could be done only for the very reddest (g − r>1.1) giant candidates by their deviation from the main-sequence locus in the ugr color diagram; this of course requires accurate u-band photometry. Early K-giant target selection included stars with (g − r)₀ (where the subscript refers to values after correcting for SFD Galactic extinction) as blue as 0.35. The final selection chose stars with (g − r)₀ between 0.5 and 1.2 and was restricted to g₀ < 18.5; this gives much cleaner samples of K giants (Yanny et al. 2009).

In order to allow users to analyze completeness and efficiency of SEGUE stellar target selection samples, the latest (v4.6) version of the algorithms (Yanny et al. 2009) was applied to all stellar objects in the imaging catalog which had g < 21 or z < 21, over the entire sky. The appropriate bits were propagated into the SEGUEPrimTarget and SEGUESecTarget fields of the photoObjAll table of the DR7 CAS. A description of the bits and the target selection algorithms is available in Yanny et al. (2009).

As described in the DR5 paper, the SDSS makes available the results of two different photometric redshift determinations for galaxies, one based on neural nets and the other based on a template-fitting approach. With DR7, we include improvements to both, as we now describe.

4.6.1. Photometric Redshifts with Neural Nets

4.6.2. Photometric Redshifts: A New Hybrid Technique

With DR7, we have made substantial improvements in the other photometric redshift code (Photoz), using a hybrid method that combines the template-fitting approach of Csabai et al. (2003; i.e., the approach used in DR5 and DR6) and an empirical calibration using objects with both observed colors and spectroscopic redshifts. We summarize the method briefly here, with details to follow in a paper in preparation.

The spectroscopic sample of SDSS contains over 900,000 spectroscopically confirmed galaxies, and the combination of the main sample (Strauss et al. 2002), the LRG sample (Eisenstein et al. 2001) and special plates targeted at fainter blue galaxies (DR4 paper) more or less cover the whole color region in which galaxies lie to the depths of SDSS. Thus, we use the DR7 spectroscopic set as a reference set for redshift estimation without any additional data from synthetic spectra.

The estimation method uses a k-d tree (following Csabai et al. 2007) to search in the ubercalibrated u−g, g−r, r−i, and i−z color space for the 100 nearest neighbors of every object in the estimation set (i.e., the galaxies for which we want to estimate redshift) and then estimates redshift by fitting a local hyperplane to these points, after rejecting outliers. If an object lies outside the bounding box of the 100 nearest neighbors in color space, the photometric redshift is less reliable, and the object is flagged accordingly. We use template fitting to estimate the K-correction, distance modulus, absolute magnitudes, rest frame colors, and spectral type. We search for the best match of the measured colors and the synthetic colors calculated from repaired (Budavári et al. 2000) empirical template spectra at the redshift given by the local nearest neighbor fit.

We have found that the mean deviations of the redshifts from the best-fit hyperplane is a good estimate of the error. That, together with the flag indicating whether an object lies outside the bounding box of its neighbors, and the difference between the estimated photometric redshift and the average redshift of its neighbors, can be used to select objects with reliable photometric redshift values.

The rms error of the redshift estimation for the reference set decreases from 0.044 in DR6 to 0.025 in DR7 with this improved algorithm (Figure 5). Iteratively removing the outliers beyond 3σ gives rms errors of 0.028 and 0.020 for the old and new methods, respectively. In addition, the reliability of the quoted errors is much higher.

Zoom In Zoom Out Reset image size

The response functions of the SDSS imager as a function of wavelength have been monitored throughout the survey. The griz responses were stable over time, although very small seasonal (i.e., temperature) variations were observed, at a level well below our typical photometric errors. However, we have found a relatively large change in both the amplitude and shape of the u-band response, which is likely due to a degradation of the UV-enhanced coating of the u-band CCD. This change in instrumental zero point is effectively corrected by the photometric calibration for objects near the mean color of the standard stars, and, in fact, the repeat photometry of stars in Stripe 82 is stable with time for stars with −0.5 < g − r < 1.5 (Bramich et al. 2008; Ivezić et al. 2007). However, the observed response changes involve a roughly 30 Å redward shift in the effective wavelengths of the u filters over the lifetime of the survey, so one would expect significant changes in the measured colors of objects of extreme color over the period, and this is being investigated. M. Doi et al. (2009, in preparation) will summarize the filter characteristics in full, including column-to-column variations within the camera and the changes with time.

Spectroscopic flat fields for the blue camera in the first spectrograph contain an interference pattern produced by the dichroic. The thickness of the dichroic coating is believed to be sensitive to the ambient humidity, and moisture which enters the system during plate changes affects the instrument response, shifting the interference pattern in wavelength in unpredictable ways on timescales comparable to the 900 s exposure time. The flats applied in processing were exposed several minutes prior to, or after, the science frames and therefore were not always representative of the true instrument response at the time of exposure. The interference pattern is most pronounced in the 3800–4100 Å region of the spectrum. If it shifts during an exposure, it will not be properly corrected by the flat field, causing significant distortion of blue absorption lines in stellar spectra, and systematically affecting estimates of metallicities and surface temperatures.

Flats obtained under different conditions were used to identify and model the stable and unstable (shifting) components of the flat, as shown in Figure 7. With this model in hand, we searched for shifts in the interference pattern over the typically 45 minute time a given plate was observed by comparing the results of the individual 15 minute exposures for each object. Thus, we took ratios of the extracted spectra from the separate exposures, and computed the median over all objects on a plate, giving results like those on the left-hand side of Figure 8. We fitted this ratio to the results expected from a shifting interference pattern (essentially a derivative of the shifting component in Figure 7), with the only free parameter being the amount of shift, and divided out this remaining component in each spectrum. The right-hand panel of Figure 8 shows that this technique removes the majority of the effects of the shifting interference. An example is shown in Figure 9, the spectrum of an A star observed on a plate where the interference term was particularly bad. The shapes of the absorption lines, especially H at 3970 Å, are much more regular in the new reductions.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The spectroscopic wavelength calibration is done quite accurately in SDSS, with typical errors of 2 km s⁻¹ or better. As the DR6 paper describes, however, detailed analyses of stellar spectra revealed occasional errors that were substantially larger than this, especially in the blue end of the spectrum. The algorithms for fitting arc and sky lines were made more robust for DR6, and this improved the situation considerably. We have implemented two further improvements for DR7:

• 1.  
  Spectroscopy is often done on nights with a moderate amount of moon. The bluest sky line used for wavelength calibration is an Hg line at 4046 Å, which is very close to a strong Fe i absorption line in the solar spectrum. Thus, when there is substantial moonlight in the sky spectrum, a fit to what is assumed to be an isolated emission line can be significantly biased, systematically skewing the wavelength solution at the blue end by as much as 20 km s⁻¹. In DR7, we now fitted this line to a linear combination of a Gaussian plus a stellar template including the absorption line, giving an unbiased estimate of the wavelength of the line. In practice, bright moon affected 10 plates (listed in Yanny et al. 2009) out of a total of 410 SEGUE plates.
• 2.  
  The sky and arc lines for each fiber are fitted to a wavelength solution; this is done independently for each fiber. This works well for the vast majority of plates. However, for a small fraction of plates, the arcs are weak (perhaps because the arc lamps themselves were faulty at that time, or because the petals which reflect the arc lamp light were not properly deployed), and the wavelength solution is poorly constrained. We therefore required that second- and higher-order terms in the wavelength solution be continuous functions of fiber number, to constrain the solution. We found that this produces much more robust wavelength solutions for those plates with weak arc observations, and has no substantial effect on the remaining plates.

The stellar spectral template library which gives the best radial velocity estimates is based on the ELODIE library (Prugniel & Soubiran 2001). We have removed one ELODIE template that gave velocities with a consistent offset from the rest of the library, as measured using the sample of ∼5000 stars with duplicate observations on each SEGUE plate pair. In order to provide more complete coverage in effective temperature, surface gravity, and metallicity for hot stars, we generated a grid of synthetic spectra using the models from Castelli & Kurucz (2003) over the same wavelength range and at the same resolving power as the spectra in the ELODIE library. This blue grid spans 6000–9500 K in 500 K increments, −0.5>[Fe/H]> − 2.5 in increments of 0.5 dex, and log g of 2 and 4. We also added a grid of synthetic carbon-enhanced spectra (B. Plez 2008, private communication, using the stellar atmospheric code described by Gustafsson et al. 2008) at values of [Fe/H] between −1 and −4, [C/Fe] between 1 and 4, log g values between 2 and 4, and T_eff in the range 4000–6000 K. With these improvements, the radial velocity scatter in repeat observations for objects that match the carbon star templates is now the same as for the full sample.

The DR6 paper describes a 7 km s⁻¹ systematic error in the radial velocities of stars (in the sense that the pipeline-reported velocities are too small). This is still with us in DR7; a correction is put into the outputs of the SEGUE Stellar Parameter Pipeline (SSPP; Lee et al. 2008a) but not elsewhere in the CAS or DAS. Beyond this problem, the plate-to-plate velocities of SEGUE stars have systematic errors of about 2 km s⁻¹ in the mean. The rms velocity error of any given SEGUE star observation is about 5.5 km s⁻¹ at g = 18.5, degrading to 12 km s⁻¹ at g = 19.5.

The spectroscopic pipeline combines observations of a given object on the red and blue spectrographs, and between the separate 15 minute exposures on the sky, by fitting a tightly constrained spline to the data, allowing discrepant points such as cosmic rays to be rejected. This spline requires as input the effective resolution of the spectra. As described in the DR6 paper, it did not do a perfect job; occasionally, very strong and sharp emission lines were erroneously rejected by this algorithm. This turned out to be due to the fact that the spline code did not adequately track the changing resolution of the spectra as a function of wavelength and fiber number. Including this effect significantly improved the behavior of this algorithm. Figure 10 shows an example spectrum of an object affected by this problem in DR6, and its improved counterpart in DR7, as is apparent by the correct 3:1 ratio of the 5007 Å and 4959 Å lines of [O iii].

Zoom In Zoom Out Reset image size

There is another problem, unfortunately not fixed in DR7, which has a similar effect. If the line is so bright that it is saturated in the individual 15 minute exposures of the spectrograph, it will also appear clipped. The flux value corresponding to saturation is a function of wavelength, but ranges from 2000 to 10,000 times 10⁻¹⁷ erg s⁻¹ cm⁻² Å⁻¹ (the units in which spectral flux density in reported in the SDSS outputs). Unfortunately, such saturated pixels are not flagged as such, although usually they are recognizable as having an inverse variance equal to zero. Luckily, objects with such strong emission lines are very rare, but the user should be aware of the possibility of objects with extremely strong emission lines and unphysical or unusual line ratios.

There have been several improvements made in the SSPP (Lee et al. 2008a; 2008b; Allende Prieto et al. 2008a) since the release of DR6. In particular, in DR6, the SSPP underestimated metallicities (by about 0.3 dex) for stars approaching solar metallicity. This was fixed in DR7 by adding synthetic spectra with super-solar metallicities to two of the synthetic grid matching techniques (NGS1 and NGS2), and by recalibrating the CaIIK2, ACF, CaIIT, and ANNRR methods. See Table 5 in Lee et al. (2008a) for the naming convention for each technique. Two new techniques (ANNRR and CaIIK3) were also added to the SSPP metallicity estimation schemes, and contributed to the high-metallicity performance improvement.

Two methods, ACF and CaIIT, have been recalibrated to the "native" g−r system, instead of making use of calibration on B−V, which required application of an uncertain transformation in color space. The ANNRR approach, which also tended to underestimate metallicity for near-solar metallicity stars, has been retrained on the SDSS/SEGUE spectra with improved stellar parameters, resulting in a better determination of the metallicity for metal-rich stars. Moreover, a neural network approach, based solely on noise-added synthetic spectra, has also been introduced. There remains a tendency for the SSPP to assign slightly higher metallicities for stars with [Fe/H] <−2.7. This offset is presently being calibrated out and will be corrected in SEGUE-2; see below. For more detailed descriptions of individual methods of the SSPP, we refer the interested reader to Lee et al. (2008a). Additionally, the pipeline now identifies cool main-sequence stars of low metallicity (late-K and M subdwarfs). The stars are assigned metallicity classes and spectral subtypes following the classification system of Lépine et al. (2007). Cool and ultracool subdwarfs are classified as subdwarfs (sdK, sdM), extreme subdwarfs (esdK, esdM), and ultrasubdwarfs (usdK, usdM) in order of decreasing metal content. The classification is based on the absolute and relative values of the TiO and CaH molecular band strengths, and derived from fits to K–M dwarf and K–M subdwarf spectral templates.

A number of open and globular clusters have been observed photometrically and spectroscopically with the SDSS instruments to evaluate the performance of the SSPP (Lee et al. 2008b). In addition, high-resolution spectra have been obtained for about 100 field stars included in the SDSS, and used to expand the SSPP checks over a wider parameter space (Allende Prieto et al. 2008a).