Observing Operations | Reviews | Survey Management

Sloan Digital Sky Survey

Review of Data Processing Operations

Overview and Status of the Spectroscopic Pipeline as of July 2000

Josh Frieman

July 7, 2000


Outline

I. Overview of the Spectroscopic Pipeline

  1. Purposes of the pipeline
  2. Structure of the pipeline

II. Status of the Pipeline

  1. A Brief History
  2. Pipeline Requirements and Current Performance wrt Requirements

III. Future of the Pipeline

  1. Tasks remaining to satisfy requirements and estimated time to completion
  2. Enhanced goals and plans to achieve them
  3. Staffing needed for long-term code maintenance and operations


Personnel: As far as I am aware, the following people have contributed to the development of the Spectroscopic pipeline (I would like to be informed of others): M. Bernardi, S. Burles, F. Castander, A. Connolly, D. Finkbeiner, L. Hui, D. Johnston, J. Loveday, R. Lupton, A. Meiksin, A. Merrelli, J. Munn, R. Nichol, A. Pope, D. Schlegel, M. Strauss, M. Subbarao, B. Wilhite, B. Yanny.


I. Overview of the Spectroscopic Pipeline

1. Purposes of the pipeline

The spectroscopic pipeline is designed to be a multi-purpose, highly automated pipeline for processing the 10^6 galaxy, QSO, and stellar spectra from the SDSS spectrographs. This is an unprecedented challenge, mitigated by the very high quality spectra these spectrographs produce.

The pipeline is designed to extract and process all spectra taken in the course of the Survey, and specifically to:

  • archive the reduced, red/blue combined, co-added 1d spectra
  • classify objects (Star, Galaxy, QSO) independently of the target selection pipeline
  • estimate redshifts and provide spectral information for all objects possible, with required redshift accuracy and success rates depending on object type (see Sect. II)

In addition to the above goals, which enable science to be carried out with the spectra, the pipeline has the following roles in survey operations:

  • provides near real-time diagnostic S/N outputs for APO observers so that exposure time for each plate can be determined under varying conditions; this is a critical element in determining Survey time to completion
  • provides diagnostic outputs for spectroscopic data processors at Fermilab so that data quality on each processed plate can be rapidly assessed as satisfying Survey scientific requirements or not (i.e., a check on the previous point); these outputs also provide QA monitoring of the spectrographic system over time
  • provides feedback to target selection on object classification and on redshift success rates for different object classes as function of magnitude, surface brightness, etc; this allows the Working groups to optimize TS parameters for survey efficiency and completeness

Aside from these functions in survey operations, the Spectroscopic pipeline is not a critical-path item for the Survey to operate.

2. Structure and Functionality of the Pipeline

N.B.: The available documentation on the design of the spectroscopic pipeline is unfortunately out of date (the last major update to the printed documentation was at the time of the Spectroscopic pipeline review in April '99). This will be rectified in the next few months, when technical papers describing the 1d and 2d pipelines will be prepared for publication.

The Spectroscopic pipeline is split operationally into two parts, 2d and 1d. The 2d pipeline reduces the raw data and calibration images from the red and blue CCD cameras from each spectrograph and outputs merged, combined, flux-calibrated spectra and noise for analysis by the 1d pipeline. The 1d pipeline determines emission and absorption redshifts, classifies spectra by object type, and outputs spectral information about each object. During normal data processing operations at Fermilab, a batch of spectroscopic plates is first run through the 2d pipeline and then through 1d.

Spectro2d

In October of 1999, it was determined that the then-current version of 2d, which was built around IRAF commands, was not flexible or robust enough to satisfy the Survey requirements. The Spectro2d pipeline was completely rewritten from scratch by S. Burles and D. Schlegel in IDL (idlspec2d). The current version of the code used for data processing (3c) was tagged in early May of this year and was then determined to meet the Survey Requirements of this year and was determined to meet the Survey Requirements in early May.

Key Tasks of Spectro2d

  • Overscan
  • Bias and dark subtraction
  • Spatial tracing of fiber images from flat-field image
  • Optimal extraction of flats and arcs
  • Wavelength calibration from arc lamps and sky lines
  • Flat-fielding: remove spectral response using flat exposures; remove pixel to pixel response using uniformly illuminated flats
  • Optimal Object and sky extraction, including scattered light removal
  • Sky subtraction using supersky built from sky fibers
  • Telluric absorption removal using spectrophotometric standard spectra
  • Flux calibration from spectrophotometric standard (F) stars spectra
  • Output spectrum, variance, and pixel mask information
  • Rebin spectra and merge red and blue halves; heliocentric correction
  • Combine multiple 15-minute exposures as needed to reach required S/N per plate (multiple-night exposures are combined if necessary, but not if a plate has been replugged)
  • Output diagnostic information on cumulative S/N per plate: this is used by the spectroscopic data processors to decide if a given plate satisfies Survey requirements and can be declared "good, done" (and unplugged). The current requirement is that the mean (S/N)^2 100 for objects with fiber magnitudes of g'=19.2, r'=19.25, i'=18.9 in each spectrograph. This limit will be reviewed and refined in the next month (see below)
  • Output other diagnostic information that provides QA monitoring on the spectroscopic system (e.g., errors in wavelength solution, QA for scattered light levels, throughput compared to PHOTO g,r,i magnitudes, etc)

For a list of the routines used in idlspec2d, see IDL Help for IDLSPEC2d.

In addition, a faster, stripped-down version of the 2d pipeline is run on the mountain during spectroscopic observations. This mountain-top version of the 2d pipeline

  • provides "real-time" information to the APO observers on the approximate S/N achieved per 15-minute exposure

Based on this information, the APO observers carry out as many spectroscopic exposures on a plate as are necessary under the given conditions to achieve the cumulative (S/N)^2 deemed necessary to meet the Survey requirements. Currently, this is set at (S/N)^2 80 at the same magnitudes as above; this translates approximately to (S/N)^2 100 in idlspec2d.

Spectro1d

The 1d pipeline, which analyzes the combined spectra output by spectro2d, is written in C and TCL. A velocity dispersion module written in IDL has been recently added. The version of the code currently used in data processing, 1d_4 was/will be cut on July 10 of this year. The code outputs a FITS image for each fiber: it includes the 1d spectrum, noise, and mask array from 2d, basic information about the target from PHOTO and TS, as well as line measurements and redshift determinations. These outputs are listed in the (unofficial) Spectro Data Model; the schema of outputs of the pipeline as they will appear to the SX database can be found at this Web site. Although out of date, general information about the structure of the 1d pipeline can be found at the Spectro Pipeline Home Page.

Key Tasks of Spectro1d

  • Fit and subtract continuum
  • Find and fit emission lines (using a wavelet filter)
  • Identify emission lines and determine an emission-line redshift and error where possible
  • Cross-correlate the spectra with stellar, galaxy, and QSO templates to determine a cross-correlation redshift and error (emission lines are removed in this procedure, except for the QSO template)
  • Classify all object spectra as Galaxy, QSO, high-redshift QSO, Star, or unknown (currently a discrete choice)
  • Output a final redshift by comparing emission and absorption redshift confidence levels; final redshift determination is denoted high confidence, low confidence, inconsistent, or failed
  • Output information on identified emission and absorption lines (equivalent widths, etc)
  • Output warning flags for spectra (e.g., blue or red half of spectrum missing; conflicting high-CL cross-correlation redshifts; outlier redshift for object class; classification differs from TS)
  • Flag low-confidence level redshift determinations for subsequent inspection and interactive z-determination
  • Provide interactive redshift determination environment for inconsistent, failed, and other spectra: to be used by Spectro pipeline developers to manually determine redshifts. It is expected that this will be necessary for a few percent of all spectra; also to be used on a spotcheck basis for QA monitoring.
  • Estimate velocity dispersion and error for galaxies
  • Provide diagnostic output for Spectroscopic data processors to monitor QA (e.g., compare emission and absorption redshifts for all objects where both are available and check dispersion and outliers; plot redshift vs. magnitude and redshift histogram for each plate, etc)


II. Status of the Spectroscopic Pipeline

1. Recent History of Pipeline Development

  • Prior to 3/99: Pipeline development. Pipeline testing by M. Strauss using simulations provided by D. Schneider.
  • 3/99: first 3 spectroscopic plates taken with SDSS 2.5m telescope
  • 4/99: Review of Spectroscopic pipeline at Chicago.
  • 9-10/99: First calibrated spectra (though shutter problems and light leaks).
  • 10/99: Spectroscopic pipeline meeting at Fermilab. Decision taken shortly thereafter to rewrite spectro2d.
  • 12/99: Spectroscopic pipeline meeting at Princeton. Task lists and schedule to pipeline completion updated. As of May 2000, the 2d and 1d pipelines were within 1-2 months of this schedule.
  • 3/00-6/00: Spectra of improved quality and in larger numbers obtained. Processing of the data through the pipeline by spectroscopic data processors at Fermilab gradually becoming routine.
  • 5/00: Spectroscopic pipeline meeting at Fermilab, shortly after cuts of 2d and 1d. 2d determined to satisfy Survey requirements.
  • 7/00: New version of 1d cut and used to reprocess plates at FNAL. Data Processing Review at FNAL.

2. Spectroscopic Requirements and Current Performance

The requirements on the performance of the SDSS Spectroscopic pipeline are described in two documents: Section 7 of the Scientific Requirements and Scientific Commissioning for the SDSS and the Offline software processing requirements. Note also that some of the requirements on the spectroscopic system are implicitly linked to requirements on Target Selection enumerated in Section 6 of the Scientific Requirements and Scientific Commissioning for the SDSS.

The current status of the Spectroscopic pipeline in terms of the Survey requirements is described in varying levels of detail in three documents:

  1. Michael Strauss' document status of requirements for the survey provides a general overview.
  2. Andy Connolly's document Requirements and Status of the SDSS Spectroscopic Systems describes the status of the SDSS spectroscopic hardware and software with regard to the requirements set out in Sect. 7 of the Scientific Requirements and Scientific Commissioning for the SDSS
  3. Mark Subbarao's webpage Status of 1d with respect to SDSS Science requirements describes the current performance of the 1d pipeline with respect to the requirements set out in Sec. 7 of the Scientific Requirements and Scientific Commissioning for the SDSS as well as those set out in the Offline software processing requirements. Parts of the information on this page have been incorporated into Requirements and Status of the SDSS Spectroscopic Systems. The numbers reported on this page are based on comparison of spectro1d outputs (using the checked-out version of the code as of 6/23/00) with manual determination of redshifts on several plates of varying S/N taken in the Spring of this year. Over the next month, all reduced spectroscopic plates taken so far (over 40) will be manually checked; this will allow more accurate statistics on pipeline and spectroscopic system performance as a function of plate S/N to be compiled.

Summary on Pipeline Status and Requirements

Based on tests carried out and summarized in the documents above,

  • idlspectro2d is deemed to satisfy the Survey science requirements
  • spectro1d satisfies most of the Survey science requirements at the current minimum acceptable S/N level per plate. This includes the requirements on redshift success rates for galaxies, BRGs, and QSOs, the redshift accuracy requirements on galaxies and QSOs (though these have only been tested internally; external comparison is difficult and remains to be carried out), and the successful classification of galaxy and QSO spectra as such.

The spectro1d requirements which are not yet satisfied or verified are the following:

  • Although 95% of QSO spectra are successfully classified as such, the requirements also state that 99% of the QSO spectra not correctly classified should be classified as unknown. Currently the majority of misclassified QSO spectra (typically 0 or 1 per plate) are instead identified as stars or galaxies by the pipeline. It is the goal of the pipeline that misclassified QSO spectra will be flagged for manual classification and redshift determination.
  • Currently 96% of stellar spectra are correctly classified by the pipeline (based on 5 plates), while the requirement is 99%. This will be addressed by the addition of hot and cool stellar templates for the cross-correlation analysis.
  • The status of the requirement on the accuracy of BAL QSO redshifts is not yet known; internal reproducibility is within the required accuracy limit, but comparison with manual redshift determinations needs to be done. This will occur in the next few weeks.

Note: internal and external verification of redshift accuracy for galaxies at the 30 km/sec level is a difficult problem. The pipeline developers will compare 1d redshifts with high-resolution redshifts from the literature where available. Aside from this, manual determination of redshifts of all the spectra taken so far (~42 plates) will be used as a "truth table"; statistical comparison with these manual redshifts will be used to determine whether the pipeline meets the science requirements of the survey.


III. Future Development of the Spectroscopic Pipeline

1. Remaining Tasks and Estimated Time to Completion

The near-term prioritized tasks for the 1d pipeline listed here are aimed at satisfying the remaining Survey science requirements. It is expected that these will be completed by 9/1/00:

  1. Add stellar templates to improve stellar classification success rate
  2. Improve QSO classification scheme so that unclassified QSO spectra are flagged for visual inspection
  3. Manually determine redshifts and classification for all objects on all plates observed so far; this will be used to check classification and redshift success rates as functions of plate S/N and thus to determine required plate S/N for survey operations. It will also be used to determine whether additional parameters beyond plate S/N need to be checked in declaring plates "done".
  4. Recalibrate emission and absorption redshift Confidence Levels using manual redshifts above.
  5. Implement redshift warning flags (e.g., missing half-spectra, discordant high-CL redshift determinations, etc)
  6. Check external accuracy of redshifts where possible from the literature
  7. Check BAL QSO redshifts against manual determination from the QSO working group to verify required accuracy
  8. Improve continuum fitting algorithm (expected to improve QSO redshift success and classification)

In addition to these tasks, both the 2d and 1d pipelines need to implement regression tests to ensure code robustness in the course of future development.

2. Enhanced Goals and Estimated Time to Achieve Them

Spectro2d:

It is expected that the following enhanced goals will be achieved by approximately 10/1/00:

  1. output estimated error covariance in addition to variance
  2. implement code to reduce binned diamond-pattern exposures
  3. improve flux calibration based on spectrophotometric standards?

Spectro1d:

The following enhanced goals of Spectro1d are expected to be achieved at various stages, but in all cases by the end of 2000; they are listed in expected order of completion:

  1. Test and finalize galaxy velocity dispersion code
  2. Output line indices
  3. Implement galaxy and QSO spectral classification using eigenspectrum (PCA) analysis
  4. Implement stellar spectral classification using non-continuum-subtracted spectra/templates
  5. Determine feasibility of galaxy spectrophotometry using diamond-pattern exposures; implement to the extent possible

3. Long-term Staffing Needs

The long-term staffing needs for the spectroscopic pipeline are based on the two roles that will be needed during Survey Operations:

  1. Long-term code maintenance: QA monitoring; possible code development necessary if hardware changes significantly (i.e., some component breaks or degrades over time); possible additional enhancements desired by the collaboration; provide assistance as needed to Spectroscopic data processors at Fermilab
  2. Interactive Redshift Determination: this will be carried out by the pipeline developers over the course of the Survey for up to a few percent of all spectra. This will provide z determinations and object classifications for objects where the automated pipeline cannot make a determination as well as provide spotchecking of the automated pipeline for QA.

The anticipated long-term staffing level (2001 and beyond) needed to meet these requirements is 1 FTE supported by ARC.



 
Submit comments and questions to the SDSS Helpdesk. Last updated 04/01/14