Astronomy · Observatories, surveys and archives

Archives, metadata and data provenance

Trace measurements through their inputs and processing, then examine how documented citizen classifications contribute evidence without becoming automatic discoveries or independent confirmations.

Continue in the approved chapter order. Necessary vocabulary, quantities and assumptions are explained in the lesson.

Back to the chapter

After reading, you can

Distinguish archive access, metadata, provenance and verification.
Recognise shared observations beneath multiple data products.
Interpret a conditional classification fraction and separate contribution, candidate and verified discovery.

A reusable file needs a documented meaning

An astronomical archive provides access to retained observations and data products. Downloading a file is the beginning of reuse: the scientific question still requires an understanding of what its values mean, how they were obtained and whether they are suitable for the proposed comparison.

Metadata describe the data and their context. Useful information includes the target or field, observing time and time system, instrument and band, coordinate frame, units, exposure information, uncertainty quantities, masks, quality flags and release identity. The exact requirements depend on the intended analysis.

Provenance describes relationships between inputs, producing activities, outputs and responsible agents. Here an agent means a person, organisation or system with a recorded role. The IVOA Provenance Data Model formalises these relationships so that processing history can be represented and shared.

A filename or timestamp alone cannot supply that history. Nor does an archive’s name guarantee that every relevant processing setting is recorded. The amount of available detail must be checked against what the proposed inference needs.

A reuse checklist for the particular product and scientific comparison; these are categories, not a universal archive schema.
Information	Question it helps answer
Observation IDs, field, band and time conventions	Which original measurements and epochs contribute?
Quantity definitions, units and uncertainty fields	What is measured or derived, and how uncertain is it?
Calibration and software versions, settings	Which relations and transformations produced the output?
Masks, flags, exclusions and selection	Which measurements or objects could be missing or unreliable?
Product/release identity and responsible groups	Which documented version is being used?
Usage terms, acknowledgements and citations	How should the material be reused and attributed?

Two products can contain the same original observations

Consider an original hypothetical archive with two target exposures, E1 and E2, each lasting 300 seconds. Processing activity P1 uses E1, E2 and calibration data C1 to produce product A. Another activity P2 uses those same exposures with calibration data C2 to produce product B.

The unique original target exposure time is 300 + 300 = 600 seconds. The archive contains two derived products, but not 1,200 seconds of new target observations. C1 and C2 are calibration inputs in this model, not additional target exposures.

The figure shows educational input-to-process-to-output arrows. The two exposure nodes are shared between branches so that their dependence is visible. It is a teaching data-flow diagram, not a complete implementation or claim about every relation in the IVOA standard.

Agreement between A and B may be useful for studying the effects of recalibration. It cannot automatically be treated as independent confirmation from new observations. Shared noise, response assumptions and other input limitations can appear in both products.

Original hypothetical provenance model. E1 and E2 identify the same observations in both rows.
Activity	Inputs	Output	Unique target exposures
P1	E1, E2, C1	A	E1 and E2
P2	E1, E2, C2	B	The same E1 and E2

Two products can share original exposures

Scroll sideways to see the full diagram. Its explanation is below.

Original hypothetical provenance diagram. Arrows show educational input → process → output direction, not a specified PROV serialization. There are exactly two original target exposure nodes: E1 with duration 300 seconds and E2 with duration 300 seconds. Both same exposure nodes feed both processing runs P1 and P2. Distinct calibration input C1 feeds only P1, and C2 feeds only P2; calibration data do not add target exposure time. P1 produces product A and P2 produces product B. The exact directed edges are E1 → P1, E1 → P2, E2 → P1, E2 → P2, C1 → P1, C2 → P2, P1 → A and P2 → B. E1 paths are solid and E2 paths dashed, in addition to colour; calibration paths are dotted. The two crossing diagonal exposure paths have no junction or extra exposure node. The original target exposure identities are shared, so their unique total duration is 300 + 300 = 600 seconds, not 1200 seconds. Two derived products do not mean two new sets of observations. Products can differ because calibration and processing differ while retaining dependence through the same original measurements. Agreement is not automatically independent confirmation. Real provenance records need input identities, observation times and time systems, band and units, calibration and software versions, processing settings and quality, responsible group and release or citation. Those extra values are not invented here. This conceptual graph is not a complete IVOA implementation, a claim of complete archive history, metrological traceability or proof that a result is true.

A traceable history supports checking but does not certify truth

The Hubble Spectroscopic Legacy Archive documentation gives a concrete example: science quantities and errors have defined units, while provenance information identifies contributing spectra, instruments, exposure details and calibration versions. Its coadded products also have documented quality and exclusion rules.

Calibration itself relates instrument indications to reference values under stated conditions with uncertainties. Knowing a calibration filename is helpful provenance, but it does not establish all those metrological conditions. A file lineage alone is not a demonstration of a measurement result’s metrological traceability.

Reproducibility asks whether a sufficiently specified analysis can be repeated to obtain its stated result under the relevant conditions. A reproducible pipeline can still use an unrepresentative sample, an inappropriate model or a mistaken calibration. Provenance helps locate and examine such possibilities rather than removing the need to test them.

Related releases can also reuse observations. The Gaia Data Release 3 collaboration paper illustrates why product definitions, input choices and limitations must be read at the release level. Counting published catalogues is not the same as counting independent observations.

Documented public contributions can become research inputs

Citizen-science projects invite members of the public to contribute to research tasks. NASA’s Galaxy Zoo introduction describes classifying galaxy images through questions. These classifications can provide useful evidence when the project documents how they are recorded, combined, checked and related to the scientific question.

The historical Galaxy Zoo 2 release paper describes image, participant and time associations, handling of repeated user-image submissions, user weighting and comparisons with overlapping expert and automated catalogues. Those documented methods describe that release; they should not be assumed to specify every later project.

An image’s apparent size, quality and viewing conditions can affect its classification. Decision-tree tasks can also have conditional denominators: a classification may reach one question only after particular earlier answers. A classification that never reaches a task is not automatically a “no” answer to it.

Raw vote fractions, weighted fractions and fractions adjusted for a documented classification bias are different quantities. Their definitions and assumptions need to be read before interpretation. Agreement between methods sharing images or assumptions is valuable evidence with dependencies, rather than automatic independent ground truth.

State the fraction and the level of verification

For an original hypothetical decision-tree example, 50 classifications concern one image. Only 20 reach a particular task, and 16 of those answer yes. The raw conditional task fraction is 16/20 = 0.8. The fraction of all 50 classifications following that yes branch is 16/50 = 0.32.

These fractions answer different questions. Neither is the percentage of galaxies in the Universe with that feature, a calibrated 80 percent probability of physical truth, or a verified discovery. Actual release columns may use weighting or bias adjustments absent from this simple unweighted example.

A participant’s documented contribution, a candidate identification, subsequent checking and a verified discovery are distinct states. A surprising classification can motivate investigation; scientific claims require appropriate validation, comparison and sometimes new observations. Credit and attribution follow the project and data-release records rather than a promise attached to every click.

Archives and public participation widen the community able to examine evidence. Their value grows when quantities, history, selection and verification remain inspectable. The next original chapter changes physical focus to gravity and orbits, while retaining this discipline of connecting measurements to the models they can support.

Invented unweighted classifications of one image; this is not a Galaxy Zoo dataset or a truth-probability calibration.
Count or fraction	Value	Denominator or scope
All classifications	50	One image
Classifications reaching task	20	Conditional path in the question tree
Yes answers at task	16	Those reaching this task
Conditional yes fraction	16/20 = 0.8	Task responses
Whole-path fraction	16/50 = 0.32	All classifications of this image

Worked examples

Follow the shared inputs before adding exposure

Two archive products each use the same E1 and E2 exposures of 300 seconds each, with different calibration inputs. How much unique original target exposure is represented?

Identify the original exposure IDs rather than counting output filenames.
E1 contributes 300 seconds and E2 contributes 300 seconds once each to the unique set.
Both processing branches use that same set, so two outputs do not double it.

600 seconds of unique original target exposure. The products can differ, but they do not constitute 1,200 seconds of new observations or automatic independent confirmation.

Keep the task denominator visible

In the synthetic classification example, why are both 0.8 and 0.32 legitimate fractions?

Sixteen yes answers among the twenty task responses give 16/20 = 0.8.
Sixteen classifications following that branch among all fifty give 16/50 = 0.32.
The other thirty classifications did not reach the task, so treating all thirty as task-level no votes changes the question.

The fractions describe different denominators. Neither directly measures a universal physical population fraction or verifies a discovery.

Try the reasoning

Does an observation date in a filename supply complete data provenance?
Show the answer and reasoning
No.
Inputs, quantities, processing, versions, conditions, exclusions and responsible roles may also be needed.
Are two reductions of the same exposures independent new observations?
Show the answer and reasoning
No.
They can reveal processing differences while retaining shared measurement noise and limitations.
Does 16 out of 20 yes votes establish an 80 percent calibrated probability that a physical interpretation is true?
Show the answer and reasoning
No.
That is a raw conditional task fraction in the invented example. Validation, image limitations and the inference model remain relevant.
Can a reproducible analysis still have a selection bias or wrong physical model?
Show the answer and reasoning
Yes.
Reproducing a calculation checks a different property from the suitability of its inputs and assumptions.

What to carry forward

Metadata give quantities and context; provenance connects inputs, activities, outputs and responsibilities.
Shared input observations remain shared evidence across different products and releases.
Citizen classifications contribute documented research inputs whose denominators, validation and attribution must be stated.

References

Teaching resources and research supporting this lesson’s explanations and examples.