Measuring an Astrophoto

Why NOVA measures images

Measurements constrain decisions; they do not replace judgment. A number can be perfectly repeatable without being a universal quality score -- NOVA reads every metric only within the image state and process context where its assumptions actually hold, not as a standalone verdict.

Six kinds of evidence

Every metric NOVA presents is one of six kinds, and the distinction matters for how much weight it can carry: a direct descriptive statistic computed straight from pixel values (a median, a percentile, a robust sigma); a model-based estimate that assumes an extraction procedure (FWHM from an extracted source ellipse, a background grid estimate); a physically motivated proxy related to a real concept without being the conventional scientific measurement of it (NOVA's whole-frame 'SNR,' which is a structure-to-background-noise proxy, not a photometric object measurement); a heuristic diagnostic score built to detect a known artifact (a gradient-severity or ringing-score value); a preservation or similarity metric that measures difference from a reference rather than quality (SSIM); and a perceptual or aesthetic judgment that asks which valid result better serves the goal, which is legitimate evidence but is never a pixel measurement.

Background and noise

Sky level and robust noise estimates (median, MAD-derived sigma) describe the current image's own background region under the current measurement procedure. They are reproducible, but they remain dependent on which region was sampled and how -- the same physical sky can measure differently under a different mask or a different background-model scale.

Stars and PSF

FWHM in pixels is only directly comparable at the same image scale and using the same star-extraction procedure. A smaller measured FWHM after deconvolution does not mean the atmosphere suddenly improved -- it means the measurement changed, and the honest question is whether that change reflects real resolved detail or an artifact of the method.

Histogram and clipping

Percentile-based measurements of highlight and shadow clipping describe pixel distribution, not detector physics -- a histogram percentile is not the same thing as a hardware clipping rail, and treating the two as interchangeable overstates what the number actually shows.

Detail and structure

Sharpness and structure metrics such as spatial-frequency power, entropy, or SSIM are not universal quality scores. Noise and ringing can inflate a sharpness or entropy measurement without adding any real detail, which is exactly why these metrics are read as safeguards inside a process-specific guardrail rather than as a general 'more detail is better' rule.

Process-specific guardrails

Different processing steps need different safety checks, because the same statistic can mean opposite things in different contexts. NOVA's current hard-rejection rules reflect this: a denoise candidate is rejected for broadening measured stellar profiles too far, a deconvolution candidate is rejected for broadening FWHM at all or crossing a ringing threshold, and a background extraction candidate is rejected for leaking measured nebulosity into the modeled background. A candidate that triggers none of these guardrails has no analytic rejection triggered -- language the Handbook uses deliberately, since a current guardrail's silence never means the candidate 'passed objective validation.'

Star removal, HDR, halo suppression, and narrowband guardrails

The same principle applies to four more steps this page's evidence set does not yet cover with a current NOVA analytic rejection rule. Star removal must guard compact nonstellar structures and real target detail, not just a reduced star count. HDR remapping and tone compression can only redistribute already-captured range -- they cannot recover source information that was clipped before the image existed. Halo suppression needs an explicit no-op control alongside ring, color, and noise checks before a difference is credited as improvement rather than mere alteration. Narrowband palette comparisons need the derived channels' own provenance -- which raw filters and recombination produced each mapped channel -- alongside any comparison, not instead of it.

Star removal and halo suppression are stated here as required guardrails without a current NOVA analytic rejection rule enforcing either one; both are named again below as open limitations, not as passed guardrails. HDR's clipped-data framing and narrowband's provenance requirement are evaluation principles this Handbook enforces in how it describes those steps, independent of whether an analytic rule for either is added later.

When comparisons are invalid

A measured comparison is only as strong as what both sides held constant. FWHM and eccentricity comparisons break when image scale or extraction thresholds differ; noise and SNR-like comparisons break when crop composition or background modeling changes what feeds the measurement; color-ratio comparisons break when filter or data type differs, when target emission contaminates the sampled sky, or when a linked versus unlinked stretch has already altered the channel relationships being measured.

Objective evidence vs aesthetic preference

Some questions are close to purely factual: did measured background sigma decrease under the same procedure, did an operation fail to produce output, did geometry change. Others mix scientific and processing judgment: did background extraction remove a gradient without removing faint real signal. And some are primarily aesthetic: which valid stretch has the most pleasing tonal balance. The Handbook states plainly which kind of question a given claim is answering, rather than dressing a taste judgment up as a physics result.

How Experiment Mode uses measurements

Measurements enter a candidate comparison in order: confirm the candidate executed at all, apply hard analytic safety gates, bring in the remaining metrics as context, then apply visual or human judgment, with repeated evidence accumulating on top of all of it. This is the same execution-then-safeguards-then-judgment sequence Experiment Mode itself follows -- measurement is the layer beneath comparison, not a replacement for it.

Provenance and uncertainty

A measured value and a fallback estimate are not interchangeable, even when they end up in the same field. Where a threshold came from, whether a metric's implementation has since changed version, and whether a number was actually measured or substituted after a failure all belong to the claim's provenance -- omitting them lets an estimate quietly read as a measurement.

Auditable claims

1. A denoise candidate is not accepted for reducing noise alone -- NOVA's current rejection rule fails a candidate whose measured FWHM broadens materially against the input, because broadening is a proxy for real detail being destroyed along with the noise.

   source-confirmed · none

2. A deconvolution candidate that appears sharper is not accepted if it broadens the measured stellar profile at all, or if its ringing score crosses NOVA's current threshold -- an apparently crisper result can still be a ringing artifact rather than resolved detail.

   source-confirmed · none

3. A flatter measured background is not automatically a better result -- NOVA's current rejection rule watches a nebulosity-leakage score specifically because an algorithm can flatten the sky by subtracting real extended emission along with the gradient.

   source-confirmed · none

4. A candidate that triggers no current hard-rejection rule is described as having no analytic rejection triggered, never as having 'passed objective validation' -- it may still carry missing metrics, an assessor failure, or a real defect that no current guardrail happens to catch.

   source-confirmed · none

5. A structural-similarity score describes how much an output resembles a reference; it does not describe whether the output is good, since processing is often deliberately meant to change the image. Similarity is preservation evidence, not a quality score.

   source-confirmed · none

6. A measured comparison between two outputs is only as valid as what was held constant between them -- color-ratio comparisons, for example, break when the filter or data type differs, when target emission contaminates the sampled sky region, or when a linked versus unlinked stretch has altered the channel relationships being compared.

   source-confirmed · none

7. A star-removal or star-reduction candidate is not judged by how many stellar sources it removed alone -- the framework requires guarding compact nonstellar structures and genuine target detail against being stripped out along with the stars, the same signal-preservation-over-reduction principle denoise uses. NOVA's current analytic rejection rules do not yet include a star-removal-specific guardrail for this; it is stated below as an open limitation, not a passed guardrail.

   source-confirmed · none

8. HDR remapping and tone compression can only redistribute the range that survived capture -- they cannot recover source information that was already clipped at acquisition, so a well-tone-mapped highlight or shadow is still built from whatever data existed there beforehand, never data that was never captured.

   source-confirmed · none

9. Judging a halo-suppression result credibly requires comparing it against an explicit no-op control alongside ring, color-fringing, and noise checks, not a single before/after glance -- a suppressed halo that only looks different is not evidence the tool improved the image rather than merely altering it. NOVA's current analytic rejection rules do not yet implement this guardrail; it is stated below as an open limitation, not a passed guardrail.

   source-confirmed · none

10. A narrowband palette comparison (SHO, HOO, or similar) is only meaningful alongside the provenance of its derived channels -- which raw filters fed each mapped channel, and what recombination or mixing ratio produced it -- because comparing two palette renders without that provenance can mistake a mapping choice for a genuine signal difference.

    source-confirmed · none