Evaluation

All evaluations run on the held-out split produced by specsr.data.splits.get_training_split(), which is shared across every training stage and analysis script.

The split is 80/20 over parent galaxies, and the evaluation partition holds original spectra only — 572 real galaxies. The built product augments every galaxy at 21 rows each, so an unfiltered partition would be ~95% synthetic rows and would compute statistics over 21x correlated samples, understating uncertainties. See ARCHITECTURE.md for why there is no separate sealed test partition in this configuration, and how to restore one.

specsr evaluate line-flux --outdir figures/
specsr evaluate line-snr  --outdir figures/
specsr evaluate residuals --outdir figures/
specsr evaluate redshift  --outdir figures/
specsr evaluate sample    --outdir figures/

Analyses

line-flux — Fits Gaussian profiles to the super-resolved output, to the low-resolution input it was given, and to the high-resolution reference, for [O II] λ3727, Hβ, [O III] λ5007 and Hα, and compares the integrated fluxes (sqrt(2π) · A · σ, in erg s⁻¹ cm⁻²) one-to-one against the reference. This is the test of whether the model recovers line fluxes, which is what downstream diagnostics (star-formation rates, ionisation parameters, metallicities) actually consume.

Each spectrum is measured through its own line profile, because the three differ by a factor of ten in resolution: a fixed velocity aperture narrow enough to isolate a line in the reference holds about a fifth of the same line at prism resolution, and would report unresolved flux as missing flux. [O II] and Hα use a single Gaussian (specsr.linefit.measure_line_fluxes). Hβ, [O III] λ4959 and [O III] λ5007 lie within 9000 km/s of one another, and the prism (median line FWHM 4800 km/s there) leaves λ4959 blended with λ5007 in 84% of the held-out galaxies and Hβ blended with the doublet in 45%, mostly below z ≈ 3. The three are therefore fit jointly as Gaussians of a shared width (specsr.linefit.fit_hbeta_oiii), so the flux of λ4959 is not assigned to a neighbour; a single Gaussian reads [O III] λ5007 about 30% high and can widen over the whole complex when centred on Hβ. A line enters the comparison when the reference fit detects it in emission at S/N ≥ 5.

On the held-out set the low-resolution input already recovers the total flux of strong lines (median ratio to the reference 0.91–1.09 for Hβ, [O III] and Hα, with 2–11% of lines off by more than a factor of two). The super-resolved fluxes are lower and more scattered (median 0.72–0.87, 24–30% off by more than a factor of two). Use the low-resolution spectrum, not the reconstruction, when what you need is the total flux of a line.

line-snr — Compares emission-line signal-to-noise between the low-resolution input and the super-resolved output, and reports the fraction of spectra improved per line, stratified by input S/N.

residuals — Two-dimensional residual maps over the evaluation set sorted by redshift, for LR−HR, LR−SR and SR−HR.

redshift — Redshift recovery from LR, SR and HR inputs. Because a learned estimator could in principle favour its own training distribution, this analysis also supports method-independent estimators so the comparison does not rest on a single approach.

sample — Characterises the spectroscopic sample against the parent photometric catalogue (redshift–magnitude, stellar mass versus star-formation rate), making the selection function and its biases explicit.

Note

Line fluxes and signal-to-noise are distinct claims. A higher S/N measured without reference to ground truth does not by itself establish that a line is correctly modelled; the line-flux analysis is what does that.