SPCC and PCC both fail with "Insufficient data" on vdB 123 — regression collapses to 0-1 samples despite hundreds of matched stars - Forums

Equipment: PlaneWave CDK12.5, ASI6200MM Pro, LRGB Target: vdB 123 (reflection nebula, Serpens Dark Cloud) — a field known for heavy interstellar reddening; the illuminating star and surrounding field carry a strong orange cast even in professionally-processed references of this same object.

The problem

Both SPCC and PCC fail at the identical step — the robust linear regression between channels — despite:

A confirmed, pixel-accurate astrometric solution (see below)
774 Gaia DR3/SP catalog sources found within the search radius
395 catalog sources matched after magnitude filtering
Up to 790 image stars successfully fit with valid PSFs

The regression still collapses to 0-1 usable sample pairs, well below the minimum of 5 required:

* Integrating spectrum data Robust linear regression (R/G): 0% *** Error: Insufficient data: only 0 sample(s) are available; at least 5 are required.

What I've already ruled out

I wanted to post this only after eliminating the obvious causes, rather than ask for troubleshooting I'd already done:

WCS accuracy — the field was blind-solved externally (ASTAP) after PI's own ImageSolver failed to converge (RANSAC couldn't find a valid star-pair match after many retries, including mirrored-coordinate and apparent-coordinate fallbacks). Confirmed correct three independent ways:
- Annotate with a Tycho-2 star-marker overlay lands precisely on real stars across the entire frame, including well away from center
- astrometricSolutionSummary() and hasAstrometricSolution both report a valid, complete solution once reloaded
- The full native PCL:AstrometricSolution:* property block (transformation matrix, reference coordinates, projection system) is present and well-formed in the file's XISF properties
Min SNR threshold — tried at both 20 and 0 (the two values suggested for sparse fields). Star counts surviving PSF fitting actually increased at lower thresholds, but the regression still failed identically.
Saturation threshold — lowered from the default 0.75 down to 0.50/0.45. No change in outcome.
BlurXTerminator (stellar-only) pre-pass — applied manually before SPCC, since tightening star PSFs is a technique that's reliably fixed similar SPCC failures for me on other targets in the past. This time it nearly quadrupled the number of PSF-fit stars (190 → 790) but the regression still returned only 1 sample.
PhotometricColorCalibration (PCC) as an alternative to SPCC — fails at the same point with the same signature, despite using a completely different calibration mechanism (broadband catalog magnitudes, no spectral integration step). Both tools failing identically suggests the issue is upstream of either algorithm's specific matching logic.
Inter-channel pixel clipping — checked min($T) per channel; confirmed 0.00000 present, though I haven't yet fully isolated whether this affects the star cores themselves or only background/sky pixels (PSF fitting succeeding on hundreds of stars suggests the star peaks themselves aren't flat/clipped).

My working theory

Given that the target sits in a region with well-documented, severe interstellar reddening, and that hundreds of real stars are being detected and matched to real catalog entries but still rejected at the regression stage — I suspect the robust regression's outlier rejection is treating the field's uniformly reddened stars as bad data, since their measured colors deviate consistently and dramatically from what Gaia's catalog says a "normal" star of that spectral type should look like. If the reddening is severe and uniform across the whole field rather than random per-star, that could look statistically identical to "every star is an outlier" to a robust regression, causing it to reject everything rather than fit a genuine (if steep) reddening slope.

Question for the forum

Has anyone dealt with SPCC/PCC failing this way on a heavily reddened field, and is there a real parameter (rejection/RANSAC tolerance, a different white-reference option, or something else) that addresses this specifically — as opposed to the general low-star-count troubleshooting (which I've already tried and which didn't help here)? Happy to share the actual data/logs if useful.

Thanks for your reply, and for the offer to take a look at the data. What you are suggesting is a very interesting angle — I hadn't thought to check whether the same star-matching mechanism was failing at both steps (ImageSolver's blind search and SPCC's regression), which does point more toward something in my own pre-processing than the field itself.

I'm going to check channel alignment on my end first (R/G/B/L masters, pixel-by-pixel on a few stars) since I've hit this kind of cross-channel misalignment before on multi-night data, but I would appreciate an independent pass on your end regardless. If you want to give it a try, I'm happy to share the four calibrated master lights (R, G, B, L), which should be enough to reproduce the issue. Let me know if that works and I'll get them uploaded for you.