I added an update about suitability of long-tube fluorescent bulbs (non CFLs) for calibration which should be interesting to @liz and maybe @cfastie as well.

That's some excellent detective work. We know there is mercury in all fluorescent lamps, so all the big mercury peaks should be present in all good spectra of fluorescent lamps. If the 542-543 nm peak is missing in your spectrum of a long fluorescent tube, then that peak is probably not a mercury emission peak. That suggests the missing peak is terbium, just like the Wikipedia article says it is. Without really careful calibration, you can't know for sure which of the two green lines is missing, but the brighter peak at 546 nm is a well documented mercury peak. So we can assume that the single peak is the 546 nm mercury peak, and the peak at 542-543 nm in most CFL lamps is not mercury. It's still puzzling why there is such poor information about the wavelength and source of that bright green peak at 542-543 nm in all CFL lamps.

I looked at a few of the cleaner CFL spectra at Spectral Workbench and the two big green peaks are always farther apart than the two green peaks in the reference spectrum displayed during calibration. That suggests that the reference spectrum being used is the one I corrected based on the Wikipedia information that the "terbium" peak was at 543 to 544 nm. It seems that CFLs actually have a peak closer to 542 nm, so the clean spectra at Spectral Workbench that show two peaks show them farther apart than in the reference spectrum. It might make sense to use another spectrum for the reference. This is one corrected for all the Wikipedia peaks but with the "terbium" peak at 542.5 nm: https://spectralworkbench.org/spectrums/show2/58696.

Making that change will not matter to most users because there appear to be very few CFL spectra in which those two peaks can be discerned. But with the current reference spectrum, the best CFL spectra will never match all the major peaks in the reference.

Well, both images were taken from the same device within about 20 minutes, so you can check their images against one another even if they're not accurately calibrated:

As to the reference spectrum in the calibration tool, I used the corrected snowsky spectrum you posted in the comments of this note: https://publiclab.org/notes/warren/09-30-2015/new-wavelength-calibration-procedure-preview-for-spectral-workbench-2-0

So if you have a reference spectrum you're more confident in now, all I need is the image plus a list of the peaks and troughs as shown here: https://github.com/publiclab/spectral-workbench/blob/master/app/assets/javascripts/spectralworkbench/SpectralWorkbench.API.Core.js#L521

And I could update the calibration code.

Great detective work @warren and @cfastie!

Jeff, are these the specs for the image you need for the reference spectrum:
1390 px wide with the blue 436 peak at 211 pixels from left, and the green 546 peak at 742 pixels from left?

Hi, Chris - because we used lots of peaks to determine "fitness", we'd need all of the following nanometer positions plus their intensity (0-255):

Then, we'd need the source image for that reference to calculate the trough positions. It's probably easiest to upload a reference and calibrate it, and I can extract the values that way.

Jeff,

I made a corrected (ideal) version of the Snowy Sky CFL spectral image when we were working through this in October. This is the basis for the Twin Peaks research note. This spectrum is an idealized model of what a perfect spectrum would look like if the major peaks matched the Wikipedia article with the one exception that the 542 nm green peak is not corrected to be at 543.5 nm. This idealized spectrum is the CLF_CFL Corrected SnowySky spectrum.

I checked Dave Stoft's CFL spectra and they confirm that the two green lines are always farther apart than Spectral Workbench's reference spectrum says they should be. If you know of some other good CFL spectra you can see if they confirm this observation.

Here's a weird thing. The spectrum that you are using as the reference spectrum for calibration is my Corrected Snowy CFL spectrum (it has the two green lines too close together). When I try to calibrate that spectrum in Spectral Workbench it's spectral image lines up perfectly with the reference spectrum because the two are the exact same image. But the "Fit" is always turned up to 11. You would think that the Fit would be zero for that calibration, or at least better than 11.

There are a few places where we do math -- like convert from 0-255 to 0.000-1.000, and when repeated over many pixels, its possible that there are accumulated, essentially rounding errors. The odd thing is that I've seen matches as low as error:4, with spectra that are not the corrected snow sky reference.

Could you make a set of good double green line references to look at?

What do you mean by references?

Oh, just the spectra you're looking at to reach those conclusions. Sorry to mix vocab :-)

Ah, set as in SW set! Here is one with CFL spectra from you, me, stoft, and spectra4cps: https://spectralworkbench.org/sets/3165. All of these show the two green peaks to be about 4nm apart. Included in this set is the recommended idealized spectrum CLF_CFL Corrected SnowySky. I never found any CFL spectra with the two green peaks closer together. But I didn't find CFL spectra with distinct green peaks from any other authors (there are too many to look through). Funny, when I use "Find Similar" on one of my CFL spectra, I find only my CFL spectra (IOW, from Ebert).

Correction: When I calibrate Corrected Snowy CFL (with itself), I can get a Fit as good as 5. But the two spectral images are not well aligned when Fit equals 5. When they are visually well aligned, Fit equals 11.