rdnoise and snoise in scombine

Jaehyon Rhee (Jay) wrote on Jun 17, 2010

IRAF:

I am trying to combine echelle spectra using the scombine command.

(1) Could you explain exactly what snoise (sensitivity noise as a fraction) is?

(2) The help page of scombine says, "Note that if the spectra have been extracted from a 2D CCD image then the noise parameters must be adjusted for background and the aperture summing." How can I get adjusted values for rdnoise and snoise?

Thanks!
Jay

---

Francisco Valdes wrote on Jun 17, 2010

Hi,

I suggest you look at the help page for imcombine from which scombine derived. The idea is that one has a analytic representation or model for the variance in a pixel which is based on the pixel value (or DN). This model is straightforward for a single CCD exposure. But if the image is processed in some way it currently requires the user to figure out how the model changes. Something simple like subtracting a constant is not too hard to understand but if pixels are combined (such as by extraction from 2D to 1D) it becomes a challenge.

In this case it may be simpler to derive coefficients for the model (the two or three parameters rdnoise, gain, and snoise) empirically.

A word about the sensitivity noise -- this was put in because it was used for HST data. As you see in the equations in the imcombine help, it is basically the higher order dependence of the pixel variance with signal level. I don't think this option is much used and you should not feel you need to use it. For the spectra you can either just use a rdnoise, which is essentially saying there is a constant variance, or figure out how the variance changes with signal level and represent that with the two component model of rdnoise and gain.

I hope this helps. There is not much more that I can contribute.

Yours,
Frank Valdes

---

Jaehyon Rhee (Jay) wrote on Jun 17, 2010

Hi Frank,

Thanks so much for your help!

(1) I can think of the following way to obtain empirical gain, and I would like to know if this method is reasonable.

- Gain (in e/DN)
Compute the mean (in DN) and RMS (in DN) over a high and flat (horizontal) continuum region (e.g., using m keystroke in splot). Then, obtain its gain by using the following relations.

SNR = mean/RMS = sqrt(gain*mean)

(2) I have given some thoughts, but I cannot think of a good method to measure empirical read noise from an extracted spectrum, because every region is contaminated by the Poission noise of the spectra. What would be a reasonable method to obtain its read noise (Gaussian noise)? Any suggestion?

Thanks!
Jay

---

Francisco Valdes wrote on Jun 17, 2010

Hi Jay,

What you describe is reasonable. In this case you are making the assumption that the noise scales with the counts as in pure Poisson statistics. The other approach is to assume the noise is constant independent of signal in which case the RMS in that flat region would be the rdnoise and the invgain would be 0. You can do it either way. In the case of assume a scaling of the noise with signal it would be ideal if you could find two clean regions at different continuum levels to see that the RMS does increase with the higher continuum. Personally unless I can really show that a gain is meaningful I would just assume a constant noise and use the RMS to set the rdnoise parameter rather than the gain.

I hope this makes sense.

Frank

---

Jaehyon Rhee (Jay) wrote on Jun 17, 2010

Hey Frank,

Thanks for your insight!

I just want to make sure if my understanding is correct.

For ccdclip and crreject, sigma = sqrt[ (RN / g) ** 2 + / g + (s * ) ** 2 ] .

So, assuming s (snoise) = 0, we have to provide g (gain) always whether we assume RN (readout noise) is zero or not.

Is my understanding right?

Best,
Jay </div> --- **Jaehyon Rhee (Jay)** wrote on Jun 17, 2010

The equation in the previous message should be this:

sigma = sqrt[ (RN/g)**2 + <data value>/g + (s*<data value>)**2 ]

--- *Last post on Jun 17, 2010*