ImageIntegration weighting?

[mschuster]

I used the noise evaluation weighting for an integration and have an observation and a question about the weighting. The table below contains data for 2 of the 18 subs integrated. Sub 1 had the least FWHM (in pixels) and so I used it as the StarAlignment reference. The Uncal, Cal and Align columns are NoiseEvaluation script results for the uncalibrated, calibrated and aligned frames, respectively (all in normalized units times e-4). The noise estimates used to determine the integration weights are in bold. The interpolation used to align sub 2 (Bicubic Spline) produced a result with a lower noise estimate causing it to be weighted more heavily than sub 1, which was not interpolated. So the end result is an integration where the sub with the least FWHM is weighted about half as much as one with a larger FWHM.

In this particular integration, had the weighting been more equal, the result probably would have been almost identical, at least visually. Sub 1 was only 1 frame out of 18 of course and in fact all subs had reasonably good FWHM and SNR. But I am wondering more generally does this make sense, that the "best" sub in an integrated set is deweighted due to its lack of interpolation?

Regards, Mike

Sub	FWHM	Uncal	Cal	Align	Weight
1	0.99	4.13	3.90	-	0.58
2	1.06	4.07	3.80	2.72	1.23

[vicent_peris]

Hi,

This is in our to do list. This problem arises from the smoothing coming from the registration of the image. IMO, the noise evaluation must be done from the second wavelet layer, because the first one is heavily affected by resampling. A workaround is to make a *very mild* smoothing (with ATW or MMT) of the first layer) of the reference image in order to increase its weight.


Regards,
Vicent.

[mschuster]

Thanks Vicent,

FYI: As a test I measured the noise after deleting the first layer with ATWT. The estimate for sub 1 remains higher than all 18 other subs, but to a lessor degree (example sub 1: 8.56e-5, sub 2: 7.50e-5). Maybe resampling affects layer 2 also somewhat, or maybe this is just a residual first layer noise that could be deleted with another application of ATWT.

Another possible workaround might be to add a FITS keyword to _c_r equal to noise of _c and have integration use that value via the keyword weighting option.

Also, I note that the rejection percentage of sub 1 is larger than the others (twice as large), due to its larger dispersion I suppose. Your mild smoothing workaround might address this also.

Regards,
Mike

[mschuster]

Vicent, I tried your smoothing suggestion. It worked well, both weight and rejection are well matched now. A MMT layer 1 bias of -0.04 was required to match the impact of resampling on the noise estimate. I agree resampling does affect the first layer rather heavily.
Thanks, Mike