I use Maxim DL to build master calibration files. It provides four main options: average, median, Sigma Clip and SD mask. Frankly not studied up on the mask approach. Of the first three the Sigma Clip seems optimal and I have the impression is frequently used for photometry. I understand the basic concept, however, there are three options related to “normalization”. The description of this in the program manual is sketchy, at least to me, and I have found nothing on the internet that elaborates on it. It is said the point is to bring the involved images to the same intensity before combining but it is not clear if this is average over the tested area or pixel by pixel.
The options are to dispense with normalization, use the background only (no definition of how that is determined) or a linear regression of all pixels in the sample area. Finally there are options to ignore zero level pixels and saturated pixels. Presumably the Sigma Clip would removes such pixels anyway, so I assume this feeds back into the “normalization” process. Not understanding that well enough I can’t judge whether to use these two options or not.
Bottom line, should I normalize and if so use he background or linear regression? And, should I excluded cold and hot pixels from that process? Can anyone with a better understanding of this provide some insight or recommendations on the normalization process options.
Thanks.
Hi James,
I usually don't recommend sigma clip, because it changes the frame characteristics on a pixel-by-pixel basis. For example, if you are stacking 7 frames, pixel (100,100) might average 7 pixels together if all of them fall within one sigma of their mean. Pixel (100,101) on the other hand, might only average 5 pixels together if two of them fall outside of one sigma. The best method of combining is usually median, especially for larger numbers of frames.
Normalization usually is defined as a single constant for the entire master flatfield frame, usually based on the average pixel value within that master flatfield. If the average pixel value is normalized to 1.000, then dividing the flatfield into the science frame won't significantly change the pixel values in the science frame, so it looks about the same as it did in raw format. For most image processing programs, dealing with frames with pixel values of 30,000 or pixel values of 2.4449 make no difference to the photometry or to a scaled display; it is more for the human who is trying to understand the observation. The important factor is a single constant per calibration frame.
Usually hot/cold pixels don't affect the average value for a frame very much, since there are typically very few of those discrepant pixels. Again, since it is a constant, the science is preserved.
There are two other parameters that sometimes come into play when stacking flatfields or other calibration frames: an offset per frame, and perhaps a scaling. Say that you are taking twilight flats, where one flat has a value of 10K per pixel, one flat has a value of 5K per pixel, and one flat has a value of 20K per pixel. You could multiply the first flat by 2 and the second flat by 4, yielding a value of 20K per pixel, and then median-combine the frames to yield a master flat. While the means of the two new frames match that of the deeper third frame, the signal/noise per pixel is quite different. That is why you try to get the mean value of flats to agree with one another as much as possible. It is much easier in this respect with a light box or panel! Your software appears to have even more customization with the combination process. My suggestion is to keep things as simple as possible. My preference within IRAF, for example, is to use median combine on the flatsubs, and to scale each flatsub by the mode of the image (the most likely value). That does a pretty good job of getting rid of stars and cosmic rays.
Arne
Thanks.