Optimal Image Combining

Part I: Aligning Images

At this point, you have a slew of short exposures that you would like to combine into a single image. The problem of course is that the spectrum of your object is in a different part of the array in each exposure. No problem! You may have noticed that phxproc claimed it was "Squashing object images to make alignment images..." while it was doing its work. What it did was take 10-rows of each processed image and averaged those rows to a 1D slice across the array -- so the continuum of our sources shows up as a peak in each slice. Once we centroid that peak for each slice, we have the information we need to align each image precisely atop one another. Additionally, the flux in each spectrum can be measured to set the scaling parameters for image combination.

So lets do that! The file containing the names of the slices is alnlist, and we will use the task splot (in the oned package) to interactively perform the centroiding and flux measurements. Inspect the file list and make sure it has the files you want (it may also contain files from a previous phxproc run). Also remove any old splot log files:

!rm splot.log

splot @alnlist

To mark the centroid and measure the flux, hit e at the continuum level on one side of the peak, then e on the other side. While you can use k instead, to fit a Gaussian profile, properly-focused pre-1999 Phoenix data had a double-peaked profile that isn't well fit by a single Gaussian. Using e just blindly sums the flux and makes the fewest assumptions about the distribution of flux, which is why I still recommend it.

Hit q to move onto the next frame after measurement.

When you're done, we're ready to assemble the images.

Part II: Combining Images

Now we're ready to assemble our images into a single frame. A lot actually happens in this step, and it's worth spending a moment to discuss what we want to do:

images should be aligned (each image should have an offset relative to some fiducial)
Images should be scaled and statistically weighted by their flux levels.
If an image has been processed with a bad pixel mask, be sure to flag those marked pixels as unusable, which will let other frames in the combining operation fill in that pixel with good data
Apply statistics for each pixel and reject errant pixels
Apply a minimum/maximum threshold for good data, reject outliers
Offsets computed by splot are floating point numbers that will otherwise be rounded to the nearest integer. To minimize "bleeding" of the spatial axis, magnify all images before computing offsets. Contract the final image to the original scale. [Example: an offset of 12.25 pixels will get integer-rounded to 12 if no expansion takes place. However, if the image is magnified by a factor of 4 (i.e. all images temporarily become 1024x4096 pixels), the offset becomes 49 pixels and no round-off occurs. Once combined, the final 1024x4096 image is contracted back to 256x1024 pixels.]
Any slope of the spectral lines should be removed (usually unmeasurable).

As you can see, we've done a lot of thinking here. :) All of this is done by the Phoenix task ircombine. First, let's use nalign to turn splot's log file into useful information for ircombine:

nalign splot.log 4

This example states our intentions to magnify images by a factor of 4 during image combination. nalign creates a new imlist file, as well as files containing the new integer offsets, flux scaling for each image, and statistical weights for each image. Now it's time for ircombine:

Click here for text parameters

Note that we are combining in the average, not in the median, and that offlist, scalelist, and weightlist contain the offsets, scaling factors, and statistical weights, respectively. Pixel values between -100 and 1100 are considered acceptable; this must be determined by inspection for each data set. We are using the avsigclip pixel rejection algorithm, with a very conservative 4-sigma cutoff.

Hit :g, and ircombine will crunch through the images and provide a final combined image. If you used a large-ish (>2) magnification factor, this process might actually take a while.

Here is the combined spectrum for S140:

Using PHXPROC

Reduction Home

Extracting 1D Spectra go
forward

Craig Kulesa

Last modified: Sat Jul 21 01:26:34 MST 2001