Reduction of Burrell Schmidt CCD images of M18

Michael Richmond
Jan 17, 2017

Katy Garmany provided a CD with raw CCD images from the Burrell Schmidt, taken on July 2, 1997, covering regions in the neighborhood of M18. This report describes the reductions I have applied to the images.

The dataset

The CD contained a set of CCD images and a text file N970702.txt , which has lines like the following:

a0977.imh       sa105   object  02/07/97  03:44:56  14:59:27  1     4. 2.5  3.

My guess is that the columns mean the following:

  1. file name
  2. target name
  3. image type (object, zero, flat)
  4. date DD/MM/YY (UT?)
  5. UT of exposure
  6. Sidereal time of exposure
  7. Filter index (1 = U, 2 = B, 3 = V, 4 = R, 5 = I)
  8. exposure time (seconds)
  9. gain (electrons per adu)
  10. readnoise (electrons)

The detector was the s2ka CCD; one can find a small amount of information about it at

The raw images fall into a series of groups:

 0953 - 0964        zero images (bias images)

 0965 - 0973        3 sky flats each in U, B, I.  No V or R flats

 0976 - 0980        SA105, one image each in UBVRI

 0981 - 0990        Serpens2A field, 2 images each in UBVRI

 0991 - 1000        Serpens2B field, 2 images each in UBVRI

 1001 - 1010        Serpens2C field, 2 images each in UBVRI

 1011 - 1016        SA105, one image each in UBVRI (plus one extra V)

 1017 - 1021        SA111, one image each in UBVRI

 1022 - 1031        Serpens2D field, 2 images each in UBVRI

 1032 - 1041        Serpens2E field, 2 images each in UBVRI

 1042 - 1051        Serpens2F field, 2 images each in UBVRI

 1052 - 1060        Serpens2G field, 2 images each in UBVI, 1 in R

 1061 - 1065        SA111, one image each in UBVRI


So far, I have examined only the images of the "Serpens2D" field, which is the only field which includes the center of M18.

Image size and layout

Each CCD image is 2080 columns by 2048 rows. The layout is like so:

                                             32 overscan cols
        <----   2048 data cols  -------------> <------>
     ^  |                      East            |      |
     |  |                |                     |      |
     |  |                |                     |      |
     |  |                |                     |      |
     |  |              | |                     |      |
        |              | |                     |      |
  2048  |  West        | |              North  |      |
  rows  |              | |                     |      |
        |              | | bad                 |      |
     |  |              | |  cols               |      |
     |  |              | |                     |      |
     |  |              | |                     |      |
     |  |              | |                     |      |
     |  |              | |                     |      |
     v  |              | |    South            |      |

There are a pair of bad columns (each bad region is actually several columns wide) near the middle of the chip, centered around cols 846 and 867.

The plate scale, based on a simple linear projection model of astrometry of SerpensD U-band images, is about 2.030 arcsec per pixel.

Removing bias via the overscan region

I used an image processing package I wrote myself, XVista, to reduce the raw images and perform most of the analysis. It has the severe limitation that it operates only on 16-bit integer FITS images. Since the raw images were in 32-bit floating point format, with data values between 0 and 65535, the first step was to convert the images into 16-bit integer format. In order to retain the information in the raw pixel data in the range 32768 - 65535, I divided all pixel values by 2 during the conversion. Thus, the entire range of pixel values was compressed into the smaller range 0 - 32767.

This division discards the lowest-order bit in the original data values. There is therefore a loss of some information. Since the noise due to the readout and sky background is typically several ADU, this loss is for the most part not significant.

The first stage in the reductions was to remove the bias contribution from each image. I used the overscan region to estimate a constant bias value for each image. In particular, I computed the mean value within this box

   rows 0 - 2047   cols 2050 - 2074

and then subtracted that value from every pixel in the image.

The bias values determined in this fashion throughout the dataset are shown below. I guess that the large jumps occur after telescope motion.

After subtracting this bias value, I removed the overscan region from each image.

I examined the "zero" images after this bias subtraction value to see if any further bias subtraction was necessary. I combined the overscan-subtracted "zero" images via a median technique to create a "master zero" image. There was no significant structure in this master image, aside from the bad columns. The histogram of pixel values showed almost all values were -1, 0, +1 counts.

I conclude that it is not necessary to subtract a "master bias" from the target images; simply subtracting a constant based on the overscan region is sufficient.


I created master flatfield images in U, B and I by combining the bias-subtracted images in each passband. After scaling each of the 3 sky flats in each filter to a common mean value, I computed a pixel-by-pixel median.

The resulting flatfield images are shown below. The orientation of these images is North to the right, East up.




The variations from corner to center are relatively large: about 25 percent in U-band, B-band, and I-band.

Cleaning the target images

After subtracting a fixed bias value from each target image, as described earlier, I then divided the result by the appropriate flatfield image; as there are no V-band or R-band flats, I have not cleaned target images in those filters.

The resulting "clean" images have an unfortunate appearance: below is a picture of one of the B-band images.

The greyscale runs from

     black = -3 counts      white = 20 counts

so this is a very contrasty display. You can see a faint dark circular region in the center of the frame, with the same size and shape as the bright region in the B-band flatfield image.

It appears that the (twilight?) sky flats may not reflect exactly the response of the system to light from celestial sources. Nonetheless, I applied the flatfield corrections to U, B, and I target images, since it should certainly improve the small-spatial-scale, pixel-to-pixel sensitivity variations, which are important for stellar photometry.

Finding and measuring stars

Once the images are clean, we can find stars in them and measure the stellar properties. I used routines similar to DAOPhot's find and phot to find stars and perform aperture photometry.

The measured FWHM ranged from about 1.8 to 2.2 pixels, better in B and I, worse in U.

Some parameters of interest in the finding and measuring process were set to

   parameter      min          max
    FWHM          1.4 pix     3.0 pix     

    sharpness     0.6         0.95

    roundness    -0.5         0.5

    minsig       10.0

I used apertures of radius 3, 5, and 10 pixels to measure the light of each star; pixels falling partially within the circular aperture had a fraction of their light included in the sum. A local sky value was estimated from an annulus of radii 20 and 30 pixels.

The number of stars found in each image was

   image      filter         stars
    1024         U           1592
    1025         U           1729

    1026         B           3931
    1027         B           3936

    1030         I           5230
    1031         I           5670

Ensemble photometry

We have two images of each field in each filter, so we can check the properties of each star. In order to compare the measurements in each pair, I employed the ensemble package , which is based on the methods of inhomogeneous ensemble photometry as described by Honeycutt (PASP 104, 435, 1992).

Two useful pieces of information created by these routines are

All the results shown below are based on measurements made in the smallest aperture: radius 3 pixels. I can quickly compute similar quantities for measurements made through the larger apertures.

Let's look first as the frame-to-frame shifts for each pair of images. First, the U-band:

This graph shows that there was an offset of about 0.025 magnitudes between the two frames, in the sense that the second image had slightly smaller instrumental magnitudes (meaning that all the stars appeared slightly brighter).

The image-offset graph for B-band shows a much smaller offset of 0.002 mag:

The image-offset graph for I-band shows an offset of about 0.010 mag:

We can also examine the common "sigma-versus-mag" graph, which shows how the scatter in magnitude (in this case, something like the offset between the two individual measurements) as a function of the mean instrumental magnitude. We expect to see small scatter for bright stars, and large scatter for faint stars.

For the U-band, we see the expected pattern:

And again in B-band:

And again in I-band:

With only two images for each filter, these graphs don't carry a great deal of information.

Astrometric calibration

In order to convert the image (x, y) coordinates to (RA, Dec), I began with stellar positions from the All-sky spectrally matched Tycho2 stars (Pickles+, 2010) . This is not primarily an astrometric catalog, but since it has photometry for stars in the fields, it is convenient.

I extracted a list of stars in the area covered by each image, projected the (RA, Dec) values onto a plane, and tried to match them to positions of stars found in the images. For details on this procedure, see the match package . The matching procedure allowed for arbitrary translation, rotation and scale, and a linear radial model. Stars had to fall within 2 arcseconds of each other after fitting to count as "matched."

   filter        number of stars      RMS in position (arcsec)
                     in fit                x          y
     U              166                  0.28        0.25

     B              168                  0.17        0.16

     I              168                  0.17        0.19


Photometric calibration

The photometric calibration of the data is, at the moment (Jan 17, 2017), in a primitive state.

I am once again using the All-sky spectrally matched Tycho2 stars (Pickles+, 2010) catalog for this purpose.

My current method is very simple:

This method has a number of deficiencies:

However, all of these improvements must wait for the future.

The results of the current method of calibration can be summarized as follows:

   filter        number of stars        delta mag             RMS of
                     in fit         (calib - ensemble)      delta_mag
     U              114                  7.67               0.29

     B              133                  7.25               0.27

     I               99                  5.99               0.09


Output files

The results of running my reductions on the Serpens2D field only, and the U, B, and I-band images only, are three data files, one per passband. Each has the same format; I'll choose the U-band file to illustrate.

First, there is a header with some information:

# astrometric calibration performed Tue Jan 17 11:51:38 2017 
# original input file solve_m183_u.sig 
# reference field center 274.75 -16.7666666666667 
# match diagnostics follow 
# sig 3.2955e-12   nr 166   nm 166   
# sx 1.3796e-06   sy 1.2052e-06     
# photometric calibration performed on Tue Jan 17 11:51:38 2017 
# photometric output in U passband 
# photometric diagnostics follow 
# num_star 114 delta_mag   7.6707 delt_mag_uncert   0.2825 

Following this header is a series of ASCII text, one line per star. The first few lines look like this:

 0 2.649 0.001 0 0.07 1.73641 436.18667  275.27958  -17.13357   10.3197   0.0010 
 1 4.761 0.001 0 0.13 2.45974 1476.69000  275.24875  -16.54796   12.4317   0.0010 
 2 7.232 0.126 0 2.57 2.52641 1366.71000  275.25192  -16.60986   14.9027   0.1260 

The columns here are: