I started with standard processing: dark subtraction and flatfielding of the data (itself non-trivial with DECam). The next step was sky subtraction. I have a good way of doing this for DECam images of ordinary star fields. I fit the sky background of each image with a 2-dimensional polynomial, typically of order two or three. I do the fitting iteratively, using customizable creeping-mean rejection of outlying points above and below the fit. In dense starfields, the creeping mean can be set purposely assymetrical: e.g., to reject the brightest 20% of pixels and the faintest 10%. A setting can usually be found that effectively rejects all the stars and accurately fits the sky. With DECam, the upper and lower halves of each image have to be fit independently, because there is a bias offset between them that does not appear to be consistent from one image to the next.
But the starfields of the Milky Way behind Messier 7 are beyond dense. There is no sky between the stars to measure. My polynomial fits were hopeless.
I had anticipated this problem, however. I made sure that I took several images of a much less dense field of photometric standard stars near the time I observed Messier 7. These were 30 second exposures vs. the successive 1, 10, and 100 second exposures I used for Messier 7, but still I believed they would give the sky background I needed. I extracted the background as follows.
First, I applied my regular program for subtracting iterative polynomial fits to the sky background on the standard star images. Then I subtracted the new, background-cleaned version of each standard star image from the original version. The result was an image consisting of only the smooth, polynomial model obtained for the sky background. I obtained 5 to 6 of these images in each of the g, r, and i filters. Then I normalized them all to have a mean of 1.0 on the DECam detector that I have adopted as my reference (detector N04), and median-stacked these normalized sky-model frames to produce a master sky-model frame for each filter.
I then identified tiny areas of very dark dust cloud in the Messier 7 images, which I believed preserved something close to the true sky background. I measured the average pixel brightness in these regions and on the corresponding regions on the master sky-model frames. The backgrounds on the 100-second images of Messier 7 were typically 250 ADU in the g filter, 700 in the r filter, and 1300 in the i filter -- thus, I could approximate the true sky background by multiplying the normalized master sky-model frames by approximately 250, 700, and 1300 respectively. However, I did not trust my dark regions to be absolutely dark, and I did not want to produce negative pixel values anywhere. Thus, I decided to scale the sky-model images to match 90%, rather than 100%, of the measured brightness in the dark-clouded regions of the science images. This is equivalent to assuming that even in the dark, dusty regions, 10% of the light comes from unresolved stars or nebula and only 90% is true telluric sky background. After scaling the sky model images, I of course subtracted them from the corresponding images of Messier 7. I did not try to correct the 1 second exposures this way, only the 10 and 100 second exposures. The sky brightness values measured for the 10 vs. 100-second exposures did differ from one another by about a factor of 10. While not in itself a proof that my method was working, this did constitute an encouraging sanity check.
I encountered a problem with the bias offsets between the upper and lower halves of each image. These were not consistent between the standard star images used to calculate the sky background and the Messier 7 images. Thus, the sky-subtracted Messier 7 images had discontinuities in mean brightness between the upper and lower halves of individual DECam detector images. Reasoning that it was impossible to accurately measure or correct the bias offsets for the science frames, I simply wrote a program to remove the mean offsets between the upper and lower images halves in the master sky frames. Thus, any bias offset remaining in the Messier 7 images would at least be intrinsic to those images and not exacerbated by the sky subtraction. In fact, any such biases turned out to be negligible at least for the 100 second exposures.
I had 5 to 6 images of Messier 7 at each exposure in each filter. I obtained an independent astrometric solution for each detector using stars from the 2MASS catalog (of which there were on average about 2600 yielding accurate measurements on each detector). I then resampled the individual detector images onto a single astrometric grid spanning 32000x28000 pixels. The individual resampled images were median-stacked, with the gaps between the images masked. I had used a well-optimized dithering strategy, so no internal gaps remained near Messier 7 after the 5-6 images in each set had been stacked.
Bright, severely saturated stars in DECam images produce long horizontal streaks due to charge-bleeding. Where one of these streaks extends to the edge of a detector, it affects the region below it by depressing the pixel values by a large and spatially varying additive constant. I removed this effect by masking all negative pixels prior to the final stacks. Even though this widened the effective gaps between detectors that had bright stars on them, it still did not create internal gaps in the final stacked images.
The starfields in the 100 second stacked images were impressive, but the long charge-bleed streaks from the bright stars were unsightly. I attempted to remove them by replacing all saturated pixels in the 100 second stacked images with 10.0 times the corresponding pixel values in the 10-second exposure stacks. This helped, but the resampling had caused the boundaries of saturated streaks not necessarily to have pixel values above the saturation threshold. Thus, uncorrected outlines remained wherever there had been saturated bleed-streaks. Additionally in many cases the pixels immediately adjacent to the bleed streaks were seriously affected even though no charge had bled into them directly. I dealt with these problems through a two-step approach. First, I wrote a program to change the brightness of pixels adjacent to saturated pixels by setting them too above the saturation threshold, so they would be sure to be identified as deviant and then replaced. I applied this program to the images prior to the astrometric resampling. Second, I changed my astrometric resampling program so that if any of the four pixels adjacent to a given location was above the saturation threshold, it would assign the pixel value at that location in the resampled image using nearest-neighbor resampling rather than bilinear interpolation. This preserved the hard edges of the bleed-streaks, preventing the streak-boundary issues previously caused by the resampling. The median-stacking could in some cases yield values of half the true saturation threshold for pixels at the edge of the bleed streaks. This last problem was easily rectified by setting the nominal saturation threshold for pixel replacement to just under half the actual saturation threshold.
This finally produced acceptable results. The saturated bleed-streaks from the 100-second images were effectively removed without noticeable residuals. Much shorter bleed streaks remained around the very brightest stars, because these had been saturated even on the 10-second exposures. I tried to correct them using data from the 1-second exposures, but it was too noisy and the background control was insufficient for these short, bias-dominated images. This is not necessarily hopeless -- for example, these exposures might be so short that polynomial-fitting sky subtraction could work on them directly despite the richness of the starfield. However, I decided my results were sufficiently good already -- in fact, the 10-second bleed streaks seem aesthetically desirable in some cases -- and I should not spend any more time on this problem. I stacked the images in GIMP, using the color assignment i=R, r=G, and g=B. The GIMP processing took hours longer than I expected due to GIMPs inefficiency on very large images. However, at last an excellent-looking RGB image was produced. I saved it as a JPEG and cropped away the ragged edges to yield a rectangular image.
Besides its potential public appeal, the fits files that went into this final image can be used to extract accurate photometry and astrometry from millions of stars. Scientific potential for such an extraction includes the identification of faint, high proper-motion objects, including previously unknown, low-mass or ejected members of the open cluster Messier 7 itself. I have no time to obtain photometry/astrometry on these images now, however, nor to process the u-band data that I also took at the same time (which may in fact have the greatest scientific value). At present, however, I already have a beautiful picture of Messier 7, and one with future scientific potential. Here is a woefully down-sampled version of it:
No comments:
Post a Comment