Background Model and Stacks

(Up to the Extra IPP Features Page)

The IPP performs the following sequence of operations to generate a stacked image:

  • We start with a raw image: Raw(x,y)
  • Burntool : We subtract fits to the persistence trails from bright stars (from the immediate image and those left behind by previous images). Affected areas have a 'suspect' mask bit raised in the mask image.
  • Detrend :
    • We adjust pixel values to compensate for a bias sag
    • We subtract a dark model of the form (C_0 + C_1 * exptime + C_2 * exptime * dettemp + C_3 * exptime * dettemp2). this is a function of (x,y).
    • We multiply by a flat-field response F(x,y). This is measured from a flat-field screen in the dome, then modified based on photometric observations of stars (only for spatial frequencies < 1/1200 pixels or so).
    • for y-band, we subtract a fringe frame fitted to the fringe pattern.
  • Each exposure has a 2D background model subtracted: this is effectively a high-pass filter. In fact, it is a high-pass / low-pass filter: the individual chip images have the low-spatial frequency model subtracted; the low-spatial frequency model itself is saved for each exposure.
  • Each exposure is warped to a standardized pixel grid in a flux-conserving process. The output image products are called 'skycells' and represent about 1/75 of the focal plane (~22 arcmin on a side).
  • sets of 'skycells' are combined in the stack with outlier rejection. This is not a median, but a weighted mean with sigma-clipping. The effect is similar to a median: the resulting image consists of the temporally static signal.

Considering the data as observed by the telescope, there are several important aspects:

  • There is the (nearly static) instrumental response.
  • In addition to an instrumenal response, the signal landing on the detector consists of a true astronomical signal, of which there are dynamic and static components; and a terrestrial and/or contamination signal, which has a signficicant dynamic component.
  • the dynamic portion of the astronomical signal is nearly all PSF-like
  • the dynamic portion of the terrestial signal has a wide range of spatial frequencies:
    • star glints can be roughly PSF in width
    • ghosts can range from ~10 pixels to a couple hundred pixels
    • moon glints tend to be hundreds to thousands of pixels in scale
    • sky gradients are large-scale, but not completely spatially
    • linear.

I contend that we can recover the low-frequency component of the astronomical signal in the stacks from the collection of background models. There is one model per exposure, but with ~25k pixels instead of 1.4G pixels. The goal is to determine the temporally static component of the sky as seen in those models. To do this, we would transform them to a binned version of the celestial coordiate system of the stack skycells. We could then simply generate a median image of the that portion of the sky. With enough inputs, this would filter out the spatially varying terrestrial / contamination signal, leaving behind just the astronomical portion (possibly with an overall gradient). This model can then be added back to the stacks.