Description

Overview

The refpix step corrects for additional signal from the electronics using the reference pixels.

Reference Pixels in Data

The WFI has two sets of reference pixels: a 4-pixel border of reference pixels around the science pixels, and the Amp33 reference pixels which are a 4096 x 128 section of the detector adjacent to the border pixels.

In the data files, the storage location of the reference pixels depends on level of processing.

A Level 1, uncalibrated image has one array that contains both the science pixels and the border reference pixel, and a separate array for the Amp33 pixels.

A RampModel, which is created during the dq_init step, represents a dataset at any intermediate step between Level 1 and the final Level 2 image. Like the Level 1 file, RampModels also contain an array with both the science and border reference pixels, and another with the Amp 33 reference pixels. In addition to these arrays, there are four more arrays that contain the original border reference pixels (top, bottom, left, and right), and an additional four for their DQ arrays. The border pixels are copied during the dq_init, so they reflect the original state of the border pixels before any calibration. The border pixels that are still attached to the science data in the RampModel will later be discarded when the Level 2 image is created. Note that the border reference pixel arrays each include the overlap regions in the corners, so that each slice contains the full span of border pixels at the top, bottom, left, or right.

In the Level 2, calibrated image, the data array only contains the science pixels. The border reference pixels are trimmed from this image duing ramp_fit. The additional arrays for the original border reference pixels (which are 3D) and their DQ arrays, and the Amp 33 reference pixels, remain in the Level 2 file.

Discretization bias & reference pixel correction

The analog-to-digital conversion in the Roman electronics performs an integer floor operation that biases the downlinked signal low relative to the actual number of photons observed by the instrument. The equation for this “discretization bias” is given by:

\[\mathrm{bias} = -0.5 - 0.5 \frac{N-1}{N} \, ,\]

in units of counts, where \(N\) is the number of reads entering into a particular resultant. This is a small effect. The constant \(-0.5\) term is degenerate with the pedestal and has no effect on ramp slopes and therefore on the primary astronomical quantity of interest. The second term, however, depends on the number of reads in a resultant and may vary from resultant to resultant in Roman. This, if uncorrected, can lead to a bias in the fluxes we derive from Roman data for sources.

However, we need take no special action to correct for this effect. The reference pixels are affected by the discretization bias in the same way as the science pixels, and so when the reference pixels are subtracted (roughly speaking!) from the science pixels, this bias cancels. Exactly when this cancellation occurs depends on the details of the reference pixel correction step. Presently the reference pixel correct includes a component that removes trends across each amplifier and frame using the reference pixels at the top and bottom of the amplifier. This removes the discretization bias.

We note that even if the discretization bias were not removed at the reference pixel correction stage, it could be corrected at the dark subtraction step. Provided that dark reference images are processed through the usual reference pixel correction step, they will have the same biases present in the reference-pixel-corrected images. We have decided to perform the dark subtraction of Roman images via subtracting precomputed images for each MA table rather than scaling a fixed dark rate image by the mean time of each resultant. These precomputed dark images will contain not only the dark current but also electronic effects like the discretization bias. However, it is better to correct this effect during the reference pixel correction so that the dark reference images better represent the dark current and can be more easily used to compute Poisson uncertainties stemming from dark current.