Image Processing
or
From photon to photograph, and back again
Andrew Bradshaw → DP0 Delegates
20211203
with thanks to the camera and DM teams
And extra special thanks to: Aaron Roodman, Yousuke Utsumi, Jim Chiang, Craig Lage, Merlin Fisher-Levine, Johnny Esteves, HyeYun Park, Adam Synder
“How the sausage is made”
Recipe for Rubin images in 3 steps:
This talk will be about the Step 2, where wonderful things happen behind the scenes to make and prepare images for science
My background: researcher at SLAC working on camera hardware & software, physics PhD from UC Davis, Texas A&M physics major
Step 0: Build the camera!
Step 1: Collect photons (tests ongoing)
Bias frame, visualized as read noise in each channel
“Dark” frame, with <3? light leak
Flat field illuminator→ →
Under the camera →→
Flat field image
Final pre-assembly imaging tests & optimization underway!
“Classical” ISR
Instrument Signature Removal
Raw images need a few calibrations before basic science can be done. In a spherical cow approximation raw exposures are composed of:
Astronomical light
x camera response to light (QE, gain)
+ camera response to dark (traps, cosmic rays)
+ camera response w/o CCD (bias, electronic effects)
= Raw science images
Therefore, to turn raw images back into astro-photons,
simply invert the process:
Science = (light - dark) …easy!
flat
Credit: Steve Majewski
ISR for Rubin (and DC2 DP0)
All the processing steps with inputs and outputs are shown on the right in a (purposefully unreadable) so-called “quantum graph”, or “directed acyclic graph”, which orchestrates jobs
Credit: Jim Chiang & DC2 team
ISR for Rubin - A checklist
CCD images have always been complicated, but until now the systematic complications were smaller than statistical errors
Rubin LSST will have tiny (<1%) statistical error bars on a wide variety of science topics, so systematics must be removed to far greater precision, so
Science image ≠ (raw - bias)/flat
And DP0 puts these complications to the test!
Image: Merlin Fisher-Levine
ISR for DC2 - Insert, Undo, Undo!!
To the beautiful images made by ray-tracing the simulations of large scale structure, the sensor simulation code (imSim) adds fixed pattern noise, charge transport, and electronic readout effects including:
All of these effects have models that realistically modify the input electron images, and ISR proceeds as it would on the mountain, except using simulated calibration data
Let’s take a closer look at the first three systematics: tree rings, brighter-fatter, and crosstalk
Making things harder for ourselves
Tree rings
In reality: silicon boule growth causes dopant (ion) impurity gradients which shifts charge
In DC2: a model of the radial displacement function modifies where the photo-electron lands
HyeYun Park
Tree rings in real image measurements
Studies show systematic 10-3 pixel shift, tenth of a percent effect on shapes
Question: is this sufficiently corrected in DC2, to the level required by e.g. 3x2pt?
Johnny Esteves
Brighter-Fatter effect
In reality: collected charges repel subsequent charges, increasing PSF size for bright vs. dim objects
In DC2: linearly superpose pixel boundary displacements to calculate which pixel electrons land in
Craig Lage
12
Poisson simulation of the brighter-fatter effect
(arxiv.org/abs/1703.05823)
No charge in central pixel 160k electrons in central pixel
Correction method: model pixel boundary shifts (WCS) or redistribute charge after exposure (kernel)
Brighter-Fatter effect in real images
Surveys are just beginning to correct for this, but the signature is usually seen to magnitude vs. size distributions
Deep LSST-like images from Suprime Cam
Lab data with stars (top) and galaxies (bottom) indicate variable correction
Crosstalk
In real images: bright objects (like satellite streaks) produce positive and negative crosstalk, ~10-4
In DC2: pre-amplifier images are contaminated with nearby amplifiers according to linear model
Crosstalk corrected
One of the first things that needs to be done, perhaps before and after bias removal & gain corr.
In reality, crosstalk correction may be flux dependent (non-linear crosstalk)
Two stack club notebooks about image processing
Lab data (IDF) stars and galaxies, brightening
Galsim stars showing small scale bias in analysis
Two stack club notebooks about image processing
Lab data (IDF) stars and galaxies, brightening
Galsim stars showing small scale bias in analysis
Summary and Conclusions
Rubin LSST will be systematics limited
DC2 simulations have many realistic effects added, but still just the beginning
Simulations based in reality, and corrected as we would if the survey started now
But are the simulations and corrections good enough? You can be the judge!
Ways to test for un-corrected systematics:
Bonus slide: DC2 universe on Rubin CCDs!
Input: DC2 @ Hubble resolution Output: blurry DC2 on Rubin pixels Setup: paper + pinhole!
(thank to Michael Troxel!)