I’m giving a talk on Tuesday (12/15/2015) at a UC Davis weak-lensing workshop on DM’s approach to correcting for sensor anomalies (i.e. departures from a square pixel grid with linear response). Here’s an outline of what I’m planning to discuss; I’m hoping those (@RHL, @KSK) who have been paying more attention than I have to e.g. DESC SAWG meetings can help point out anything I’ve forgotten or characterized incorrectly.
Our Model for Image Data
We generally consider the image data we see to be the result of applying the following operations discretely to the true sky:
- Convolution with a PSF.
- Coordinate system transformation (WCS).
- Scaling by photometric sensitivity (flat-fielding).
- Adding noise.
We additionally frequently approximate by assuming:
- the PSF is locally constant (on the scale of an individual object)
- the coordinate system transformation is locally affine (on the scale of an individual object)
- the PSF is Nyquist-sampled on the pixel grid
- the noise is independent between pixels
Astronomers have historically compounded some problems by treating WCS effects as photometric sensitivity effects, but that’s just a mistake, not a problem that’s fundamentally difficult (i.e. we just have to be careful).
It’s also been common in the past to treat all of the effects as wavelength-independent within a single filter. We know we’ll at least need to have wavelength-dependent PSFs and flat-fields for LSST; it’s not clear whether we also need wavelength-dependent WCS.
The sensor effects that worry us are the ones that (ordered roughly by level of concern) break this model of the observational system, void our appoximation assumptions, or make it harder to constrain the model by adding many new parameters.
Lateral electric fields
Edge distortions are a pure WCS effect that’s fixed on the chips, so we’re not worried about being able to constrain our model for it. But they may void the assumption that the WCS is locally linear; as a result we may need to mask the most affected pixels in some contexts.
Tree rings / impurity gradients are a pure WCS effect, fixed on the chips, moderately easy to constrain (lots of parameters, but lots of data). And they’re unlikely to void the assumption that the WCS is locally linear. The result is that I’m not worried about this. May not be significant for LSST anyway.
Tape bumps / lattice stresses are a pure WCS effect, fixed on the chips. Almost certainly breaks linear WCS assumption, but small area, so we can mask them if they’re present for LSST.
Brighter-fatter / charge-correlation effects are a big concern, because they don’t fit into our typical model (they’re not a convolution or a coordinate transformation). For characterization, we’ll rely on lab experiments and physical simulations to come up with a parameterized model, then constrain the model from flat-field image correlations and potentially stars on science images. We’d like to correct as much of this as possible at the pixel level at a very early stage, so most downstream processing can continue to use our traditional model of the observational system. If we can’t completely correct it there, we’d include it in a forward modeling of stars that are being used to constrain the PSF. We would very much like to avoid including it in forward modeling of WL source galaxies, and we think that’s unlikely since they’re faint and hence the effect is small (and hopefully mostly corrected at the pixel level anyway).
Pixel area variations
Like tree rings, this is a pure WCS effect that adds many new parameters, but it’s fixed in the chips so we’ll have a ton of data to constrain it. Important question is whether it breaks linear WCS assumptions at a level that matters; if it does, we may have to resample data at an early stage rather than just include it in the WCS.
Easy to correct at the pixel level; probably not too hard to constrain the model.
Easy but computationally expensive to correct at the pixel level; easy to constrain the model.
Charge Transfer Inefficiency
Negligible for LSST? Lots of literature from HST if we do need to account for it.