Models, PSFs, Deconvolution, etc

Is there a paper which describes, in detail, how the DECaLS models are created?

I see that there are PSF, EXP, DEV, and COMP. When I look at the models in the Browser, some look good, but others are obviously very unphysical (as already noted in another thread), and I’d like to learn more about how they were created.

That includes understanding what deconvolution was used, and which PSFs were part of that process.


We have a paper in prep on that. In short, it’s not deconvolution, it’s forward-modeling – we make pixel-by-pixel predictions of the flux in each pixel in each exposure, based on parametric models, and then optimize those models. The PSFs are spatially-varying pixelized model estimated using PsfEx.


1 Like

Thanks! :slight_smile:

Let’s see if I understand: suppose there’s one g-band exposure, and one r-band one. You create two models from the (raw) data, one for each exposure. If there’s more than one exposure for a band, then you have one model for each. And somehow you “optimize” these many models. Is that about right?

I guess you have a method for handling artifacts and transients, and there could be very different models in different bands, for example a large EELR (extended emission line region) which is invisible in all but one band,

Oh, and what’s “PsfEx”?

Looking forward to reading the paper (I hope it won’t be behind a paywall).

Oh, I forgot, we already posted the article to the arxiv:

We’ll have one model for each galaxy. The model says what the intrinsic distribution of light is for that object, including its brightness in each band. Then, for each image we have the atmospheric transparency, the filter, and the PSF for that image (and the astrometry), which allows us to transform model space into pixel space. We then optimize (tune) the model parameters (positions, shapes, and brightnesses of galaxies) so that the predictions best match the observed pixel values.

We do have artifact flagging, and then we just ignore pixels that we think are contaminated by artifacts. Currently we don’t have any special handling for transients, so if there is a transient in one image, we usually fit it as a point source with a weird color.

PsfEx is


1 Like