search

dstndstn / tractor

The Tractor: measuring astronomical sources via probabilistic inference
Other
86 stars 24 forks source link

photometry on fields with large residuals #75

Open knyland opened 8 years ago

knyland commented 8 years ago

Sorry, this isn't really a software issue, but a question on software usage . . .

I'm currently using tractor to obtain improved photometry on fields with optical, near-IR, and mid-IR data, using Ks-band data at the highest resolution as a prior to constuct the model. In some cases, the sources in each band are well represented by the model and the residuals are noise-like and smooth. However, in other cases, particularly if there are sources visible in the mid-IR that are not visible in the Ks-band data (or if the structure is overly complicated), there is a lot of unmodeled flux left over in the residuals. Clearly in these cases the fitting must be performed again until all of the flux (or as much as possible) is properly modeled and the residual image is noise-like.

This seems like it must be a common issue, and so I'm wondering if there is a recommended strategy for approaching this in Tractor?

dstndstn commented 8 years ago

That is indeed a common issue, but not one I have any worked-out strategies for. Initialization is hard, and so is changing the model complexity. For my SDSS/WISE work, I included all SDSS sources PLUS any sources detected in the WISE images that were not within a few pixels of an SDSS source. Maybe something like that would work for you also -- include mid-IR-only detections in the list of sources.

dstndstn commented 8 years ago

PS, this is something I do hope to work on soon for the http://legacysurvey.org project -- we often get deblending failures leading to only one of a pair of blended sources being detected and a resulting poor fit. I have ideas but no implementation yet.