Wednesday, October 24, 2012

Photometry: linear regression fitting of sky

Yesterday we agreed that the photometry procedure should have as little arbitrary procedures, constants, etc., as possible. So I'm back to the photometry measurements again...but that's a good thing, as I didn't feel entirely happy about them.
The idea I had this morning on the train was simple -- I don't know what conceptual block prevented me from going this way sooner.
Basically, having a cumulative flux curve (the growth curve) is not generally helpful, as the growth of flux outside is non-linear (depends on geometry as well as on the sky value). However, if I normalise the flux profile to _flux_per_pixel, it should theoretically flatten far away from the galaxy. The slope of a linear regression fit should show the quality of the sky value -- the level of variations. If the sky slope is negative and pretty large, then we probably still are within the galaxy.
If the slope is reasonably small (here go the arbitrary parameters again..), simply taking a mean of all measurements within the ring would give a reasonably good sky value.
The catch is getting the width of the elliptical ring used for fitting right. (I can get its distance by waving my hands, taking the maximum distance from my previous measurements, multiplying it by pi/2 or something. We're testing for it anyway).
However, the width of this ring is a tradeoff between accuracy and precision. Taking a wider ring would surely help to reduce the scatter due to random noise, arbitrary gradients and so. However, the possibility to get a (poorly) masked region or some sky artifact, etc. inside this ring also increases.
I tested it a bit using scipy.linalg routines, so far the slope was below 10^-4 counts.
The growth curve itself is useful as a sky subtraction quality check.

No comments:

Post a Comment