Tile dependence of star-galaxy classification

- Alex Drlica-Wagner and William Wester

William was examining the performance of his star-galaxy classification algorithm in the SN-C field when he noticed strange (rectilinear) spatial variations in his star-class objects. Following up in the SV-A1 Gold Catalog v1.0 using the modest star-galaxy classifier, the same features are found.

The region in the center of the field where the most distinct over-density of star-class objects occurs correspond to the coadd tile 'DES0335-2832'.

Tile 'DES0335-2832' spans both the deep field and one of the shallow fields, as does the tile below it (also showing an over-density of star-class objects). At first, this struck us as a potential issue with PSF homogenization; however, the DESDM page says (color included) No PSF normalization was performed. Coadd images are non-homogenized. The SV-A1 Gold modest star galaxy classifier depends heavily on spread_model_i and spreaderr_model_i, specifically in the combination spread_model + 3*spreaderr_model (with a cuts at +/-0.003). Looking at the distribution of this variable across the full SN-C field, things look reasonable (with the star peak being captured).

However, when I recreate this distribution for tile DES0335-2832 and the tiles on the right (DES0339-2832) and left (DES0332-2832) of it, I find that the distribution in DES0335-2832 is quite different from it's neighbors. The modest classifier is thus missing the star peak in DES0335-2832 and selecting dominantly galaxies. Breaking this cut down even further, it looks like the dominant change is in the spread_model distribution, which has shifted to lower values for DES0335-2832

spread_model + 3*spreaderr_model spread_model
So the question now is "What is causing spread_model to change in tiles that overlap to fields of differing depth?" So far, the most basic ideas are:
  • Something in SExtractor determining global parameters for the tile
  • Something in the coadd process itself

One strange qualifier is that the deep portion of DES0335-2832 appears to have more consistent SG separation. This suggests that somehow the deep portion of the tile is dominating the fit/calibration even though it occupies a subdominant region of the tile. Eli suggests that in fact the lack of homogenization may be causing this feature. He has a few additional studies here (and the impact on photometry here).

P.S. It looks like the HST ACS data for the CDF-S field is outside this trouble tile.