Detections with SExtractor notes

[andrevitorelli]

Discussion 2021-04-01 10:00

Axel SExtractor:

• before: bkg estimation
• oy can set the size of patch for bkg estimation
• formula with 3median - 2mean
• usually, two passes, second removing the objects detected in first. Not in sx, do it yourself.
• sx does not load entire image to memory
• types of detection: absolute or relative to background
• DETECT_THREH x times sigma of noise
• you can set de min area for detection
• Deblending: each detection, split light profile in bins DBLEND_NTHRESH, min contrast to consider them different. (avoid noise creating false blends)
• sx is very powerful because its optimized, fast in detections/second
• estimate positions:
  -- default: barycenter light profile
  -- improvement: windowed positions - weighted moments, limits effects of noise.

Bastien: window function ~ barycenter but reduces noise.

• detection with 2 images: 1 detection 2 measurements

• even if same image, use 2 copies. slightly different results.

• important: provide weights, flags, etc.. (saturated pixels, masks, etc)

• threshold for flagging: if 10 % of flux comes from other objs -> blended.

• propagate flags if finds objects inside flagged regions.

• sx: star/gal sep: pre-trained NN

• doesn't work at today's depth, trained on old images. works upt to mag 21 at most.

Cecile: can you give another set weights to NN?
Axel: cannot be trained inside sx, but change the file.
Alexandre: deprecated for years...

• some use stellarity(class_star) more than .99 for initial PSF estim.

Alexandre: it still works in bright stars - use in PSFEx.
Axel: PSFEx doesn't use it.

• SNR in sx windowed.

• sx + PSFEX - run sx twice sx -> psfex -> sx

• configure interpolation - "Experimental Stuff"

• sextractor only use certain inputs (assoc, psf) if you ask for results related to them.

default: sex -d, full default: sex -dd

• sourcextractor++ version: 0.14
  -- plug&play for measurements
  -- can provide stack
  -- multiepoch, multiband...

Alexandre: devel by euclid?
Axel: Euclid has own sextractor, but is not public.

• original reworked in C++ fixing weird behaviour.
• sx is detection-deterministic, but not in fluxes.

Alexandre: euclid sx -> euclid wanted C++

Axel: astromatic sofware are made to work together.

• SCAMP (Astrometric solutions and photometric calibrations),

• SWARP (Coadds),

• WeightWatcher (weights, flags),

• SExtractor,

• PSFEx

• SEP: expose functions to work as a library.

• Missing: photometry. (Have the tools to do it, but doesn't, for now)

Andre: How to choose convolution filter for detections?
Axel:

• use simulate images
• mostly you gain faint obj, but prob discarded later because bad quality

-checkimage: can return bkg, img with bkg subtracted, img with detection removed, segmentation map.

Bastien:

• segmentation map: each object contributes to some set of disjoint pixels.
• window function can cause bias: is there an (net) effect... an example?

Axel:

• for positions, it's tricky, but most cases is a good effect, minimises noise. most cases will be better
• in faint objects
• Difference of center is very very small

Bastien:

• future lensing will have many faint objects, may become important

Axel:

• personal preference on windowed position, from experience.
• regularization, safety checks.
  Axel:
• worse thing about sx, some information is missing, you have to look into the code.

Axel:

• documentation (for Dummies) is kinda outdated, but good for regular stuff.