Multi-threshold vs. Topological Clustering: A deblending comparison - - PowerPoint PPT Presentation

1

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Multi-threshold vs. Topological Clustering: A deblending comparison

Valerio Roscani PhD student – INAF/OAR LSST2018 Project & Community Workshop August 15, 2018

2

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Blending has a non-trivial impact on the Euclid survey, even more on LSST survey (PSF 3-4 times wider) Estimated number of blended sources will be about 40% of all LSST objects (Dawson et al. 2016) Main Survey LSST overlapping Euclid area 7000 deg^2 (46%) - 11000 deg^2 (73%) (Rhodes et al. 2017) The ideal situation is to find an algorithm valid to both the surveys Deblending in the Euclid survey as a starting point for deblending in LSST overlapping areas can be considered

Euclid & LSST

• Euclid VIS ≤ 24.5 AB Magnitude (Rhodes et al. 2017)
• Euclid VIS PSF = 0.18’’ (Cropper et al. 2016)
• Euclid Survey Area 15000 !"#$ (Euclid site)
• LSST ≤ 27.5 AB Magnitude (Rhodes et al. 2017)
• LSST median seeing 0.7’’ (Rhodes et al. 2017)
• LSST Main Survey Area 18000 !"#$ (LSST site)

3

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

• Status: First version implemented in the Euclid MER pipeline, tested on MER-PHZ Data-

challenges and ran in Scientific Challenge 3 (SC3)

• (SExtractor) with parameters optimized after tests on Euclid-like images

SExtractor

SExtractor (Bertin and Arnouts 1996) Multi-threshold approach Two Main Deblending parameters: DEBLEND_MINCONT: Minimum contrast parameter for deblending (% flux for single branch) DEBLEND_NTHRESH: Number of deblending sub-threshold

4

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

ASTErIsM

5

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

ASTErIsM-DENCLUE

6

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Datasets: We create cubes of simulated images with related “true map” Our datasets are used to measure the deblending efficiency of different algorithms taking care of

• ver/under deblending and fraction of correctly

reconstructed objects Datasets are produced:

• with bulges or disks objects using Skymaker

(Sersic models) (E.Bertin 2009)

• with real CANDELS-GS cutouts

Datasets

N i m a g e s N t r u e m a p s

The objects have different: magnitudes, sizes, distances, morphology, etc. Examining the ambigous blends in HST catalog ∼75% of blendings are composed of just two

• bjects

(Dawson et al. 2016)

7

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Metrics

8

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Preliminary Results

Parameters map of deblending performance Reddish zones mean higher deblending performance 2-objects dataset single object dataset

ASTErIsM

9

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Preliminary Results

Reddish zones mean higher deblending performance Parameters map of deblending performance 2-objects dataset single object dataset

SExtractor

10

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

• riginal

True map asterism Asterism_overlap sextractor

• riginal

True map asterism

• riginal

True map asterism

• riginal

True map asterism Asterism_overlap sextractor Asterism_overlap sextractor Asterism_overlap sextractor

11

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018 Future developments:

• New datasets will be used for tests
• Comparison with other Deblenders
• Possible implementation of CNN (M.Huertas-Company, A.Boucaud)

Conclusion & Future developments

Conclusions: ASTErIsM and Sextractor have similar performances with the right set of parameters For ASTErIsM a set of parameters which reduces over/under deblending has been found For Sextractor it’s seems to be not so trivial

12

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Backup

13

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

ASTErIsM SExtractor

Backup

14

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Reconstruction/overlap indicators

• Sim_msk == True where (sim[image_ID] == sim_ID and sim[Image_ID] == overlap_id )
• Det_msk == True where (Det[image_ID] == det_ID
• Roi_msk == True where np.logical_and(sim_msk, det_msk) :

Reconstruction/overlap indicators

• Overlap_th : A threshold to remove sources detected out of the simulation area, and to asses that a det_ID
• verlaps sim_ID
• Recovery_sim_frac = roi_msk.sum() / sim_msk.sum() :
• Recovery_det_frac = roi_msk.sum() / det_msk.sum()
• Overlap_size = Roi_msk.sum()

Candidates

• Candidates = all detected objects which have a number of overlapped pixels > overlap_th are the, and
• verlapping a simulated source

15

LSST Project and Community Workshop 2018 • Tucson • August 13 - 17

#lsst2018

Association and contamination

• clean_candidates_list = List of candidates which are filtered by a second overlap_th
• Contaminated_list : list of all the candidates not associated to the sim_ID
• Clean_contaminated_list : list[from 0 to N_candidates or to N_candidates - 1] = list of all the candidates

which have a contamination_frac > contamination_th

Deblending quality indicators and thresholds

• Purity_sim_th = The minimum threshold defined on the recovery_sim_frac
• Purity_det_th = The minimum threshold defined on the recovery_det_frac
• Debl_ok = Number of assoc_id == N_sim and clean_contamination_list = [empty]
• Delta_debl = Number of assoc_ID + length(clean_contamination_list) - N_sim
• Debl = average(Debl_ok over all the N_imgs)
• Delta_debl_avg = average(Delta_debl over all the N_imgs)