The New SOSIEReduction/Analysis Pipeline

Christian MaroisNRC - HIA

AMNH, May 2010

(See also my SPIE 2010 paper on the same subject)Bruce MacintoshJean-Pierre Veran

1

What is ADI/SSDI imaging & LOCI processing?

I1

Same wavelength, same object, but at different time = ADI

Same object, simultaneous, but at a different wavelength = SSDI

Same wavelength, different time and different object = Archive

How to combine a bunch of reference images to subtract the “noise” of an image of interest

(while keeping the flux of what you are looking for)?

Sources of reference images

2

ADI

t1 t2 t2 t1-

SSDI

Archive

1 2 2 1-

O1 O2 O1 O2-

FOVRotationwith timeof Alt/Az

PSF chromaticity

3

I1

LOCI processing

• Select a section (purple) on all images.• Calculate cross-correlation matrix.• Inverse that matrix.• Obtain coefficients to multiply j, k, m, ... to minimize the noise in the purple section of i.• Repeat for many purple sections until all image i has been subtracted.• Repeat for all images of object of interest.• Combine all subtracted images of object of interest.• Find planets/disks/bg objects/...

Sequence of images

Trying to remove noise of image i using j, k, m, ... while keeping the planet flux

i kjm

4

ADI example

68 coadded

K-band, 10s exposure

~10x improvement per image, then ~sqrt(n).For 100 images, gain of the order of 50-100x (5 mag).

5

SPHERE

GPI

Keck K-band 1h ADI HR8799

6

ADI/SSDI & LOCI1. Easy to make a detection.

2. Harder to optimize & trust the photometry and astrometry (especially if your object transits a few degrees from Zenith...).

3. Quest for <10% abs photometry and mas astrometry is not yet a done deal... but we are getting there.

7

LOCI many biases

• Candidate misalignment: image-to-image, PSF smearing, distortion, rotation axis offset, North angle determination.

• LOCI non-uniform sensitivity (correlation matrix bias).

• LOCI box finite size.

• Non-symmetric dark shadows.

8

LOCI box finite size

r

dr!"#$%&'(

)*+,-*.

ADI “gap” variations in an optimization box.

For sections of annulus, the effect is constant, does not average with FOV rotation.

Requires very narrow optimization boxes/sections of annulus or a throughput calibration function

9

Non-Symmetric Dark Shadows

Potential photometry & astrometry bias.

Occurs in near Zenith transits and with non-symmetric obs. sequences (more data before/after transit).

10

Non-Uniform SensitivityHR8799 HST data

LOCI optimization annulus

Dark

Bright

Regions that are bright have more weight in the LOCI correlation matrix, resulting in more candidate self-subtraction in those areas (lower throughput).

11

SOSIE Architecture

!"#$%&'

()*+'*&',-*),%+

.*/,0

1"'20),%+

34*5"

1"5,/)&*),%+

6783

9&%0"//,+5

:%&);

<=,5+4"+)

*+'

34*5"

8%4>,+*),%+

? @ A B CNIRI@GN

NICI@GS

NIRC2@Keck

CONICA@VLT

GPI@GN/S

Smaller size for 1 time use. Lasts longer and cleans stronger. Take care of your tough messes. 15 pads to a box.

IE

Mainly developed for the ongoing IDPS survey.

12

SOSIE For each instrument, all reductions are saved in an archive for future

use as reference images (planet & disk).

Multi-wavelengths (SSDI) using “spectrum templates”, followed by an iterative LOCI analysis to recover the real spectrum.

SSDI, ADI and Archive PSF subtraction are performed in a single step.

A masking technique is used to avoid non-uniform sensitivity.

Astrometry & photometry are derived using PSF fitting with LOCI ADI+SSDI+Archive template PSFs (remove most of the biases discussed earlier).

LOCI parameter optimization using Simulated Annealing.13

SOSIE Local Masking Technique

!"##$%&'(")*+&'#(,*&#$&

-()&%*#$.$#$)/$*(+&0$

1&23

4#(+&#5

67!8*9#"/$22()0

14

No mask With mask

b bc c

11

2233

44Uni. noise,

but var. throughput

Better & uni.

throughput, but

var. noise

15

SOSIE SSDI+ADI+Archive

i

i+1

i+2

i-1

i-2

Ref. Objects:

Time (ADI)

Wavelength (SSDI)

I jI j-1 I j+1 I j+2I j-2

ADI gap

i

O1 O2 O3 O4 O5

SSDI criterion

For each LOCI box (after spatially scaling images to align speckles at all wavelengths)

...

......

...... Once a subset of images

has been selected, the best ones are kept for the final LOCI reduction (for speed & stability)

Both the ADI and SSDI criteria can be asymmetric

in time/wavelength.

16

Opt. SNR with SA Param. Optimization

!"##$%&'(")*+&'#(,*&#$&

-()&%*#$.$#$)/$*(+&0$

1&23

4#(+&#5

67!8*9#"/$22()0

abc

da: Central mask sizeb: Cor. matrix inner radiusc: Cor. matrix outer radiusd: Cor. matrix azimuthal radiuse: ADI motion gapf: SSDI flux subtraction criterion

Optimization is performed at several separations

17

With mask, all ref. images With mask, not all ref. images

No mask, all ref. images

b

b

bc

c

c

d?

d?

18

19

SOSIE PSF FittingOnce a candidate has been found, its photometry and astrometry can be derived by subtracting a template LOCI PSF obtained from unsaturated data (much better than Gaussian fitting or aperture photometry). The candidate can then be removed from the reference images to produce nice dark-shadow-free images.

Template LOCI PSF

20

Other Upgrades• PSF smearing deconvolution (marginal SNR gain).

• Final image combination: Weighted mean instead of median (marginal gain).

• LOCI Matrix inversion with SVD+cutoff (no gain vs “sequence truncation” - potential large gain vs simple LOCI implementation).

21

• L-/M-band LOCI sky subtraction

22

Future Upgrades

• Finishing implementing SSDI+ADI(~80% done)

• Finishing archive-based scheme (+Archive)(~90% done)

• Multi-processor & Optimization (currently 20 min for 500x500x70 images)

23