Blind Detection of Ultrafaint Streaks with a Maximum Likelihood … · Real GEO Data: PSF convolved...

Post on 18-Jan-2019

215 views 0 download

Transcript of Blind Detection of Ultrafaint Streaks with a Maximum Likelihood … · Real GEO Data: PSF convolved...

LLNL-PRES-692194 This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE-AC52-07NA27344. Lawrence Livermore National Security, LLC

Blind Detection of Ultrafaint Streaks with a Maximum Likelihood MethodCASIS

Dawson (LLNL), Schneider (LLNL), and Kamath (LLNL)

18 May 2016

1

Real Data GEO

Real GEO Data: PSF convolved detection seems to work very well.

• Consistently detects streaks and stars

• Star/streak separation based off multivariate Gaussian fit

Three streaks.

Real Data LEO

PSF convolved detection worked in some cases

Failed on some LEO data

• Shorter exposures = lower SNR = detection “issues”

• Many fainter streaks that could be seen by eye were not detected

Failed to detect this streak.

Maximum Likelihood Detection

lnL (x, ✓) = � 12 [d(x)�m(x, ✓)]T ⌃�1 [d(x)�m(x, ✓)] + C

m(x, ✓) =

Zs(x, ✓)�(x)dx

m(x, ✓) = ↵µ(x,⇥)

�(x,⇥) = µ(x,⇥)T⌃�1µ(x,⇥)

(x,⇥) = d(x)T⌃�1µ(x,⇥)

Work with Jim Bosch and Michael Schneider

@lnL@↵

= ↵�(x,⇥)� (x,⇥) = 0 ↵ML = (x,⇥)

�(x,⇥)lnLML =

2(x,⇥)

2�(x,⇥)⌫(x,⇥) =

(x,⇥)p�(x,⇥)

source model convolved with PSF

Log likelihood

separate the flux from the other model parameters

noise weighted model auto-correlation

noise weighted data-model cross-correlation

Significance

Faint Source Pipeline: Tests on Simulated Data

Testing detection pipeline on faint simulated streakStreak just to left of black line.• Create simulated image with 4 high SNR stars

and one faint streak.

• Convolve image with initial PSF guess. This creates image

• Estimate (currently constant across image), and combine with for

• Detect sources in image above a SNR threshold (footprints)

• Star/streak separation, estimate PSF, repeat above with more accurate PSF.

• Mask detected footprints in image

• Convolve image with streak model (i.e. line), and repeat detection process. ( )

�PSF

⌫PSF

PSF

PSF

PSF

PSF ⌫PSF

⌫streaki

Testing detection pipeline on faint simulated streak• Create simulated image with 4 high SNR stars

and one faint streak.

• Convolve image with initial PSF guess. This creates image

• Estimate (currently constant across image), and combine with for

• Detect sources in image above a SNR threshold (footprints)

• Star/streak separation, estimate PSF, repeat above with more accurate PSF.

• Mask detected footprints in image

• Convolve image with streak model (i.e. line), and repeat detection process. ( )

�PSF

⌫PSF

PSF

PSF

PSF

PSF ⌫PSF

⌫streaki

Can just start to make out streak.

Testing detection pipeline on faint simulated streak• Create simulated image with 4 high SNR stars

and one faint streak.

• Convolve image with initial PSF guess. This creates image

• Estimate (currently constant across image), and combine with for

• Detect sources in image above a SNR threshold (footprints)

• Star/streak separation, estimate PSF, repeat above with more accurate PSF.

• Mask detected footprints in image

• Convolve image with streak model (i.e. line), and repeat detection process. ( )

�PSF

⌫PSF

PSF

PSF

PSF

PSF ⌫PSF

⌫streaki

Point sources detected at high significance but streak below

threshold, SRN < 5.

Sign

al to

Noi

se R

atio

Testing detection pipeline on faint simulated streak• Create simulated image with 4 high SNR stars

and one faint streak.

• Convolve image with initial PSF guess. This creates image

• Estimate (currently constant across image), and combine with for

• Detect sources in image above a SNR threshold (footprints)

• Star/streak separation, estimate PSF, repeat above with more accurate PSF.

• Mask detected footprints in image

• Convolve image with streak model (i.e. line), and repeat detection process. ( )

�PSF

⌫PSF

PSF

PSF

PSF

PSF ⌫PSF

⌫streaki

Testing detection pipeline on faint simulated streak• Create simulated image with 4 high SNR stars

and one faint streak.

• Convolve image with initial PSF guess. This creates image

• Estimate (currently constant across image), and combine with for

• Detect sources in image above a SNR threshold (footprints)

• Star/streak separation, estimate PSF, repeat above with more accurate PSF.

• Mask detected footprints in image

• Convolve image with streak model (i.e. line), and repeat detection process. ( )

�PSF

⌫PSF

PSF

PSF

PSF

PSF ⌫PSF

⌫streaki

Iteratively determine the Maximum Likelihood Model (major room for computational optimization but fundamental method solid)

Streak now detected at high significance!

Faint Source Pipeline: Tests on Real Data

After worked on real data with faint streaks we ran it on an image with no apparent streak.

Nothing in the PSF convolved image.⌫PSF No “footprints” with SNR > 5

After running iterative maximum likelihood search… found a very faint but significant streak!⌫PSF

Detected FootprintSignificant Streak