Source detection at Saclay

1
Source detection at Saclay
Jean Ballet and Régis Terrier, CEA Saclay
LAT consortium, 29/09/04

• Look for a fast method to find sources over the
  whole sky
• Provide list of positions, allowing to run
  maximum likelihood locally

• Tests on DC1 data
• 6 days data set
• Work in Galactic coordinates
• 4 energy bands (32 MeV / 100 MeV / 316 MeV / 1
  GeV / 10 GeV)
• Pixel adapted to each band (0.5 / 0.3 / 0.2 /
  0.1)
• Cartesian projection around the Galactic plane
• Polar projection (r 90-b or 90b, ?l) around
  the poles

2
Source detection using wavelets. Description

• Iterative algorithm
• Select relevant scales
• Wavelet Transform
• Threshold for each scale
• Detect relevant strucure to compute
  multiresolution support M
• Reconstruct solution S
• Compute residuals
• Wavelet Transform on residuals
• Detect structures belonging to M
• Reconstruct and update solution S
• Iterate until convergence
• Can be applied to continuous WT (reconstruction
  via wavelet packets)
• dyadic WT (a-trou algorithm)
• Tests using MR1 software package (developped by
  J.L. Starck).
• Actual source detection on the smoothed image
  with SExtractor.

3
Source detection using wavelets. Maps

• DC1 sky with mr_filter using iterative filter,
  Poisson noise, 4 sigma threshold
• 100 MeV 1 GeV keep scales 0.8 and 1.6. 135
  candidate sources
• 1 GeV 10 GeV keep scales 0.4 and 0.8. 210
  candidate sources

4
Source detection using wavelets. Results
Sources found using the mr_filter (yellow
crosses) compared with the 3rd EGRET catalogue
(green circles). The image is the sky at 100 µm
(to show the Galaxy). 0.1 10 GeV 108 correct
identifications, 27 spurious 1 10 GeV 168
correct identifications, 48 spurious, one
duplicate Together 205 correct identifications,
75 spurious
5
Source detection using wavelets. Summary
Strong points

• Already existing maintained package, immediately
  available, fast
• Good detecting power
• Method already used in other contexts (for
  example, XMM large scale survey of M. Pierre et
  al.)
• Can detect extended sources as well (if any)

Weak points

• Finds many spurious sources. Because the
  detection bears on wavelet coefficients (not on
  sources directly), raising the threshold does not
  give very good results. Maximum likelihood is
  necessary to weed out the false detections.

In progress

• Wavelet filtering in 3-D (X, Y, E) currently
  being developed as a general tool. Not clear this
  will be available soon enough for GLAST.

6
Source detection using optimal filter. Description
Idea Determine optimal filter using (known)
power density spectrum of the background
(Galactic diffuse emission) and Point Spread
Function. Generalisation of the matched filter
technique (Vio et al., AA 391, 789). PSF
averaged over off-axis angle and energy. Source
detection in Poisson regime compute probability
that local photon distribution follows background
shape (like wavelet transforms do).
Example 316 MeV to 1 GeV band
7
Source detection using optimal filter. Results

• Threshold such that probability times number of
  resolution elements is 1
• 0.032 0.1 GeV 29 correct sources, 6 spurious,
  2 duplicates
• 0.1 0.316 GeV 52 correct sources, 6 spurious
• 0.316 1 GeV 86 correct sources, 5 spurious
• 1 10 GeV 108 correct sources, 7 spurious, 4
  duplicates
• Together 166 correct sources, 24 spurious

Below Raw map sources (Galactic plane).
Above Filtered map (truncated at 10 sigma)
8
Source detection using optimal filter. Summary
Strong points

• Simple method, handy to experiment on pixel size,
  sky projections,
• Reasonably fast (1 hour on my laptop for the
  whole sky with 0.1 pixels)
• Reasonably powerful
• Gives direct source significance

Open issues

• Optimal filter varies a lot with energy. Split
  into even more energy bands ? Need to combine
  likelihood images.
• Not the same structure in latitude (sharper) and
  longitude near the Galactic plane. Use different
  filter in both directions ?
• Optimal filter depends on amplitude of background
  structures (balance with Poisson noise). Not the
  same in plane and at poles. Use smaller latitude
  intervals ?
• Galactic power density spectrum must be
  extrapolated to shorter wavelengths (currently
  masked by Poisson noise)

9
Source detection at Saclay
Jean Ballet and Régis Terrier, CEA Saclay
LAT consortium, 29/09/04
Methods exist work quite fast give a
reasonable list of sources but there is still a
long way to go.
Several open issues

• Would like to reduce pixel size to 0.05 above 1
  GeV. Becomes RAM hungry
• Get significances of sources after wavelet
  detection ? Much easier for setting threshold,
  also useful for setting position error.
• Adding likelihoods in several energy bands at map
  level will require storing even more data in
  parallel, and could become CPU consuming. Need
  reasonable approximations to avoid full Poisson
  probability computation at all pixels.