Image Subtraction & Transient

Detection Techniques
A.C. Becker (U. Washington), A. Rest (CTIO), G. Miknaitis (UW),
R.C. Smith (CTIO), C.Stubbs (Harvard), LSST Collaboration
The process of image subtraction drastically changes the characteristics of the signal and noise in the images being analyzed.
One or both images are resampled to a fiducial astrometric system; one image is additionally convolved to match PSFs; and
one image is subtracted from the other. Without careful propagation of image artifacts and noise, the output image may be
littered with false positives and have a poorly defined measure of significance. We outline improvements made to the
SuperMACHO/ESSENCE pipeline that have helped us improve the quality of our difference images. These improvements are a
significant step along the route to robust image subtraction algorithms for the LSST variability pipeline.

Masking : Bad pixels In difference imaging, the primary goal is the determination of the convolution kernel that
matches PSFs. This requires comparing pixel regions found in both images, which should optimally contain an object or set of objects at
high S/N. If there are bad pixels (or variable objects) in these regions, the process can fail. It is of utmost importance to know the quality of
the input images and to use only pristine pixels in this calculation. In addition, bad or saturated pixels leave a large footprint in resampled
images, and an even larger footprint in difference images where they have been convolved twice. We explicitly propagate the influence of bad
pixels through our pipeline, using a bit-wise image mask that tells us if a pixel was found in the detector's bad pixel mask, was saturated, or
received flux from any such pixel during the convolution steps. Our object detection algorithms have been modified to use these masks.
The difference image contains
Template Image – Raw Template Image –Resampled artifacts from both images. In
Input Image – Raw Input Image –Resampled addition, one set of artifacts are
Difference Image
convolved with the difference
imaging kernel. In this case, the
template was convolved before
subtraction from the input image.
Note the mask from the template is
spread, yielding 3 shades of grey.
The whitest regions represent
regions dominated by bad pixels
during the convolution. These are
The template image is resampled The input image is resampled to the
Template Mask masked out in the difference image Difference Image Mask
to an astrometric reference frame. same astrometric reference frame. Image Mask and in object detection. The middle
This process extends the footprint Due to pointing differences, the bad and outer regions correspond to
of bad pixels. In the mask image at pixel region lands on a different part destination pixels that received any
right, the white area represents the of the sky. Mask levels are the flux from masked pixels during the
region where bad pixels dominate same as in the template image. convolution (2 levels for the 2 levels
the convolution. These will be of masking in the input image). These
masked out in the difference image. are suspect but not masked out in
The grey area represents less the difference image, allowing us
contaminated pixels. These will not sensitivity to variability near bad
be masked in the difference image, pixels. We increase the noise in
but will be excluded from the difference imaging kernel fitting. Note the these pixels accordingly.
bad pixel region is interpolated over in the resampled image.

Masking : Diffraction spikes

Template Image Difference

Features around saturated stars lead to

significant numbers of residuals in difference images. To compensate, additional bits are added to our
mask to model haloes and diffraction spikes. The image on the right shows the resampled template and ESSENCE SN
2004-G-29
input images and their masks, as well as the difference image and its mask. Template Mask Image Mask Difference Mask

Noise propagation : The propagation of noise is important to quantify detection

limits in the difference images. The resampling and difference imaging stages both accept input noise
images, and produce output noise images. Object detection stages use this information to determine the
significance of detected objects. We currently ignore pixel covariance.
Result of resampling a FITS file whose pixels are all the same value. The process was done using the Swarp package, which uses WCS information
in each header for the transformation. The input image had TNX-format WCS distortions, and was flattened out to a TAN projection. All figures have Histogram of pixels in an input image containing purely
the same scaling. A horizontal slice through these images is seen to the right, which indicates that the Sinc resampling yields less systematic Gaussian noise, and in output images resampled with a Difference and associated noise images. The amplitude of the systematic noise
distortions (although at the expense of a larger footprint). The observed features also reflect the spatial distortions of the PSF introduced by the variety of algorithms. The input and output coordinate variations is ~5%. We enhance the noise around masked pixels to compensate for
resampling process, which must be minimized for the image subtraction stage. systems are the same as in the figures to the left. Note the their influence during convolution.
Lanczos4 resampling kernel distorts the image statistics
the least.
Bilinear Interpolation

Sinc Interpolation
Lanczos3

Sinc Interpolation
Lanczos4

Efficiency analysis : The determination of image detection efficiencies requires the addition of fake stars and an
inventory of fraction recovered as a function of brightness. To enable this for our image subtraction pipeline, we save the convolution
coefficients in the FITS header of the difference image. In this way, we can use the same kernel that determined the original difference image
to yield the efficiency image – otherwise, the efficiency stars may dominate the convolution solution.
Input Image Difference Image Input Image + Fake Stars Difference Image + Fake Stars