| Version 1.9 Build 1367
|
|
NOTE 219 - AIPS++ Mosaicing development plan
Tim Cornwell
Draft 1997 April 30
A postscript version of this note is available (100kB).
This document describes a phased plan to realize the Mosaicing
specifications as described by Holdaway et al.
(AIP++ Specification memo
120)
(hereafter HSCR). The concentration here
is on the grouping of work to be done. For any necessary clarification
of the definition of the work, see HSCR.
sky currently (April 1998) provides simple mosaicing
capabilities. Specifically, the program can correct for VLA and WSRT
primary beams in either a linear mosaic or via a clean-based
algorithm. In the clean-based algorithm, the minor cycles are
performed using an average PSF, and major cycles using the fully
correct PSF (including any time-variation). Both these capabilities
result from an initiative to demonstrate that the synthesis framework
is capable of supporting mosaicing data reduction.
A significant design decision in sky is that all deconvolution
is performed inside sky. This is contrary to the model used in
Miriad mosaicing, whereby the dirty image, PSFs and primary beams are
written out and processed separately in another program. The rationale
for the choice not to do this in sky is that the ancillary
information (arising from the MeasurementSet) that must be carried
along may be quite complicated (internally the VisibilityIterator
class is used to encapsulate this information).
Other points relevant to HSCR:
- The SkyEquation formalism naturally decomposes primary
beams into antenna-based terms, as desired in HSCR.
- Tracking of e.g. solar system objects is
available via the Measures system.
- The calibration object, cal, knows nothing about
mosaicing: all prediction of model visibilities is done by sky.
Visibility-plane calibration is thus independent of whether or not the
data are mosaiced. This means that the forms of visibility-plane
calibration envisaged by HSCR should be straightforward to
implement. The exception may be derivation of a relative
calibration between single dish and interferometric data.
- The one key capability yet to be developed fully and
demonstrated in the synthesis framework is the solution for image
plane-based calibration effects such as non-isoplanicity or pointing errors.
- Optimization of performance and disk space usage will
continue to require special attention. One strategy advocated by
Sault, Staveley-Smith and Brouw is to grid at a fixed wavelength (thus
implying frequency-scaled uv and sky position coordinate systems).
This would reduce the disk space needed for scratch images such as the
transfer functions at the cost of causing problems elsewhere with
a non-physical coordinate system.
- The advent of image regions will also help
performance since currently the entire model image is transformed
even if the primary beam covers only a fraction of the image.
To build on the current capabilities of sky, we need to do the
following:
- Rework the existing PBSkyJones class hierachy to allow the types of
PB models described in HSCR. In this phase, we will only allow PBs
with known form. Solveable primary beams will come later.
- Optimize the use of FFTs in the Gridded transform (class GridFT)
to avoid FFTing regions outside the primary beam model.
- Post-deconvolution primary beam correction, both single field
and multiple field.
- Provide different deconvolution methods:
- inverse and Wiener filtering (including production of
the effective uv coverage i.e. the FT of the average
PSF).
- SDI-CLEAN
- MEM using Cornwell-Evans algorithm (as modified by Sault)
- ComponentModel processing should be optimized and extended to
include a number of effects such as time and bandwidth smearing.
- Wide-field imaging and mosaicing should be decoupled.
A wide-field image will be constructed from a number of facets cleaned
simultaneously. The SkyEquation mechanism will automatically apply
the correct primary beam correction to each one. Sault, Staveley-Smith
and Brouw show that the correction to place different facets on the
same ultimate tangent plane can be performed by the appropriate
rotation of the uv data during the gridding and degridding steps.
This should be put into place in UVWMachine, and then sky
changed to make the appropriate facet images. The missing ingredient
is a simple facility to insert the facets into a final large image.
A suitable concluding milestone for this phase would be to able to
make mosaics from a very large spectral-line mosaic dataset from
WSRT. Another milestone would be to correct known assymetries in the
primary beam model e.g. the VLA primary beam squint when
observing circular polarization on the Sun.
Single Dish On-The-Fly data reduction can be accomplished using the
SkyEquation formalism but to acheive satisfactory performance when
imaging only OTF data, some changes may be required. In particular,
a different version of SkyEquation::gradientsChiSquared is probably
necessary since a Fourier transform is not required. In the
case of mixed single dish and interferometric data, the
current gridded transform machine should be adequate.
An important dependency is that MeasurementSet V2 is needed in order
to hold OTF data in a reasonably compact form.
A concluding milestone would be to construct an OTF image from
12m spectral data.
Once the changes above have been made, we can move on to develop
a simulation capability using the tools refined in the previous
work. Specifically, the SkyEquation can be used to provide
transforms of models, and the refined PBSkyJones can be used to
implement various primary beams in the mosaicing.
I suggest that the simulator be a separate object, with an interface
similar in spirit to that of sky i.e. tending towards
atomic, low-level operations rather than high-level operations.
VisJones classes for simulating additive noise and gain errors
now exist and are used in the current simulator. Some work is
needed on an interface for defining observing patterns
(such as mosaicing sequences).
Part of the simulation tools should be devoted towards examining
the three measures of accuracy of reconstruction described
by Cornwell, Holdaway and Uson (1992):
- Dynamic range
- : Peak on-source brightness divided by
rms off-source rumble.
- Fidelity
- : Median of on-source brightness divided by
error in reconstruction.
- Visibility SNR curve
- : Radial average of visibility
divided by rms error in reconstruction.
A concluding milestone would be to simulate a simple mosaiced
observation from one of the consortium telescopes, and make an image
using sky, and evaluate the imaging performance. As a check on
performance, the simulation should be done for separate time-variable
pointing errors per antenna.
There is a need for a number of basic tools for aiding mosaicing.
- Summary of a mosaic observation, both textual and graphical.
- Display pointing centers graphically. Possibly overlay with
e.g. spectra, or visibility curves.
- For any given point in an image, display which pointings (field-ids)
are relevant (and with what weight).
- Selection tools, textual and graphical.
- Based on field-id, etc.
- Based on proximity to a given point on the sky
- Based on data criteria e.g. peak visibility
- Weighting tools to control the weighting of various
parts of the mosaic.
- tools to show effect of weighting on sensitivity and
beam shape
- relative weighting of interferometric (i.e.
sensitivity-based or sidelobe-minimizing
- relative weighting of SD and interferometric data
Development of advanced mosaicing tools will require the
simulation capabilities and tools developed in phases III and
IV.
- On-the-Fly interferometric imaging. The computational load
seems to be immense but the formalism should handle this
adequately.
- Adaptive weighting of data during the deconvolution
process.
- Calibration determination between SD and interferometers
by, for example, regression in the overlap region of the
uv plane. Bob Sault has done work in this area that we should
emulate.
- Self-calibration for Primary beam parameters such as
scaling of the PB (e.g. probably adequate for WSRT UHF feeds?),
and pointing offsets. This is simpler for componentmodels and should
be done that way first. Note that if done correctly, non-isoplanatic
imaging should be a special case. By using bright sources at known
positions, one can determine antenna- and sky position-dependent gain
terms (phases in the non-isoplanatic case) on the bright objects, and
then take these phases into account when forming an image of the
residuals. In this way, a corrected image of the entire field
can be formed. The next step is to extend this to the emission
modelled by images (not just the componentmodels).
The phasing is not strict. Phase I and II can proceed
concurrently. Phases II, IV can be interchanged. Phase V obviously
must come last.
I'd estimate phases I-IV are several FTE-months of effort each.
Phase V is probably substantially longer, maybe 6 months.
Hence the entire effort is probably 18-24 FTE-months. Phase
II should be done by someone familiar with OTF data, phase
III requires a moderate but not sophisticated understanding
of interferometry, and the other phases require someone with a
deep understanding.
Please send questions or comments about AIPS++ to aips2-request@nrao.edu.
Copyright © 1995-2000 Associated Universities Inc.,
Washington, D.C.
Return to AIPS++ Home Page
2006-03-28