In that earlier post, I described a method of specifying a zero boundary condition in which the regions of the maps with low hits (the edges) are set to zero after each iteration. While this helped significantly with smaller maps (particularly pointing observations), it wasn't much help for significantly larger maps (i.e., covering much more area than the array footprint). This would lead to "saddle" shapes all over the map as the number of iterations was increased. We have therefore added two new masking options.
The first is to define a circular masked region:
ast.zero_circle = (LON,LAT,RADIUS)
or
ast.zero_circle = (RADIUS)
In this case, the portion of the map beyond a circular region centered over LON,LAT with the given radius (all quantities in degrees) is constrained to zero. Alternatively, only a single value may be supplied, in which case it is interpreted as the radius, and the LON,LAT are taken to be the pointing centre of the map (usually the compact source of interest for which this option is generally useful).
The second method automatically determines the masked region based on a SNR threshold.
ast.zero_snr = SNR
In this case, after each iteration of the map, all map pixels above the threshold SNR are included in the "source mask", and map pixels outside this mask are constrained to zero. As bright structures "grow" with respect to the neighbouring zero-constraiend map pixels, each iterations of the source mask will also grow to include more pixels.
For both of these options, the ast.zero_notlast flag applies: for the final iteration the zero constraint will not be applied if this is set.
So from a numerical point of view, what does this constraint do in the regions that were constrained to zero until the last iteration? In its simplest form the map-maker is trying to solve for several signal components in the bolometer time series:
Signal = ASTronomical + COMmon + FiLTered
where AST is the signal produced by astronomical sources, COM is the average signal seen by all of the bolometers, and FLT is generally residual low-frequency noise. By setting AST to zero at any location in the map, the only other places that signal can go are COM and FLT. COM only contains the average signal seen by all of the detectors, and FLT only contains low-frequency signal components. Therefore this operation, perhaps surprisingly, will not remove any compact structures in the zero-padded regions; the effect seems to be roughly equivalent to setting the mean value on array-footprint-sized patches within those regions to zero.
The updated dimmconfig_bright_compact.lis is a good default configuration for small/bright sources; in this case a circular mask is used with a radius of 60 arcsec. Similarly, there is an updated dimmconfig_bright_extended.lis that masks based on a SNR threshold of 5-sigma. We illustrate the effect of the SNR masking using the short observation of Orion mentioned in the previous blog post, s4a20100219_00039.
First, a map produced without any zero constraints and 20 iterations, clearly showing the large-scale degeneracy that results in a strong gradient in the map:
Here is the same reduction, but setting ast.zero_snr=5 and ast.zero_notlast=1:
and the mask indicating which pixels are constrained to zero until the final iteration (white) -- this is the QUALITY component of the final image:
Note that 20 iterations was chosen in this example to illustrate the improved convergence properties when using the zero mask. In fact, only about 8 iterations are needed in this particular case. A better automatic convergence test is still an outstanding issue for SMURF, and at present we recommend checking the maps after each iteration visually.
No comments:
Post a Comment