Recall that after the noise level has been estimated, we add a constant value, offset (for input to the module, but recorded as C_NOISE in the keywords), to determine which pixels will be annealed. Choosing this value too low results in many noisy pixels being annealed; choosing this value too high results in leaving out pixels that have a good signal.
A mask is constructed consisting of those pixels with transverse field strength above the value of the noise plus the offset. This mask is then eroded to remove isolated pixels above the noise, then grown, to provide a buffer around the pixels with well measured field. All pixels within the buffer are annealed, but the annealed solution is only returned for those above the noise threshold; a nearest neighbor acute angle (smoothed) solution is returned for pixels in the buffer area.
In the following plots, + represent the number of pixels in the buffered area, ◊ represent the number of pixels where the annealed solution is returned.
Consider first the variation through half a day (or through roughly the extent of the orbital velocity variations) for 2012.01.01. Black is offset=20, dark blue is offset=40, red is offset=60, light green is offset=80, light blue is offset=100. There are large variations over a few hours in the number of pixels being annealed when offset=20, and the variations decrease in amplitude with increasing values of offset.
To understand this, consider the distribution of the pixels being annealed at two times, 6 hours apart, for two values of offset: 20 and 60. For offset=20 (top row), there are clear residual patterns of the noise, which change between the two times. For offset=60, these residual patterns are no longer clear. Recommendation: take offset≥50 to avoid the residual patterns.
To estimate the run time for the disambiguation, we also need to consider how the number of pixels being annealed varies over real changes to what's on the disk. The following plot shows the number of pixels as a function of offset for 00:00 TAI on the first day of the month for roughly a year, with each color corresponding to a different date. I'm still working on characterizing this distribution, but the number can vary by at least of order a factor of 2.
