Reports are given in reverse chronological order
A plot of the percentage of deblends vs. average aspect ratio (at Ks band) summarizes what deblending has done for sci and cal scans in the 1997 data. Values here are derived from nights run through the first round of v3 OPS testing.
Cal blocks and sci blocks are plotted as blue points (for 9706xxn and 9707xxn data) or green points (for 971216n) at the average Ks aspect ratio for the block. The percentage of deblends was computed only for sources have Ks<13.0 mag and was calculated by dividing the total number of sources into those having the BlT flag like '%2%'. To avoid confused regions, sci and cal scans were used in the computation only if log(dens)<3.5.
Cal blocks of the same field are connected by lines of the same color.
The first thing to note here is that the green points (from 971216n) generally fall lower on the plot than those from the earlier data. For reasonably round images (near I2min/I2maj = 0.90), the blue points bounce between deblend fractions of 10% and 30%, which seems very high from an astrophysical viewpoint. For this same aspect ratio, the green points have reached a floor of around 7%.
The second thing to note is that repeats of the same cal fields can have strikingly different deblend percentages for similar aspect ratios. Field 90533 (magenta line), for example, has deblend percentages between 5% and 22% for very similar aspect ratios.
In a nutshell, deblend percentages vary in ways that are not tied to the astrophysics but instead to aspects of the data itself (poor focus, for example). Deblending such data is likely doing more harm than good, so my recommendation is that deblending should be turned off for v3 OPS processing.
My assertion yesterday that the deblend fraction increases with degrading seeing is not true. The correlation I saw yesterday was on a limited data set, and I was fooled by one (apparently anomalously high) point.
This plot gives the Ks<13.0 active deblend percentage for cal field 90533, used repeatedly in the OPS test nights. Except for one high point (near 20%), the deblend fraction is roughly constant for aspect ratios between 0.97 and 0.85 and for Ks seeing shapes between 0.92 and 1.15.
Sorry for the false alarm.
After working all day on deblends in the calibration standards list and looking back at the data from my deblend analysis last week, I've discovered the following. There is a correlation between seeing shape and the percentage of deblended sources. However, it doesn't go in the sense I'd expect. Naively, I'd think as seeing improved you might be able to more easily split close binaries and the percentage of deblends would increase. Not so. The observed effect is that the percentage of deblending increases as the seeing *degrades*.
Last week I noticed that cal field 90533 on 971216n had deblend percentages (for Ks<13.0) anywhere from 5% to 22% among its four sets of six-time repeats on that night. Well, the 22% occurs in the set with the worst seeing (but not the worst elongation), and I'm seeing the same correlation with secondary standard deblends across nights.
Can anyone explain this?
Starting with some of the very early northern data (970611n and 970614n), I've looked at cal sets whose scans have log(density)<3.5 and computed the mean I2min/I2maj ratios in the set. Then I looked at all the sources in that set with Ks-band PSF mag<13.0 to figure out how many of those were successfully actively deblended in any band (i.e., the BlT flag has some digit that's a "2"). The results are shown as blue points in this figure.
Note that the distribution of points is rather flat until you hit the point at which we (in ad hoc fashion, as I recall) in v2 downgraded sci scans to quality=1 due to bad image elongations. This number is I2min/I2maj<0.81 as shown by the red line. At face value, this would indicate that we'd like to turn deblending off near the same point before it goes wild, corresponding to the ramp up in the blue points seen in the right half of the plot.
However, it seems to me that normal, low galactic latitude data such as these shouldn't have 15-30% of their bright stars as close doubles. To check this, I took a night later in the year (971216n) when the images were much improved overall. Those points are shown in green on the same figure. Note that the percentage of deblends is considerably lower on average than for the early 9706xx nights.
I'm thinking that the different floors seen in the percentage of deblends at good aspect ratios may be correlated with different collimation periods. Perhaps that floor rises in data with poorer collimation because the elongated PSFs at the array edges are more susceptible to deblending.
I should also note that on Laurent's page, there's a plot that also shows the percentage of active deblends as a function of source density. For scans in the density range I've considered above, he found active deblend fractions of around 3% (any band deblended). This fraction is much smaller than the one I'm finding. One difference is that he looked at sources with Ks<14 as opposed to my Ks<13 cut above, and for badly elongated images the chi-squared values (see e.g. 970611n) jump markedly upward between Ks=14 and Ks=13. Another difference is that he looked at an ensemble of nights for his plot, including later northern nights as well as southern nights which have better images overall anyway. So, I'm not surprised that the fraction of deblends I'm finding for 971216n (~7%) is somewhat higher than what he found. If anyone else disagrees with this, please let me know.
