Automated Bright Star Masking

Bright stars create havoc within coadd images and UNLESS handled aggressively they are the primary source of contaminents to the extended source database. They are the progenitors of "halos" (low surface brightness emission extending well beyond the PSF), diffraction spikes, horizontal stripes, glints and persistence residual ghosts -- a real horror show. For further sunny information on these deamons, see GALWORKS Bright Star Cleansing .

After months of 3-channel data reduction and analysis (not to mention a year or two of work with protocam data), I have come to the conclusion that using the R1 mags to generate 'blanking parameter values' (e.g., confusion radii, diffraction spike lengths, etc) via lookup tables or mathematical functions is NOT dependible (re: robust) under a wide variety of conditions -- stellar number density and sky background. The reasons are many (one of which is that the R1 mags are not reliable, particularly near saturation) of which I will not discuss here. Instead, I maintain that lookup table values suffice as "first" guesses to the ultimate final solution we desire. The masking parameters (confusion radii, diffraction spike lengths, horizontal stripe(s) and persistence ghost blanking) are found by measuring and scoring the corresponding features for each object (per band). We take a cpu hit since we require a fair number of complicated computations, but we gain in the end with uniform performance regardless of the quality of the R1 mags or other input (first guess) parametric values. This memo describes the algorithm and performance on a number of fields, including high glat fields, low glat (3 to 7 degrees above the plane) and a field directly in the plane (our old friend, MSX, taken way back in April).

Algorithm

It would be fair say that the set of algorithms can be described as the "bright star processor' given the relative complexity compared to lookup tables or characteristic functions. In brief:

Input: J,H,K read-1 mags & positions, various confusion flags (from MAPCOR (?) )

Image parameters: global sky background and histogram values for select points on the distribution curve (to derive the sky noise).

Step 1: derive sky noise (per band)


The "noise" of the coadd image can be characterized by the 84% and 16% quartiles of the distribution histogram. This mimics a gaussian distribution:
noise = (qt84-qt16)/2.0
where qt84 and qt16 are the 84% and16% quartiles, respectively.

Step 2: Derive Confusion Radii

'predict' confusion radii via fit to lookup tables;
Modify first-guess confusion radii by the "confusion noise" factor;
using modified first-guess confusion radii, derive boundary limits to be used in the fits described below. The confusion noise factor is the ratio of the sky noise in the image and the nominal sky noise attributed to low stellar density images (current values: 0.6, 1.1, 1.15 dn, for JHK, respectively).
confusion noise factor = noise_0 (band) / noise (band)
where noise_0 is the low density sky noise (confusion noise) and noise is the sky noise in the image containing the bright star.

determine confusion radii:


Compute median values in annuli extending away from the bright star, convergence occurs at the point where the 'SNR' of the annular surface brightness falls below some threshold (default = 1-sigma); the SNR is computed by normalizing the median value in the annulus by the confusion noise (i.e., the "noise"). The radii are then used to mask pixels comprising the central bright source and its 'halo'.

Step 3: Horizontal Stripe Terminator


Bright stars exhibit 3 horizontal stripes, one centered on the star (primary), one 256 coadd pixels below the star (secondary) and one 256 above the star (secondary), extending fully across the image. The primary stripe is about a factor of two brighter than the secondary stripes (both of which have about the same surface brightness). This step is designed to measure the SNR of the stripes and 'score' them accordingly. We then blank the stripes if the score is above some threshold limit.

'predict' stripe SNRs via fit to lookup tables;
modify first-guess stripe SNRs by the "confusion noise" factor;
using modified first-guess SNRs, decide whether further computation is required (default threshold: 1-sigma).

determine SNR of primary and secondary stripes:

for primary stripe, compute median value in rectangular region comprising the stripe -- 512 pixels in cross scan, and +-4 pixels inscan, centered on y-position of bright star. The central pixels surrounding the star, comprising the star + halo, are excluded from the computation (since they add nothing to the stripe).

To measure the relative (or local) 'background' value, a similar computation is made on an adjacent stripe region. The SNR is then computed by normalizing the median value (minus the background value) by the confusion noise. The SNR is then compared to some threshold. The threshold is 0.4 dn, modified by the confusion noise factor as follows:

threshold = 0.4 / sqrt(confusion_noise_factor)
Thus, for more confused regions (i.e., near or in the galactic plane) the threshold is raised accordingly (so we are less likely to pass the threshold). If the SNR is greater than the threshold, the primary stripe is masked from the coadd.

For the secondary stripes, the procedure is the same as that for the primary stripe, except that we use both secondary stripes in the median value determination (since they have roughly the same surface brightness we improve our statistics by root 2).

Step 4: Spike Hammer

Bright stars generate vertical/horizontal diffraction spikes due to the secondary support structure of the telescope. The horizontal spike is part of the primary horizontal stripe for the brightest stars (K < 4), thus it is automatically masked with he primary stripe masking. For moderately bright stars, (K > 5), the horizontal stripe is not prominent (but the spike is) and is typically not masked. Thus, we treat the horizontal spike according to whether the horizontal stripe has been masked or not. We measure the SNR of the verti/hori spikes (added symmetrically) as a function of radius from the central source, and compare this with some threshold limit. We blank the spikes (symetrically) at the point where the SNR falls below the threshold.

'predict' the spike lengths via fit to lookup tables;
modify first-guess lengths by the "confusion noise" factor;
using modified first-guess values, derive boundary limits to be used in the fits described below.

determine diffractin spike lengths:


compute median values in small rectangular wedges (3X5 pixels) centered on the spikes; the vertical spikes are added symmetrically and the horizontal spikes are also added symmetrically ONLY IF the primary horizontal stripe is NOT masked (see previous step). The snr is computed by normalizing the median value by the noise. The procedure continues at sequentially larger radii from the central source until the SNR falls below some threshold. The spikes are masked from the image at this radius.

The threshold is 0.75 dn, modified by the confusion noise factor as follows:

threshold = 0.75 / sqrt(confusion_noise_factor)
Thus, for more confused regions (i.e., near or in the galactic plane) the threshold is raised accordingly.

By actually measuring the halo, stripe and spike features of bright stars, we can reliably judge if the regions comprising the features require masking and thus avoid the often misleading information that look-up tables and characteristic functions give (the axium, GIGO, applies to this process).

Persistence ghosts have not been addressed thus far in this discussion. Given that these objects are (or should be) well characterized early in the pipeline (well before galworks), it is possible to eliminate them by simply looking for the appropriate flag generated by (?MAPCOR?). If a detection is a highly probable persistence ghost, then GALWORKS will mask an appropriate region comprising the ghost. This operation will occur before bright star processing.

Look-Up Tables

See the following link for a discussion of generating the look-up tables. Bright Star Parameter Tuning .

Additional fields have been added to the databank since the previous memo was whipped up, including MSX -- the galactic plane. Soon, I will add the scan containing Maffei 1. The latest greatest plot of the confusion radii for bright stars is given below, as well as he spike lengths plot and the horizontal stripe plots.

The confusion radii, spike lengths and horizontal stripe ratings are summarized in the following plots. The horizontal stripe analysis is summarized by a "rating". By this we mean the number of stars (per mag bin) in which a horizontal stripe is observed, normalized by the total number of sources. Thus, for the brightest stars, these stripes are always observed, so they have a appearance rating of 1.0. For the faintest stars, these stripes begin to dissappear and the rating approaches zero. We also compute the SNR of the stripe itself -- this is the parameter used in the algorithm described above.

Note the large scatter in the plots. This is due to the uncertainty in the R1 mag (due to saturation, clipped images and other effects) and the variable confusion noise (and changing background level). It is this scatter that renders look-up tables unreliable as the only source for masking information.

• Confusion Radius
  white points == low source density
  purple crosses == moderate to high source density (glat=3-7)
  yellow crosses == very high source density (msx)
  

• Diffraction Spike Length

• Horizontal Stripe SNR & Rating
  This plot shows both the measured SNR for primary (white filled points) and secondary (blue triangles) stripes. A quadratic has been fit to the points. Shown in green (dashed lines) are the horizontal stripe "rating" determined independently of the SNR business. A high rating means that the stripe should be blanked, a low rating means it should not be blanked. Notice that the SNR and the rating are correlated. It appears that an SNR limit of 2-sigma corresponds to the point where the rating falls off sharply. Thus a limit of 1 or 2-sigma should be applied to the data (stripes).

Performance

Several bright stars were examined in detail with the new automated bright star masking algorithm. The stars are scattered widely across the sky, including the high glat fields of of Hercules, Coma, Hercules supercluster and a couple other anonymous fields, and low glat fields including scans located between 3 and 7 degrees glat, and finally one MSX scan located around 1 degree glat. Thus, we have a wide range in confusion noise and sky background variation -- the two most troublesome components to robust bright star masking. We will focus on a couple test cases from a high glat field to the low glat fields, and include a gaggle of additional gif images showing other bright stars.

Note that some of the R1 mags appear to be far from realistic -- this demonstrates one of the problems with the photometry.

High GLAT -- confusion noise minimal

• JHK images of bright star; J=4.0 , H=3.8, K= 4.5
• JHK images showing masked regions
  blue == confusion radius
  red == diffraction spike lengths
  green == primary/secondary horizontal stripes
  

  This star is clearly saturated and thus the R1 mags are probably not very reliable. Given that, the predicted confusion radii are 79.1, 83.2, 55.6, for JHK respectively -- much MUCH to small. The surface brightness of this very bright star does not fall below the 1-sigma threshold for 160, 150 and 103 radii, respectively. The JHK images showing the confusion radii (in blue, see above gif) clearly are a much better boundary for masking.

  The predicted stripe SNR for the primary/secondary horizontal stripes are well above the threshold limit. The measured SNR are in accordance with the prediction.

  The predicted spike lengths are 115.3, 124.3, 89.5, respectively. Again, these are much too small due to poor R1 mags (saturation). The actual spike lengths measured are 159.5, 159.6 and 177.8 pixels, respectively.
  

  
  

• JHK images of bright star; J=3.7 , H=4.0, K=4.4
• JHK images showing masked regions

  
  

• JHK images of bright star; J=3.3 , H=3.5, K=4.6
• JHK images showing masked regions

  
  

• JHK images of bright star; J=5.4 , H=4.4 , K=4.5
• JHK images showing masked regions

  
  

• JHK images of bright star; J=3.5 , H=4.0 , K=3.8
• JHK images showing masked regions

  
  

• JHK images of bright star; J=3.6 , H=6.7 , K=6.3
• JHK images showing masked regions

  
  

• JHK images of bright star; J=6.4 , H=5.4 , K=5.1
• JHK images showing masked regions

  
  

• JHK images of bright star; J=5.0 , H=6.9 , K=6.6
• JHK images showing masked regions

  
  

• JHK images of bright star; J=7.7 , H=7.0 , K=7.0
• JHK images showing masked regions

  
  

• JHK images of bright star; J=8.8 , H=8.5, K=8.4
• JHK images showing masked regions

• JHK images of bright star; J=1.7 , H=6.9, K=6.3
• JHK images showing masked regions
  blue == confusion radius
  red == diffraction spike lengths
  green == primary/secondary horizontal stripes
  

  The source density is very high in this field. The confusion noise quantifies what the eye sees: 1.0, 1.8 and 1.5 dn, JHK respectively. Compare this with the nominal high glat noise values of 0.6, 1.1 and 1.2, respectively. The confusion noise factor is then 0.59, 0.60 and 0.87, respectively. This factor will be used to modify the thresholds for spike length convergence and stripe score.

  Predicted confusion radii: 126.7, 32.4, 31.4, respectively. The J value is clearly out of whack -- due to a poor R1 mag. The actual confusion radii measured turns out to be 67.5, 54.6, 41.6, respectively.

  Only the J and K primary stripes have measured SNR greater than the modified thresholds (H just barely fails the thresh). For the secondary stripes, only J has a large enough SNR for blanking.

  The predicted spike lengths are 173, 48 and 51, respectively. The J value is clearly bogus. The actual spike lengths are 82, 65 and 62, respectively.

  
  

• JHK images of bright star; J=3.1 , H=7.1, K=6.5
• JHK images showing masked regions

  
  

• JHK images of bright star; J=3.4 , H=4.2, K=4.4
• JHK images showing masked regions

  
  

• JHK images of bright star; J=3.5 , H=3.9, K=4.4
• JHK images showing masked regions

  
  

• JHK images of bright star; J=3.6 , H=4.0, K=4.4
• JHK images showing masked regions

  
  

• JHK images of bright star; J=5.9 , H=4.2, K=4.6
• JHK images showing masked regions

  
  

• JHK images of bright star; J=6.9 , H=6.2 , K= 5.7
• JHK images showing masked regions

  
  

• JHK images of bright star; J=7.6 , H=6.9, K=6.8
• JHK images showing masked regions

  
  

• JHK images of bright star; J=7.9 , H=7.0, K=6.8
• JHK images showing masked regions