Background

The IPP automatic processing of PS1/GPC1 data is currently blocked by excessive false positive rate in Magic for short exposures. The assessment by Paul Sydney as of late May was that Magic is able to detect satellite streaks, but that the effort has reached the limit of its ability to distinguish real streaks from other contaminating sources.

At the time, the top candidate for the contaminating source was the correlated read noise seen especially in certain chips. This signal has the appearance of positive or negative trend in the bias for each row, with a small amount of correlation between neighboring rows. The cause of this signal has been extensively investigated by the camera group, and is apparently due to the interaction between the video signal of multiple chips readout at the same time on the same flex-print circuilt. Without going too much into the underlying electronic issues, the signal consists of a row-by-row offset in the effective bias of the row, along with a possible slow drift of that bias. The camera software currently subtracts a row-by-row constant bias to make a first-order correction for this effect. However, the (unpredictable) slope results in a slow drift for each row from this corrected value. As a result, the cells & chips which suffer most from this effect have a generally good appearance on the end of the row, trending to significal offsets for the beginning of the row. The amplitude of the effect (the 'streakies') varies from chip to chip and from cell to cell within a chip. Although the row-by-row amplitude varies and is unpredicatble, the range of tilts for a given cell appears to stay consistent over many reads and many nights.

The camera group is studying possible mitigations in the hardware, including use of the pedestal signal and/or measuring a pre-scan bias level as well as a post-scan bias level. Both mitigataions will require research in the lab with the test setup, though it is reasonable to expect that a solution can be implemented without removing the camera from the telescope -- the implementation would be completely in software. However, because of the research needed (and finite available labor by the camera group) no such implementation can be expected for 2-4 months. In addition, any hardware mitigation will only allow for future improvements; none of the data taken would be addressed by this work. To this end, we have been working in two veins to implement a solution in the processing, even if such a solution would be less that the perfect solution.

The two processing fixes we have been pursuing attempt in one case to correct (remove) the signal and in the other case to incorporated the effect into a more complete description of the noise model.

In the first case, John Tonry has implemented a program called 'tiltystreak' which attempts to subtract the slope from each row, while at the same time attempting to avoid contamination from real signals in the data. This program takes the difference between the image and a 1 pixel shifted version, in order to avoid sensitivity to gradients in the input image. What is left is the 1st derivative of ths signal. The program attempts to determine the row-by-row slope. It attempts to reduce further the impact of real sources by performing the analysis on all 8 cells in a row at once. It has the option of only applying the fitted trend to a specified subset of cells. The IPP team has incorporated 'tiltystreak' into the IPP pipeline, and defined recipes that allow the analysis to be turned on optionally, and for the desired cells and chips to be specified for processing.

The second processing mitigation we have pursued is at attempt to model more accurately the effective readnoise introduced by this structure. By default, the camera reports a single read-noise value for each cell. We have created a tool (ppNoiseMap) to measure a robust variance for collections of pixels in a grid across each cell. This measurement can be made on a number of (eg) very short dark exposures, with a single high-quality measurement generated from the median of the multiple measurements. The resulting information can be used as a more accurate model of the read noise, with the noise level modeled on the subcell level. We have added an option to the standard processing to use this noise model to determine the noise characteristics of any given science image (combining in quadrature for each pixel the read noise from the model and the contribution from the poisson noise).

Testing

We have preformed a series of tests applying each of the two mitigations to the same collection of data. The input data for this test consisted of a series of short g-band exposures for which the Magic analysis had previously given very poor results. Unfortunately, as discussed below, these tests show that neither of these two mitigations is sufficient, though it appears that they substantially improve the situation.

We have run 3 tests with the g-band sequence. In the first test (noise.20090618), we only applied the resolved noise model; In the second test (tilty.20090619), we only applied the 'tiltystreak' correction. The critical measurement of success or failure is the number of streaks detected by magic. Both tests result in 1000 - 2000 streaks detected. Since there are only at most a few real streaks in these images, the vast majority of these are false positives. (HOW MUCH DID it improve relative to neither applied??). The tilty.20090619 test gave marginally better results than the noise.20090618 test, but both were clearly unacceptable. We ran a third test in which both the noise model and 'tiltystreak' were applied, in addition to a somewhat modified static mask, as discussed below. This third test (noise.20090620) was again a modest improvement over tilty.20090619, but the number of detected streaks was still too high (approaching 2000).

Detailed Inspection.

Below are a series of plots and images which illustrate the noise properties of the GPC1 images and how they are affected by the mitigations discussed above. The first series shows the evolution of the noise properties of a detrended chip from a single image in the native chip pixel coordinate system. All three plots show the log(Npixels) as a function of the predicted signal-to-noise (background subtracted). A Gaussian has been fitted to the core of this distribution, and is over-plotted in blue. If the noise model and background subtraction were perfect, the fit would result in a Gaussian with sigma 1.0 and a mean of 0.0, and the profile would follow the fitted curve all the way down. This image is a positive image of the night-time sky, so we expect pixels with high significance due to the presence of stars.

The first plot shows the result for noise.20090618, which only the noise model applied. The second plot shows the result for tilty.20090619, with only tiltystreak applied. The third shows the result for noise.20090620, with both applied and the additional masking. This chip is known to have significant correlated read-noise.

The set of images below show the S/N image used for the measurements above. In the first image, the correlated read noise is evident. In the second, tiltystreak has improved the correlated read noise, but has introduced an artifact on the first and last few rows of the chip. We noticed that this effect was a typical result from tiltystreak, so for the third experiment (noise.20090620), we extended the static mask by a few rows on each cell. This was an attempt to make a crude fix, with the intention of doing a better job if the impact seemed promising. The third image in the set shows that this additional masking was not quite sufficient in all cases.

The next set of plots and images are equivalent to the above set, but for the warped images generally corresponding to the above chips. The goal is to show that the warping does not significantly affect the noise predictions, and indeed it does not seem to have an adverse impact.

The final set of plots and images shows the equivalent to the above for the difference images for this exposure. The critical difference between this set at the warped version is that the real astronomical source should be subtracted, so that the high end deviations should be removed. It is quite clear that in all cases, the amount of signal at the high end is reduced. There is also an increase in the effective noise level. Although the third experiment (noise.20090620) is the best of the three, it still has a noise level which is unexpectedly high (sigma = 1.24).

Although the elevated Gaussian noise is a concern, and needs to be studied further, the observed magic-detected streaks do not seem to be the result of poorly estimated Gaussian noise. The images below show the location of the detected streaks, along with small segments corresponding to the pixels responsible for the streaks. These streaks are not randomly distributed across the focal plane as would be expected if they were caused by Gaussian noise. Closer examination of the images shows that other remaining linear features are largely responsible for triggering magic.

It seems that the static mask current in place is not sufficiently conservative. The next few images show some of the types of features which are not covered by the static mask. The first shows the vignetted edge of the focal plane. The next shows a glowing pixels around a bad cell. The third shows some hot columns. Our next step is to improve the mask by examining a set of processed science images as shown here and to manually mark poor pixels where were not caught in the initial mask creation.

At this point, I conclude both 'tiltystreak' and the noise model are needed to produce generally well-characterized images, but that two possible remaining issues are contributing to the poor behavior of magic:

1) insufficiently conservative masking of some features in the camera.

2) excessive Gaussian noise in the difference images relative to the warps.

The former can be addressed with ~1 day of work by the IPP team in manually reviewing the static masks. The latter needs some additional investigation, and may be due to a software error, or may simply be a result of the correlated noise properties of the pixels in this camera.

Attachments