Bad pixels

The recipes apply a predetermined bad-pixel mask with the aim of removing the bulk of `hot' and `cold' pixels. The approximate percentages of pixels masked for each instrument is as follows: Classic Cam1.1, INGRID 1.0, IRCAM 0.1, IRIS2 0.1, ISAAC 0.4, UFTI 0.7, and UIST 0.4.

The Michelle bad pixel mask is time dependent; initially 12 aberrant pixels were identified, but when the new detector returned on UKIRT in 2004, the mask merely ignored the top sixth of the array.

For the test data available, NACO had large numbers of very noisy pixels to the foot and sides of the detector, hence the default bad-pixel mask has 9.9% bad pixels. If your NACO data do not show this, you can form your own mask from a long-exposure dark frame, looking for the highly deviant pixels. See Creating a bad-pixel mask for suggestions.
[_MASK_BAD_PIXELS_]

There are two problems. First, the pre-calculated mask only accounts for 95% of UFTI's problem pixels. The other 5% are occasionally deviant on timescales of days. The variability of IRCAM, Michelle, UIST, and IRIS2 bad pixels is unknown at the time of writing. In addition the bad-pixel masks have not been regularly monitored prior to 2000 August. The result is that non-physical values could appear in the processed data, some as extreme as $-10^{-31}$ causing automatic registration and image display to go awry.

Therefore, after dark subtraction, recipes apply thresholding which flags non-physical values as bad, meaning undefined. This is just augmenting the bad-pixel mask, and no valid data are lost. The upper limit is above the nominal saturation levels: 16000 for Classic Cam; 30000 for INGRID; 20000 for IRCAM in STARE or NDSTARE mode, and 33000 using the Deepwell; 200000 for IRIS2 and ISAAC; 100000 for Michelle; 4300 for NACO in FowlerNsamp and Double_RdRstRd modes, but 12400 for Uncorrelated reads except for the $M^\prime$ band where it is 28000; 17000 times the number of coadds for NIRI; 15000 for UFTI; and 20000 for UIST. The lower limit is the 2-, 3-, 3-$\sigma$-clipped mean, approximating to the mode, less five times the clipped standard deviation, $\sigma$. While a positive threshold looks attractive, small negative values, while appearing non-physical, can arise through noise. Therefore, to avoid a bias (mainly in the $J$ band), a further constraint is that the lower limit lies in the range $-$100 to 1.
[_SUBTRACT_DARK_, $<$instrument$>$/_GET_SATURATION_LEVEL_]

Recipes issue warnings if the dark-subtracted frame's mode is negative, allowing for the error in the mode. It aborts with an error message if the modal dark-subtracted signal is more than one standard deviation negative. These states usually arise because of an aberrant dark.