Read Noise

From the Noise data from CD15,  (cd15_13jun/41/*618 to 622 (room) and 623 to 627 (mirror)), I've done some analysis.  I think the mirror data are totally dominated by read noise and the room data are dominated by non-linearity.  This post is for the read noise.

The five FITS files were each NSNSNSD x 16 repeats stored as 1024,1024,64 fits files.  ctdelay = 70 which corresponds to a frame time of 2.055 seconds.

I read all files in.  The first two NSNSNSD patterns for both room and mirror had clear offsets, so I discarded those.  Now I had two 1024,1024,56,5 arrays.

For the mirror, I took the difference of N_2 and N_1 and N_4 and N_1.  I now have two 1024,1024,14,5 arrays.  I also noticed the median of each difference was different for the different files by a few DN.  I subtracted off the median of each file's difference and added the median of the first file to try and remove this offset.

The signal should increase by a factor of 3 between these two differences.  The actual value is closer to 2.8, indicating non-linearity hits early.

I did the same processing on the room data for N_2 - N_1 and adjusting the medians of each file.

I reformed the data to 1024,1024,70 arrays so that I could look at the median and standard deviation of each pixel through the time series.  The median and noise images show patterns, but since I'm looking at each pixel through the 70 images separately, I don't think that matters too much.

The read noise is clear in the mirror data.  That figure is below.   There is some difference in the level of read noise that depends on the amplifier chain (columns 1,5,9...; 2,6,10...; ...).  Clearly, the noise is about the same for both differences, even though the signal is almost 3x larger.  This makes me think we're limited by read noise.

The noise vs signal for each pixel in a 1024,1024 by 70 time series.  The blue is the difference of sample 4 and sample 1 in a NSNSNSD pattern.  The black (left hand side) is the difference of sample 2 and sample 1.  

• all pixels: median noise = 8.99
• columns 0,4,8,...: median noise = 8.82
• columns 1,5,9,...: median noise = 11.06
• columns 2,6,10,...: median noise = 8.92
• columns 3,7,11,....: median noise = 8.07

Since the noise is coming from a difference, it is square root 2 higher than for a single read.  Taking the median = 9.0 ADU, the single read = 6.4, and the number of electrons = 222 e-.  In comparison, John Lacy found read noise around 3.5 ADU for 2x2 pixel averaging (=> 7 ADU for difference => 175 e-). We really need to get this level of read noise down.  I thought we were at 70 e- for a single read.  I suppose there could be some other noise source that I'm not correcting for that is dominating.

From the time series looking at the room, I got the corresponding noise vs signal plot below.  The noise is starting to come off the 9 ADU level.

My whole point of looking at these data was to try and recalculate eta, g, and beta.  I don't think that is possible with readnoise dominated data.  I'll try to linearize the room data and see if I can come up with a good answer.