Eagle smear correction incomplete

[mggr]

Smear correction is the process of compensating for light continuing to accumulate on the detector while the data are read out row by row. Eagle has a frame shift read out and thus must have smear correction. Hawk does not.

Smear correction should be done prior to radiometric correction (i.e. at level 0) as pixel values will be incorrect if the smear is not removed.

The Eagle detector has 1024 bands, of which the middle 500 or so are used. All of the 1024 bands should be accounted for, as they may be accumulating light despite being unused. The 2009 Edinburgh calibration shows many of these bands do accumulate light.

The full correction can only be done if all 1024 bands are recorded, which only happens in the radiometric calibration experiments. This omission will result in errors.

We should seek to characterise the size of this error using the calibration data. Fixing it is impossible with the current hardware (which cannot keep up with recording 1024 bands). Some improvements may be possible if reducing this error is considered important enough:

• create a bandset that bins all the unused bands together to account (grossly) for the light accumulated in them
• characterise the light present in these bands for a number of situations and use this to make approximate corrections.

[mggr]

Additional points from Bill:

• the order of processing is not necessarily fixed (fsm before dark) - this appears to vary in the literature
• overflow will cause additional and possibly uncorrectable smear problems (per data quality reports on Reports)

[mggr]

iopa's report on smear correction

[mggr]

Attached a report from Ioannis detailing the predicted effect of unrecorded smear.

Snippet summary: It was discovered that with the lowest used integration times smear can account for up to 70% of the recorded light for certain (edge of capability) bands, whereas for mid-range times that can reach up to 15% for those bands.

However, in normal use and for mid-range bands, the effect is of the order of 0 - 2%.

[mggr]

Variable (software) binning does appear to work and is practical from a hardware-keeps-up point of view. Got a sample dataset to test whether the software binning is doing the right thing.