| Version 1 (modified by , 12 years ago) ( diff ) |
|---|
PV3 Stack depth issue
DRAVG noted that the coverage in PV3 did not agree with the coverage in PV2. The cause of the differences between PV2 and PV3 is the LAP bad exposure filter code. The LAP code attempts to determine at various stages if a given exposure should be abandoned for being bad based on the subcomponent quality values. Each of the chip, camera, and warp stages have different thresholds for how many subcomponents can be bad before that exposure is rejected. Processing may continue for this exposure, but it will not be included in the stacks. The goal of this is to exclude obviously faulty exposures. The thresholds are (0.05, 0.0, and 0.2) for the three stages, representing 3 OTAs at the chip stage, any problem in the astrometry, and 20% of the variable number of skycells.
For the 137354 exposures currently part of PV3, I compared the quality distributions between the PV2 and PV3 releases. Of these, 2013 exposures changed whether they were retained (less than 20% bad) or rejected, with 1138 exposures that were excluded from PV2 now being included in PV3, and 875 exposures that were included in PV2 that have now been excluded from PV3. Splitting according to quality issue:
| Reason | PV2, not PV3 | PV3, not PV2 |
| overlap/8007 | 5 | 11 |
| pixels/3006 | 489 | 468 |
| unknown/257 | 381 | 534 |
| unhandled/case | 0 | 125 |
In any case, it's clear that the 20% threshold at the warp stage was too restrictive, as it appears to have caused ~5% of exposures to be excluded from the LAP processing. For all future processing, this threshold has been raised to 30%, which will exclude <0.1% of exposures.
For the stacks that have already been constructed, we will recreate the stacks including the exposures that had been excluded before. This will ensure that PV3 has a consistent depth across the sky. As all warps are already on disk, no chip to warp processing is necessary for this operation. After doing the same set of cuts, and ensuring that stacks have more than a single input, this will require making 115963 replacement stacks, about a third of the current total of 356821.
