Uploaded image for project: 'Data Management'
  1. Data Management
  2. DM-21329

Determine appropriate pixel scale for DECam coadds

    Details

    • Type: Story
    • Status: Done
    • Resolution: Done
    • Fix Version/s: None
    • Component/s: None
    • Labels:
      None
    • Story Points:
      16
    • Sprint:
      AP F19-4 (2H September), AP F19-5 (October), AP F19-6 (November), AP S20-1 (December)
    • Team:
      Alert Production

      Description

      https://jira.lsstcorp.org/browse/DM-21230 identified that our DECam HiTS templates were using the pipeline-default 0.33"/pixel, and that 0.2"/pixel provide a 2x decrease in the number of (false) DIASources. Subsequent Slack discussion suggested that the pixel scale should not matter so long as the PSF was Nyquist sampled and the correlated noise was handled correctly in the coaddition. The latter condition is not met, so while awaiting future algorithmic advances this ticket is to determine an empirical best choice for DECam.

      This ticket is to produce HiTS templates for a small area for several pixel scales, ranging from (perhaps) 0.1"/pixel, including the instrumental pixel scale (0.26"/pixel), and up to and exceeding the Nyquist scale for the HiTS seeing distribution. The output is a plot of the number of detected DIASources as a function of the template pixel scale.

        Attachments

          Issue Links

            Activity

            Hide
            mrawls Meredith Rawls added a comment -

            I am nearly done making CompareWarp coadd templates with HiTS 2014 dataset at 6 different resolutions. I am using a best seeing criterion and restricting the dataset to a handful of ccds in one field (Blind14A_04, which corresponds to Blind15A_26). This results in 3-9 coadd patches depending on the configured skymap resolution. Nearly all of the regions with data have 21 constituent visits. The processed images are on lsst-dev in `/project/mrawls/hits2014/rerun/2019_11` and the coadds are in reruns named `coadds1` through `coadds6`. The steps are:

            • ingest the raw HiTS 2014 data and calibs and run `processCcd.py`
            • `makeDiscreteSkyMap.py` with --config skyMap.pixelScale set to the present default (0.33), 0.2, 0.23, 0.26, 0.29, and 0.32
            • `makeCoaddTempExp.py` with the ap_pipe makeCoaddTempExp_goodSeeing config as well as --config select.nImagesMax=1000 and select.maxPsfFwhm set to 4.2, 6.93, 6.03, 5.33, 4.78, and 4.33, respectively (corresponding to the pixelScales above)
            • `assembleCoadd.py` with --warpCompareCoadd and --id manually set to include all patches which have a tempExp from the previous step (there are least 3 and at most 9 data-containing patches)
            • `ap_pipe.py` using the HiTS 2015 dataset as science and the new coadds as templates
            Show
            mrawls Meredith Rawls added a comment - I am nearly done making CompareWarp coadd templates with HiTS 2014 dataset at 6 different resolutions. I am using a best seeing criterion and restricting the dataset to a handful of ccds in one field (Blind14A_04, which corresponds to Blind15A_26). This results in 3-9 coadd patches depending on the configured skymap resolution. Nearly all of the regions with data have 21 constituent visits. The processed images are on lsst-dev in `/project/mrawls/hits2014/rerun/2019_11` and the coadds are in reruns named `coadds1` through `coadds6`. The steps are: ingest the raw HiTS 2014 data and calibs and run `processCcd.py` `makeDiscreteSkyMap.py` with --config skyMap.pixelScale set to the present default (0.33), 0.2, 0.23, 0.26, 0.29, and 0.32 `makeCoaddTempExp.py` with the ap_pipe makeCoaddTempExp_goodSeeing config as well as --config select.nImagesMax=1000 and select.maxPsfFwhm set to 4.2, 6.93, 6.03, 5.33, 4.78, and 4.33, respectively (corresponding to the pixelScales above) `assembleCoadd.py` with --warpCompareCoadd and --id manually set to include all patches which have a tempExp from the previous step (there are least 3 and at most 9 data-containing patches) `ap_pipe.py` using the HiTS 2015 dataset as science and the new coadds as templates
            Hide
            mrawls Meredith Rawls added a comment -

            I completed the processing as described above and examined the results. The DIA Source counts are closer than I expected between each template, but it does appear that the instrumental pixel scale of 0.26 arcsec/pixel is a good choice.

            The plot below shows source and object counts for each template pixel scale. It shows three different things: total number of DIA Objects (blue), DIA Sources with the bad, suspect, and/or saturated center flag (red), and DIA Sources without any of those flags (orange). Since we believe many of the sources are erroneous (i.e., not true detections of astrophysical variability), especially when a sub-optimal template resolution is used, fewer sources and fewer objects are generally seen as good. 

            I also plotted the location of all these sources on the sky to get a feel for the relative small size of the dataset used (just a few CCDs), where the obviously bad sources are, and see if anything else stands out. Below, blue and red points represent "good" (none of the three aforementioned flags) and "flagged" DIA Sources, respectively. Each point is plotted with high transparency, and all six reruns are plotted on top of each other. As a result, darker points mean that a source was found in most or all of the six reruns, and lighter points mean a source was only found in one or two reruns (so it is more likely to be spurious). For the most part, the same sources are found in each rerun, although they are not always precisely on top of each other.

            Show
            mrawls Meredith Rawls added a comment - I completed the processing as described above and examined the results. The DIA Source counts are closer than I expected between each template, but it does appear that the instrumental pixel scale of 0.26 arcsec/pixel is a good choice. The plot below shows source and object counts for each template pixel scale. It shows three different things: total number of DIA Objects (blue), DIA Sources with the bad, suspect, and/or saturated center flag (red), and DIA Sources without any of those flags (orange). Since we believe many of the sources are erroneous (i.e., not true detections of astrophysical variability), especially when a sub-optimal template resolution is used, fewer sources and fewer objects are generally seen as good.  I also plotted the location of all these sources on the sky to get a feel for the relative small size of the dataset used (just a few CCDs), where the obviously bad sources are, and see if anything else stands out. Below, blue and red points represent "good" (none of the three aforementioned flags) and "flagged" DIA Sources, respectively. Each point is plotted with high transparency, and all six reruns are plotted on top of each other. As a result, darker points mean that a source was found in most or all of the six reruns, and lighter points mean a source was only found in one or two reruns (so it is more likely to be spurious). For the most part, the same sources are found in each rerun, although they are not always precisely on top of each other.
            Hide
            mrawls Meredith Rawls added a comment -

            The output from running ap_pipe is on lsst-dev in /project/mrawls/hits2015/rerun/coaddtestN, where N goes from 1 to 6 corresponding to the template pixel scales 0.33 (default), 0.2, 0.23, 0.26, 0.29, and 0.32, respectively.

            Show
            mrawls Meredith Rawls added a comment - The output from running ap_pipe is on lsst-dev in /project/mrawls/hits2015/rerun/coaddtestN, where N goes from 1 to 6 corresponding to the template pixel scales 0.33 (default), 0.2, 0.23, 0.26, 0.29, and 0.32, respectively.
            Hide
            mrawls Meredith Rawls added a comment -

            After some discussion, we decided to add a few more pixelScale resolutions: 0.13" (extremely high-res), 0.52" (extremely low-res), and 0.8" (ridiculously low-res). The maxPsfFwhm cut for the images that will make up the template was adjusted as before - 4.2 * 0.33 / pixelScale - yielding 10.67, 2.67, and 1.73, respectively.

            When this ticket was originally made, a larger difference was seen between ap_pipe runs on the HiTS dataset with lower- vs higher- resolution templates. Much of this difference was likely due to misconfiguration, and in particular the need to scale the allowed maxPsfFwhm for selecting constituent images for a template along with the desired template skymap pixel scale.

            There is only a small difference in source (and object) counts between the old default, 0.33"/pixel, and the instrumental pixel scale for DECam, 0.26"/pixel. Nevertheless, the instrumental pixel scale is a sensible default, and there are about 5% fewer DIA Objects than the 0.33"/pixel case.

            Going down to an extremely high resolution template (small pixel scale ~ 1/2 the instrumental value) is computationally inefficient and does not improve the source or object counts. Going up to an extremely low resolution template (large pixel scale > 2x the instrumental value) does eventually produce a large quantity of spurious detections.

            In generating these final plots and source/object counts, I have also taken care to consider only a spatial region that overlaps all of the reruns.

            The takeaway is unchanged - the instrumental pixel scale for DECam still seems to be a good choice.

            Show
            mrawls Meredith Rawls added a comment - After some discussion, we decided to add a few more pixelScale resolutions: 0.13" (extremely high-res), 0.52" (extremely low-res), and 0.8" (ridiculously low-res). The maxPsfFwhm cut for the images that will make up the template was adjusted as before - 4.2 * 0.33 / pixelScale - yielding 10.67, 2.67, and 1.73, respectively. When this ticket was originally made, a larger difference was seen between ap_pipe runs on the HiTS dataset with lower- vs higher- resolution templates. Much of this difference was likely due to misconfiguration, and in particular the need to scale the allowed maxPsfFwhm for selecting constituent images for a template along with the desired template skymap pixel scale. There is only a small difference in source (and object) counts between the old default, 0.33"/pixel, and the instrumental pixel scale for DECam, 0.26"/pixel. Nevertheless, the instrumental pixel scale is a sensible default, and there are about 5% fewer DIA Objects than the 0.33"/pixel case. Going down to an extremely high resolution template (small pixel scale ~ 1/2 the instrumental value) is computationally inefficient and does not improve the source or object counts. Going up to an extremely low resolution template (large pixel scale > 2x the instrumental value) does eventually produce a large quantity of spurious detections. In generating these final plots and source/object counts, I have also taken care to consider only a spatial region that overlaps all of the reruns. The takeaway is unchanged - the instrumental pixel scale for DECam still seems to be a good choice.
            Hide
            sullivan Ian Sullivan added a comment -

            Thank you for your additional analysis with very coarse and very fine resolution. In the future we should come up with a better metric for assessing the difference imaging algorithms' performance other than just the number of DIA objects and sources that are detected, but this is good for now.

            Show
            sullivan Ian Sullivan added a comment - Thank you for your additional analysis with very coarse and very fine resolution. In the future we should come up with a better metric for assessing the difference imaging algorithms' performance other than just the number of DIA objects and sources that are detected, but this is good for now.

              People

              • Assignee:
                mrawls Meredith Rawls
                Reporter:
                swinbank John Swinbank
                Reviewers:
                Ian Sullivan
                Watchers:
                Eric Bellm, Ian Sullivan, John Swinbank, Krzysztof Findeisen, Meredith Rawls
              • Votes:
                0 Vote for this issue
                Watchers:
                5 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved:

                  Summary Panel