In this sprint I did some investigations into the LSST deblending algorithms, with three main outcomes.
1. The SDSS deblender is not the same algorithm as the current LSST deblender. The LSST deblender will remove pixels from the boundaries of footprints that it thinks are not important and should not have been included in the footprint. The deblender modifies both the child deblend footprint AND the parent footprint. The current LSST deblender also does not do a linear fit to the amplitudes of each component before reassigning fluxes. The algorithm as also implemented with control parameters not exposed to configuration, and in a sufficiently obfuscated way that I am not confident the other functions are performing the same as the SDSS version. They may be, but it is hard to verify.
2. Measurement algorithms run on scarlet output models follow the trend of outputs produced by standard deblending. A giant caveat to this is that some clumping can be observed in certain outputs, but this is an already known issue related to small box sized used in the scarlet modeling stages and should soon be addressed. Most observed deviation in measurements will need to be explored using fake objects where a concrete "better" can be determined. Further investigations into any subtle biases that may be present will also be contingent on knowing the "truth" of a measurement.
3. Looking at the outputs produced by flux measurement routines revealed an issue present in calculating the errors on fluxes. There seems to be definite clumsiness to greater or lesser extents present in flux errors that does not seem present in the flux values themselves. These show up in both forms of deblending leading to the conclusion that further work on these flux algorithms is warented.
Story
In this sprint I did some investigations into the LSST deblending algorithms, with three main outcomes.
1. The SDSS deblender is not the same algorithm as the current LSST deblender. The LSST deblender will remove pixels from the boundaries of footprints that it thinks are not important and should not have been included in the footprint. The deblender modifies both the child deblend footprint AND the parent footprint. The current LSST deblender also does not do a linear fit to the amplitudes of each component before reassigning fluxes. The algorithm as also implemented with control parameters not exposed to configuration, and in a sufficiently obfuscated way that I am not confident the other functions are performing the same as the SDSS version. They may be, but it is hard to verify.
2. Measurement algorithms run on scarlet output models follow the trend of outputs produced by standard deblending. A giant caveat to this is that some clumping can be observed in certain outputs, but this is an already known issue related to small box sized used in the scarlet modeling stages and should soon be addressed. Most observed deviation in measurements will need to be explored using fake objects where a concrete "better" can be determined. Further investigations into any subtle biases that may be present will also be contingent on knowing the "truth" of a measurement.
3. Looking at the outputs produced by flux measurement routines revealed an issue present in calculating the errors on fluxes. There seems to be definite clumsiness to greater or lesser extents present in flux errors that does not seem present in the flux values themselves. These show up in both forms of deblending leading to the conclusion that further work on these flux algorithms is warented.