Recently we had a query about how SkyView combines data where there is an overlap of plates in the Digitized Sky Survey datasets. The short answer is we don’t, but I thought the longer response might be of general interest.
For [any given pixel on] the DSS (or DSS2) plates, SkyView only ever samples data from a single plate. It never combines data from multiple images. The ‘best’ plate is always selected for each individual pixel, where best is typically defined as the plate where a given pixel is farthest from the edge of the plate (though that can be overridden).
There are a few of reasons for this…
1. Simplicity in the code.
2. Adapting to the different characteristics of multiple plates is non-trivial. Adjacent plates can have very different backgrounds and somewhat different resolutions. These are not necessarily constant over a plate so they would have to be adaptively solved for so that we would know how to deal with edges properly and combine data effectively.
3. The edges of the plates have many gross artifacts (e.g., scans beyond the edges of the emulsions, labels, calibration wedges, etc.). Detecting these is non-trivial and so simply avoiding the edges as much as possible is desirable. There are also a fair number of defects in the interiors of the plates (emulsion flakes, hairs, dust, scratches), and combining plates combines all of the defects.
4. Given the non-linear nature of the plates, analysis of combined data is non-trivial and the increase in sensitivity is at best <~2 in the rare areas where 4 plates overlap (though in that case at least one is likely to be near the edge so that its defects would dominate any small gain). One thing that I have not considered is whether the fact that the images are being regenerated from a lossy compressed image would need to be addressed. I don’t know if anyone has studied that.
SkyView does try to minimize the appearance of edges between DSS plates using a de-edgeing algorithm. SkyView recognizes that the zero point of individual plates are arbitrary. SkyView‘s deedging algorithm looks at the edges between each pair of plates and looks at the histogram of jumps between adjacent pixels. It then adds an offset to each image to make the median offset between images 0. The offsets added are given in the associated FITS header. Other algorithms are also available.
Generally this simple algorithm works so long as there are only a few plates involved. When very large numbers of plates are involved, it breaks down. Some kind of relaxation method is probably appropriate there (or possible adding a non-constant backgrounds) but we haven’t adopted any.
Note that for CCD images the arguments against adding plates are much weaker, and we are looking at adding a new mosaicking tools in SkyView that would support image addition. If you have suggestions or comments about how we might do this, I’d be very interested.