Deblending Overlapping Objects

One of the jobs of the frames pipeline is to decide if an initial single detection is in fact a blend of multiple overlapping objects, and, if so, to separate, or deblend them. The deblending process is performed self-consistently across the bands (thus, all children have measurements in all bands). After deblending, the pipeline again measures the properties of these individual children.

Bright objects are measured at least twice: once with a global sky and no deblending run (this detection is flagged BRIGHT) and a second time with a local sky. They may also be measured more times if they are BLENDED and a CHILD.

Once objects are detected, they are deblended by identifying individual peaks within each object, merging the list of peaks across bands, and adaptively determining the profile of images associated with each peak, which sum to form the original image in each band. The originally detected object is referred to as the "parent" object and has the flag BLENDED set if multiple peaks are detected; the final set of subimages of which the parent consists are referred to as the "children" and have the flag CHILD set. Note that all quantities in the photometric catalogs (currently in the tsObj files) are measured for both parent and child. For each child object, the quantity parent gives the object id (object) of the parent (for parents themselves or isolated objects, this is set to the object id of the BRIGHT counterpart if that exists; otherwise it is set to -1); for each parent, nchild gives the number of children an object has. Children are assigned the id numbers immediately after the id of the parent. Thus, if an object with id 23 is set as BLENDED and has nchild equal to 2, objects 24 and 25 will be set as CHILD and have parent equal to 23.

The list of peaks in the parent is trimmed to combine peaks (from different bands) that are too close to each other (if this happens, the flag PEAKS_TOO_CLOSE is set in the parent). If there are more than 25 peaks, only the most significant are kept, and the flag DEBLEND_TOO_MANY_PEAKS is set in the parent.

In a number of situations, the deblender decides not to process a BLENDED object; in this case the object is flagged as NODEBLEND. Most objects with EDGE set are not deblended. The exceptions are when the object is large enough (larger than roughly an arcminute) that it will most likely not be completely included in the adjacent scan line either; in this case, DEBLENDED_AT_EDGE is set, and the deblender gives it its best shot. When an object is larger than half a frame,the deblender also gives up, and the object is flagged as TOO_LARGE. Other intricacies of the deblending results are recorded in flags described on the Object Flags section of the Flags page.

On average, about 15% - 20% of all detected objects are blended, and many of these are superpositions of galaxies that the deblender successfully treats by separating the images of the nearby objects. Thus, it is almost always the childless (nChild=0, or !BLENDED || (BLENDED && NODEBLEND); equivalently status = GOOD) objects that are of most interest for science applications. Ocassionally, very large galaxies may be treated somewhat improperly, but this is quite rare. For examples of the deblender at work, see Robert Lupton's page of all the Zwicky galaxies in run 756/20.

The behavior of the deblender of overlapping images has been further improved since the DR1; these changes are most important for bright galaxies of large angular extent (> 1 arcmin). In the EDR, and to a lesser extent in the DR1, bright galaxies were occasionally ``shredded'' by the deblender, i.e., interpreted as two or more objects and taken apart. With improvements in the code that finds the center of large galaxies in the presence of superposed stars, and the deblending of stars superposed on galaxies, this shredding now rarely happens. Indeed, inspections of several hundred NGC galaxies shows that the deblend is correct in 95% of the cases; most of the exceptions are irregular galaxies of various sorts.