There are a number of considerations when it comes to point source removal from a mosaic! If you are interested in producing the cleanest possible image of the diffuse emission, then combining the source lists from the individual observations will probably not be satisfactory; even fainter sources below the threshold of a single observation will suddenly become very noticeable once you have stacked several observations together. On the other hand, pushing the point source detection limit as deep as possible will produce a very non-uniform point source removal limit, which may or may not be an issue. So what are our choices?
First, you have the point source lists (and masks) from each individual observation. One can apply to each observation the mask from that observation, and then merge. This has the difficulty that a source detected in one observation, but not detected in the other observations of that location, may be just below the detection thresholds in those observations. After mosaicking, it can be statistically significant and, due to the fact that you've removed that region in the one observation where it was detected, accentuated. On the other hand, if the source is well below the threshold in all the observations but one, when the source was flaring, for example, it does not make much sense to remove that region from the observations when it was not detected.
A second option is to produce a list of the unique sources from those lists using emlfill and emlmerge, and then recreate the source masks for the individual observations from that list of unique sources. The emlmerge routine is undergoing revision and is not currently available.
A third approach can be taken if the mosaic is not much bigger in area than a single observation. The observations of M101 are a good example. In this case you can merge the event files for each instrument over all of the observations, and then use that merged event list for source detection with cheese. (Remember that you will also want to merge the atthk.fits files as well.) There are a number of caveats!
First, edetect_chain, which is called by cheese does not handle strong changes in effective exposure very well, so you may want to fiddle with edetect_chain directly to get the parameters correct. (The emask_threshold1 parameter may need to be reduced.)
Second, although this process will produce an emllist.fits file with the deepest possible detection limit in each location, it may not pick up the sources in regions that have low relative exposure. You may need to augment this list with sources from the individual exposures.
Third, this list is not without its perils as well; since it has been constructed from overlapping images with (possibly) very different PSFs at the same point, the probabilities determined by emldetect may not reasonable. And the output list will require a great deal of hand selection! This can be done if you are looking for a really smooth image, but does not allow one to set a uniform source elimination criterion. In this third approach, you can use the emllist.fits from the merged observations to create masks for the individual obsids (i.e., §5.12.2).
Finally, the SAS routine emosaicproc should be considered as well. (We have not tried using it yet.)