The plate processing pipeline produces a variety of image features
for each object, thirty of which are used by the classification
task. Five oblique decision trees vote to produce a classification
of star, nonstar, or defect. Before the catalog
is exported, multiple observations of each object are combined to
produce a final classification of star or nonstar.
As one would expect, the performance
of the classifier varies with galactic latitude and with magnitude:
as the problem gets harder (crowded fields, fainter objects), the
error rate goes up. This is illustrated in the following comparison
of SPM stars matched to the GSC2.1. Galactic latitude is shown on
the y axis, GSC2 F magnitude is shown on the x axis, and completeness
is shown on the z axis.
|