Training Sets for Statistical Feature Recognition in Multidimensional Solar Imagery

We have previously reported the multi-dimensional extension of a statistical maximum likelihood algorithm for segmenting images into different feature classes developed by Turmon, Pap, and Mukhtar (2002, ApJ 568, p. 396). The method works best for features which have overlapping but nonetheless distinct distributions of observed variables. Developing these empirical class-conditional distributions from independently classified training sets depends sensitively on the match of spatial scales between the training segmentations and the desired feature classes. We discuss recent progress in extracting well-posed class distributions even when the training segmentations are mixtures of the classes which we wish to identify. For example, in addition to large-scale labelings, Harvey and White (1999, ApJ 515, p. 812) provide finely grained information which we use to help isolate areas of pure quiet Sun. Quiet Sun distributions of observed quantities can then be separated from distributions derived from areas labeled as network which also include quiet Sun. Similarly, these distributions can then be isolated from those mixed with active regions and/or sunspots. This research is funded by a NASA Supporting Research and Technology grant.