Mixtures of projection pursuit models: an automated approach to land cover classification in Landsat Thematic Mapper imagery

Unsupervised projection pursuit methods and principal component analysis are compared for extraction of features from Landsat Thematic Mapper imagery. On sequestered test data, PP projections improved separation of individual categories from all other categories; improvement ranged from a few percent to as much as ≈22%. End-to-end classification of land-cover, combining these features with supervised classifiers, is also described. For sequestered test data, this approach obtained 96.8% pixel classification accuracy in identifying the 14 land-cover categories. A mixture of experts model is also described for automatically partitioning the problem domain