Dynamically reconfigurable projection pursuit ensembles for cloud detection in AVIRIS imagery

A Projection Pursuit method is used to find structure and reduce the complexity of high-dimensional data (input patches from AVIRIS imagery) by discovering a low-dimensional set of statistically interesting data projections. Individual Projection Pursuit networks in an ensemble focus on one of four spectral bands used in the study. We use preprocessing steps such as Gray-Level Difference Vectors and Sum and Difference Histograms, which are useful for cloud detection. In the past, most work with these particular histogram techniques has involved the extraction of pre-specified moments of the histogram. We show that when each input histogram is treated as a point in a high-dimensional data space, projection pursuit techniques can be used to find an underlying multi-modal structure useful for cloud detection. Because these projection techniques typically have a very large number of parameters, we also examine an online perturbation analysis technique that assesses the relative importance of projection parameters (dynamic reconfiguration). Ensemble methods combine features extracted from AVIRIS imagery by multiple Projection Pursuit networks to obtain pixel classifications using backward error propagation with a cross-entropy objective function. Predicted cloud masks are compared against human interpretation.