Ensemble methods for automatic masking of clouds in AVIRIS imagery

We describe the first-phase of an investigation into techniques for automatic cloud masking in remote sensing data. BCM Projection Pursuit networks are explored as a method of unsupervised feature extraction from AVIRIS images. Search vectors in this method discover directions in the data in which the projected data is skew or multi-modal, by minimizing a projection index which depends on higher moments of the projected data distribution. Ensemble methods are used to fuse information from extracted BCM features and to smooth the mapping of these features to classification of image pixels. Ensemble hierarchies contain multiple levels of networks, combining BCM at the lowest levels with backward propagation algorithms, based on cross-entropy minimization, at higher levels in the ensembles. Predicted cloud masks are compared against cloud masks derived from human interpretation; ensembles achieve better overall classification accuracy than single BP networks.