Planery speaker: Nonlinear manifolds for feature extraction: Opportunities and challenge

Hyperspectral data provide enhanced capability for characterization of spectral signatures and thus, potentially improved discrimination of targets of interest. However, to fully exploit its potential, issues related to the high dimensionality of the data, correlated spectral bands, and nonlinear spectral responses must be addressed. Signatures represented in the narrow spectral bands are also more sensitive than their multispectral counterparts to nonstationarity of samples over extended areas and time. Increased availability of hyperspectral data and greater access to advanced computing have recently motivated investigation of nonlinear manifold learning approaches, which were originally developed within the machine learning community. Although primarily investigated for dimension reduction in a supervised classification framework, manifolds are also useful for semi-supervised classification and for representation of nonlinear structures in predictive models. Popular local and global manifold learning methods are summarized, with illustrations of their application to classification of remotely sensed data. Issues related to parameter settings are discussed, strategies to mitigate computational load are described, and potential extensions to accommodate local spatial variability and semi-supervised learning in dynamic environments are explored.