Unsupervised Learning and Feature Extraction in Hyperspectral Imagery

Nakao, Eduardo K.; Levada, Alexandre L. M.

Remotely sensed hyperspectral scenes are typically defined by large area coverage and hundreds of spectral bands. Those characteristics imply smooth transitions in the spectral-spatio domains. As consequence, subtle differences in the scene are evidenced, benefiting precision applications, but values in neighboring locations and wavelengths are highly correlated. Nondiagonal covariance matrices and wide autocorrelation functions can be observed this way, implying increased intraclass and decreased interclass variation, in both spectral and spatial domains. This leads to lower interpretation accuracies and makes it reasonable to investigate if hyperspectral imagery suffer from Curse of Dimensionality. Moreover, as this Curse can compromise linear method’s Euclidean behavior assumption, it is relevant to compare linear and nonlinear dimensionality reduction performance. So, in this work we verify these two aspects empirically using multiple nonparametric statistical comparisons of Gaussian Mixture Model clustering performances in the cases of: absence, linear and nonlinear unsupervised feature extraction. Experimental results indicate Curse of Dimensionality presence and nonlinear adequacy.