Unsupervised classification of hyperspectral data: an ICA mixture model based approach

Abstract

Conventional unsupervised classification algorithms that model the data in each class with a multivariate Gaussian distribution are often inappropriate, as this assumption is frequently not satisfied by the remote sensing data. In this Letter, a new algorithm based on independent component analysis (ICA) is presented. The ICA mixture model (ICAMM) algorithm that models class distributions as non-Gaussian densities has been employed for unsupervised classification of a test image from the AVIRIS sensor. A number of feature-extraction techniques have also been examined that serve as a preprocessing step to reduce the dimensionality of the hyperspectral data. The proposed ICAMM algorithm results in significant increase in the classification accuracy over that obtained from the conventional K-means algorithm for land cover classification.

A general procedure for learning mixtures of independent component analyzers

“…Note that both f(s (n) k ) and p(s (n) k )are actually computed in a non-parametric manner (kernel density estimates) from (21) and (14), respectively, given a suitable algorithm to a more extensive field of applications. The estimation is asymptotically unbiased and efficient, and it is shown to converge to the true pdf under several measures, when a suitable kernel is chosen [41].…”

“…ICAMM has been applied in real applications such as: learning of natural image codes [5], image classification, segmentation and denoising [20], separation of voices and background music in conversations [5,11], unsupervised object image classification from multispectral and hyperspectral sensors [21,22], separating several sources with fewer sensors in teleconferencing [23] and extending the classical continuous hidden Markov model by modeling the observation densities as a mixture of non-Gaussian distributions [24]. ICAMM has also been applied in biosignal processing: separation of background brain tissue, fluids and tumors in fMRI images [6], analysis to identify patterns of glaucomatous visual field defects [12,15], assessment of EEG to detect changes in dynamic brain state [16], classification of breast cancer diagnostic data [13], and analysis of multi-phase abdominal CT images to highlight liver segments [17].…”

Active spectral imaging for standoff detection of explosives

“…The data present in the hyperspectral cube were analysed using the main-component method [45,46]. At the same time, it was noted in [38] that, applying the independent component method [47,48], one can obtain better results when separating the spectral characteristics of each individual component of a mixture.…”

Stochastic Resonance In Signal Detection And Human Perception

“…Our preliminary results show that suitable Gaussian SR noise can improve the accuracy of the Gaussian Maximum Likelihood Classifier (GMLC). The role of SR in this area will be further examined utilizing the extensive techniques developed at Syracuse University [Varshney and Arora, 2004]. Possible enhancement of other processing algorithms by adding SR noise will also be evaluated.…”