Processing spectral data

Spectral images offer more information on complex probes than either conventional imagery (yielding only one or few measurements per voxel) or conventional spectroscopy (resulting in only one or a few spectra per probe) can. Spectral imaging is thus starting to become ubiquitous in areas ranging from materials science and process control to remote sensing and medicine. However, the information concealed in the massive amount of data generated by spectral imaging is not as easily accessible as in conventional (gray value or color) images and as in conventional spectroscopy, hence calling for new methods to analyze and visualize spectral images. Broadly speaking, analysis methods can be classified in terms of the information used in the training phase, and by the extent to which they use spatial context in the analysis. This article gives a brief overview of 'unsupervised' methods that require sample spectral images during the training stage as well as specification of a degree of complexity of the sample, either in terms of number of cluster centers or intrinsic dimensionality of the data [principal component analysis (PCA)]. 'Supervised' methods, on the other hand, require a training set in which a class membership, or label, is available for each pixel. Our discussion includes methods that deal well with the high dimensionality and high degree of correlation among individual features, such as partial least squares (PLS) and linear discriminant analysis (LDA). All of the above methods ignore the spatial context when applied to individual spectra. Often, the label map can be assumed to exhibit some spatial coherence, a fact that can help in classifying low quality spectra and that can be exploited using conditional or Markov random fields. The methods are illustrated using examples from automated process control and quantitative medical diagnostics. Copyright c 2009 John Wiley & Sons, Ltd.