Multispectral data classification using similarity factors

The problem of classification through spectral information can be addressed using many different techniques, from traditional k-nearest neighbor or linear (or quadratic) discriminant analysis to partial least-squares discriminant analysis, or more sophisticated approaches such as support vector machines or wavelet-based methods. In many cases the available spectra, independently from their origin (e.g. mass spectroscopy, light spectroscopy, hyperspectral imaging), require several pretreatments before any classification method can be applied, and typically the most appropriate preprocessing of raw spectra is determined by trial-and-error. In this paper, we propose the use of similarity factors based on principal component analysis to classify different types of spectral datasets. The proposed classification technique has a very intuitive graphical interpretation and works through an assigned sequence of pretreatment steps on the raw signals, which avoids the trial-and-error selection of the most appropriate preprocessing method. The proposed strategy is successfully validated through two food engineering case studies. A third case study illustrates how the similarity factor-based strategy can be extended to the classification of multi/hyper spectral images.