Pattern recognition applied to spectroscopy: Conventional methods and future directions

Spectroscopic techniques have gained importance in a wide range of fields because of their appeal as rapid, reliable and nondestructive analysis, in most cases requiring minimum sample pre-treatment. These features make spectroscopy techniques suitable for routine analysis in online mode processing facilities. However, due to the high complexity of spectral data, which contain a large number of variables, multivariate statistical analysis is required to recognize patterns from samples. The multivariate nature of these methods makes evaluation of the robustness a much more complex task in comparison to classical ruggedness testing, as applied in univariate methods. In this review, unsupervised and supervised algorithms conventionally used for qualitative and quantitative analysis are explored. Among them, artificial neural networks, hierarchical clustering, linear regression extensions and principal component analysis are highlighted. Following the recent breakthrough in powerful and fast growing spectroscopic technologies, such as hyperspectral imaging, new challenges in pattern recognition are emerging. Thus, we present an overview of the new and promising developments in pattern recognition methods for complex spectral data, including support vector machines and penalized regression methods for robust variable selection.