Independent component analysis and multivariate curve resolution to improve spectral interpretation of complex spectroscopic data sets: Application to infrared spectra of marine organic matter aggregates

Abstract

The aim of this study is the identification of the most reliable multivariate technique for an efficient structural characterization of complex IR data sets. Different samples of organic matter (OM) aggregates were chosen as an object of investigation. In fact, the simultaneous presence of different biomolecules such as carbohydrates proteins and lipids makes difficult the interpretation and the comparison of the most significant fractions and structures, which describe the mechanisms of OM aggregation. With this aim, we collected the FTIR spectra of several samples of normal and anomalous size aggregates of organic matter and then submitted them to independent component analysis (ICA) by means of the algorithms fast independent component analysis (FastICA), joint approximate diagonalization of eigenmatrices (JADE), mutual information least dependent component analysis (MILCA) as well as to multivariate curve resolution alternate least squares (MCR-ALS). Among ICA algorithms, the MILCA was the most efficient because it always allowed a good spectra resolution and a higher number of significant and chemically interpretable components than the FastICA and the JADE algorithms, avoiding in addition the presence of some spectral ambiguities often observed when the two last ICA algorithms were applied. For the examined spectral sets, MCR-ALS gave comparable results with MILCA in terms of the number of identified components and for the exclusion of spectral ambiguities, though we observed some differences between MILCA and MCR-ALS. For instance, MCR-ALS spectra generally resulted more resolved than MILCA ones for all spectral sets, supporting the reduction of the spectral noise, but for one specific spectral set, it showed some misleading spectra, not observed in the corresponding MILCA treatment.The different abilities of these multivariate methods were briefly discussed. The influence of different pre-processing techniques (mean-centering, Pareto scaling, derivatives) on the decomposition results was also investigated. ICA and MCR-ALS deconvolution combined with appropriate preprocessing are capable to deal with complex IR data sets regarding separation of meaningful information from noise and its interpretation.