Discrimination and classification of FTIR spectra of red, blue and green spray paints using a multivariate statistical approach.

The use of multivariate techniques was investigated in a forensic paint analysis context. The data set consisted of the infrared spectra of 74 spray paint cans, corresponding to three colors code, respectively red, green and blue. Two aspects of the forensic procedure are studied, respectively, the discrimination of paints coming from a market study through exploratory techniques, and the source prediction of unknown samples in database using classifiers. The exploratory discrimination capabilities of principal component analysis (PCA) and hierarchical clusters analysis (HCA) were compared to a visual comparison of the spectra. Iterative PCA was found to be the most adapted solution for exploratory analysis of the samples. Very few differences were found compared to a visual comparison of the samples and the statistical foundations behind the method ensure that no errors are due to a misclassification of the samples. Market studies and joint PCA also represent a significant gain of time. Following that, classification and prediction of future samples were evaluated by means of supervised techniques of classification such as linear/quadratic discriminant analysis (LDA/QDA), support vector machines (SVM), soft independent modeling of classes analogies (SIMCA) and partial least squares discriminant analysis (PLS-DA). SIMCA was the preferred method, as it provided the smallest false negative rates together with a correct classification rate of about 95%. From an investigative point-of-view the presence of false positives was considered acceptable, as it is preferable to have a longer list of possible sources but have confidence that the true source belongs to it.