Application of multivariate chemometrics in forensic soil discrimination based on the UV-Vis spectrum of the acid fraction of humus

Forty-four soil samples from five different areas were examined on the basis of the UV-Vis spectrum of the acid fraction of humus with a view to achieving good discrimination between them. Fulvic and humic acids were extracted from the samples into an alkaline aqueous solution and absorbance values, after appropriate transformations, were subjected to a K-mean cluster analysis (CA) over the objects (samples) for an initial feature reduction (20 variables retained). This was followed by principal component analysis (PCA) for the removal of outliers (four samples removed). The same statistical technique was used on the remaining samples to decide how many variables to enter into the linear discriminant function analysis (DA) and whether original variables or component scores should be used. It was found that the scree test was a good criterion for deciding on the number of components to extract (nine components extracted) and that the use of component scores instead of original variables led to a lower average redundancy (20.6%) of the variables in the discriminant model. From the components entered into the model and their loadings, it was concluded that the discrimination achieved was due to the relative concentration of aromatic groups and other fragments in the samples as well as the degree of soil humification. An overall 85% correct classification of the training dataset was observed (Wilks' lambda = 0.0420) and the method was judged satisfactory for supporting exclusionary forensic purposes.