Geographical classification of weathered crude oil samples with unsupervised self-organizing maps and a consensus criterion

Abstract

Weathering processes considerably change the chemical composition of crude oils spilled on the environment. The impact of this effect on the classification capabilities of Kohonen neural networks (self-organizing maps, SOM) trained to predict the geographical origin of crude oils, was studied. A set of 38 worldwide crude oils from six distinct geographical regions was considered and each product was weathered to build a database of 263 samples. Gas Chromatography–Mass Spectrometry was used to monitor weathering and to determine a suite of 22 ratios of PAHs (polycyclic aromatic hydrocarbons) and alkanes, which are currently considered resilient to weathering. Three different experiments were performed with SOMs to study the effect of oil weathering and analytical conditions on the final results: a) weathered and un-weathered samples were randomly distributed in the training and validation sets, b) the training set included only unweathered samples while the validation set consisted of all the weathered samples, and c) the operational conditions (instruments/technicians) were different for the samples of the training and validation sets. After training the different SOMs, the percentage of correct classifications was in the 83.8%–95.7% range for the validation sets. A consensus SOM associated to a confidence level is proposed in which the decision is taken based on the number of votes of the winning class. The lowest success on the classification of the most weathered samples was expected and can be explained easily by the important changes on the chromatographic profiles due to the disappearance of the lightest alkanes and the relative enrichment on the more substituted PAHs. Most incorrect classifications occurred because the samples activated neurons at the close neighbourhood of the regions where they should belong to. The ‘consensus’ approach improved the capability to predict unknown samples. Results were compared to those from discriminant partial least squares and a variable selection step before the development of the SOMs was investigated.