Classification with Artificial Neural Networks and Support Vector Machines: application to oil fluorescence spectra

This paper reports on oil classification with fluorescence spectroscopy. The investigations are part of the development of a laser-based remote sensor (laser fluorosensor) to be used for the detection and classification of oil spills on water surfaces. The polychromator of the fluorosensor has six channels for measuring signals that represent the spectral fluorescence signature of the detected oil in the UV/VIS wavelength range following excitation at 355 nm wavelength. The investigation of the oil classification is based on the shape of the signature of the oil detected by these channels. The investigation uses three methods to examine crude oils, heavy refined oils, and sludge oils: the channels’ relationships method (CRM); artificial neural networks (ANNs); and support vector machines (SVMs). This was done on a laboratory database of oil fluorescence spectra. The database and the input fluorescence signature of the oils play a very important role in the efficiency of the classification method. If the input fluorescence of the oil does not fit into one of the classes already included in the database or if it is a completely new and previously not considered signature, then the training process for classification must always be redone. Generally, all three methods yield promising results and can be used for the detection and classification of oil spills on water surfaces. The channels’ relationship method provides a meaningful classification of the available spectra, according to a rough oil type estimation. More specific substance information can be achieved with ANNs and SVMs. Both SVMs and ANNs prove to be efficient, fast, and reliable and have real-time capabilities. The SVM method is faster and more stable than ANN. Therefore, it is considered to be the most convenient method for classifying spectral information.