Study of preprocessing sensitivity on laser induced breakdown spectroscopy (LIBS) spectral classification

Laser induced breakdown spectroscopy (LIBS) is an atomic emission based spectroscopy that uses a laser pulse as the source of excitation. The laser is focused to form hot plasma, which atomizes and excites the sample. In the LIBS spectrum each “feature” is the amplitude or intensity detected at different wavelengths in the range of 200-1000 nm. Pattern recognition techniques were applied on samples with similar elemental composition resulting in almost similar LIBS spectra which are visually very difficult to differentiate. It was observed that the classification results obtained from different classifiers were sensitive to data preprocessing. The outlier detection and removal techniques PCA, Dendrogram using Agglomerative Algorithm, Editing by Nearest Neighbour (NN) and Distance Matrix approaches were used in preprocessing step. After removing outlier(s) the resulting training patterns were used to model the k-Nearest Neighbour (k-NN), Principal Component Analysis (PCA), Dendrogram, Multiclass Support Vector Machine (SVM) and Decision Tree classifiers. In k-NN after removing outlier(s) the average classification accuracy was increased by 2% for high energy materials (HEM), but no improvement in non high energy materials (Non HEM) or in top level classification (decide either HEM or Non HEM). But, for other classifiers the classification accuracy gets reduced. Finally instead of removing outlier(s) dimensionality reduction by thresholding was applied and the classification accuracy increased by 4% in k-NN for HEM and 38% in multiclass SVM for HEM and 4% for Non-HEM.