A modified backward elimination approach for rapid classification of Chinese ceramics using laser-induced breakdown spectroscopy and chemometrics
A modified backward elimination approach was proposed for feature selection (FS) to eliminate the redundant and irrelevant features from laser-induced breakdown spectroscopy (LIBS) spectra for rapid classification of Chinese archaeological ceramics. The major elements (Fe, Ca, Si, Al and Mg) of LIBS spectra in the ancient ceramics identified using the National Institute of Standards and Technology (NIST) database. The six different pre-processing methods were used to reduce the error caused by random factors and the influence of various non-target factors on the classification results, which could increase the comparability among the Chinese archaeological ceramics from different dynasties. The input features for random forest (RF) model were selected by a modified backward elimination approach and three assessment criteria of sensitivity, specificity and accuracy from full spectra. LIBS spectra pre-processed by mean centering with the optimal input feature were used to construct a RF classification model for different dynasty ceramics. As indicated by research results, the sensitivity, specificity and accuracy of RF model for the ceramics samples in the test set are 0.9526, 0.9910 and 0.9782, respectively. In this sense, available statistics proved excellent performance of Chinese archaeological ceramics classification. Compared with the predictive result using RF, VI-RF and SBS-RF models, the sensitivity, specificity and accuracy of a modified SBS-RF model are higher than the results by others. The results demonstrate that the proposed algorithm is more efficient to reduce the redundant features, computational time and improve the model performance, and it is a good alternative for rapid analysis in multivariate classification.