Spectral clustering based on histogram of oriented gradient (HOG) of coal using laser-induced breakdown spectroscopy

The quantification accuracy of laser-induced breakdown spectroscopy was limited due to matrix effects. In this work, a method named unsupervised-clustering-based quantification (UCQ) was proposed to reduce the matrix effects by clustering the spectra by k-means combined with the histogram of oriented gradient (HOG). The calibration spectra were clustered by k-means into k categories, and partial least squares regression (PLSR) models were established on each category. The unknown spectra were classified, and the models related to their categories were used in prediction. Leave-one-out cross-validation (LOOCV) was used to evaluate the accuracy of quantitative analysis. By clustering the spectra into three categories using HOG as the clustering features, the accuracy of distinguishing samples from different sources was 100%, and the accuracy of distinguishing samples from the same source with volatile content higher and lower than 20 wt% was 96.84% and 97.75%. When the number of clusters was three, the best prediction accuracy of Aad was achieved, the R2P was increased from 0.7677 to 0.8178, and the RMSEP was decreased from 2.0733 wt% to 1.8362 wt%. The highest accuracy of predicting Vad was achieved when the number of clusters was two, the R2P was increased from 0.9692 to 0.9786, and the RMSEP was decreased from 1.3066 wt% to 1.0879 wt%. The prediction error of Qbad was minimized when the spectra were clustered into five categories, the R2P was increased from 0.8093 to 0.8306, and the RMSEP was decreased from 0.8892 MJ kg−1 to 0.8382 MJ kg−1. The results demonstrated that HOG has great potential in the classification of spectra, and the influence of the matrix effects can be reduced by unsupervised clustering.