Multi-order moment fusion of laser-induced breakdown spectroscopy (LIBS) and Raman spectroscopy for mineral classification

Traditional laser-induced breakdown spectroscopy (LIBS) methods rely primarily on spectral intensity, overlooking the rich statistical information embedded in higher-order moment features such as mean, variance, skewness, and kurtosis. This limitation hinders both the accuracy and generalizability of LIBS in analyzing complex mineral samples. To address this challenge, we introduce a novel approach that extracts and integrates multi-order moment features to enhance spectral representation and improve classification performance. Specifically, we compute higher-order statistical moments from LIBS spectra and standardize them using Z-score normalization to eliminate dimensional bias. A random forest model is then used to assign feature importance weights, guiding the feature fusion process. The resulting LIBS features are further fused at the feature level with Raman spectral data, allowing for multi-parameter representation of each sample. A neural network classifier is subsequently employed to evaluate the model's performance. Experimental results demonstrate that our fusion-based method achieves classification accuracy, precision, and specificity exceeding 99 %, significantly outperforming conventional LIBS-based approaches, which attain only 83.11 % accuracy. These findings highlight the effectiveness of multi-order moment fusion in enhancing spectral analysis of complex samples, and demonstrate its broad potential for applications in mineral identification and beyond.