Adaptive nonlinear mapping for feature extraction and fusion in mineral classification based on laser-induced breakdown spectroscopy (LIBS) and Raman spectroscopy

Abstract

Efficient extraction of key features from LIBS (Laser-Induced Breakdown Spectroscopy) and Raman spectra, while simultaneously eliminating redundant background information, is crucial for enhancing the applicability of spectral data. To address this, we propose an adaptive nonlinear mapping method based on multi-function optimization, tailored for feature extraction from different types of mineral spectra. Specifically, the method first computes the mean and variance of each spectrum, and then evaluates the deviation between key features and the mean-variance using feature importance analysis. Based on these deviation parameters, the method adaptively selects the most suitable mapping function. The algorithm incorporates four built-in mapping functions that amplify significant information while compressing irrelevant features. These functions map spectral attributes such as intensity, width, area, and position, and convert the data into a two-dimensional image representation. Subsequently, a convolutional neural network (CNN)-based image classification approach is employed to accurately classify the extracted information. Using this feature extraction and classification framework, we achieved a classification accuracy of 99.70% for the fusion of LIBS and Raman spectral data, significantly outperforming conventional methods such as principal component analysis (94.55%) and random forest-based feature importance evaluation (92.12%). The results demonstrate that the proposed method provides a substantial advantage in processing complex spectral data, enhancing spectral analysis performance, and offering strong support for broader applications of LIBS and Raman spectroscopy.