A novel feature screening algorithm for low-resolution LIBS spectrum elemental quantification

This study presents an innovative approach that integrates Laser-Induced Breakdown Spectroscopy (LIBS) with chemometrics for the quantitative analysis of Si, Ca, Al, and Mg in geological samples. Given the spectral redundancy in low-resolution LIBS devices, the study employs pre-processing techniques, such as AirPLS, Wavelet Transform (WT), and normalization to mitigate spectral noise. Enhanced feature threshold searching is achieved by incorporating SHapley Additive exPlanations (SHAP) and LightGBM into the Boruta algorithm, substantially improving quantitative analysis models based on Support Vector Regression (SVR) and Partial Least Squares Regression (PLSR). The modified Boruta-SVR model demonstrated remarkable robustness, with R² values of 0.9862, 0.9873, 0.9882, and 0.9916, and RMSE values of 0.8099, 0.324, 0.1378, and 0.2382, respectively, for Si, Ca, Al, and Mg. The results confirm that the Boruta-based feature selection method, when applied to low-resolution LIBS spectra, outperforms traditional methods, capturing unique sample features under mixed spectral peak conditions, thereby enhancing the robustness of quantitative analysis models.