Improved data augmentation and feature selection strategies for enhancing LIBS analysis in biomass higher heating value prediction

Rapid and accurate prediction of biomass higher heating value (HHV) is essential for optimizing energy utilization. Laser-induced breakdown spectroscopy (LIBS) combined with machine learning enables rapid quantitative analysis of HHV, but its predictive accuracy often suffers from insufficient sample size and spectral redundancy. To address these issues, this study proposes an optimized random forest (RF) framework that integrates data augmentation and a two-stage feature selection. First, LIBS spectra of biomass samples were acquired. Then, a gradient boosting decision tree-guided synthetic minority oversampling technique (GB-SMOTE) was developed to augment the spectral data, thereby improving the quality of the training samples. Subsequently, a two-stage feature selection approach was implemented to eliminate spectral redundancy. In the first stage, an attention gradient (AG) method was used for preliminary selection, reducing the features from 5784 to 1735. In the second stage, an improved subtraction average-based optimizer (ISABO) was employed to further refine the feature subset to 122. Finally, the RF model established with the selected features achieved excellent prediction performance, with a coefficient of determination for prediction ( R P 2 ) of 0.9441, a root mean square error for prediction (RMSE P ) of 0.3555 MJ/kg, and a mean relative error for prediction (MRE P ) of 1.63%. These findings demonstrate that the proposed strategies can effectively overcome the limitations of small sample sizes and spectral redundancy in LIBS quantitative analysis, providing a reliable approach for rapid and accurate assessment of biomass fuel quality.