Near-infrared spectroscopic prediction of gasoline olefin content: A systematic approach using continuous region feature selection and region-sensitive ensemble learning

Abstract

This study addresses the challenges of high-dimensional collinearity and regional information heterogeneity in near-infrared spectroscopy for gasoline olefin content prediction by proposing a systematic optimization approach combining a Continuous Region Utilizing Integrated Spectral Evaluation for Near-Infrared (CRUISE-NIR) algorithm with a Region-Sensitive Adaptive Ensemble Learning (RAEL) framework. The CRUISE-NIR algorithm shifts spectral analysis from a “point” to a “region” perspective, fully considering the physical correlation of adjacent wavelengths and chemical prior knowledge, reducing 4443 original variables to 16 key features. Meanwhile, the RAEL framework dynamically adjusts prediction weights according to sample performance characteristics in different spectral regions, achieving sample-specific precision prediction. Experimental results demonstrate that the proposed method achieves a root mean square error (RMSE) of 0.2795 and a coefficient of determination (R²) of 0.9646 on the test set, significantly outperforming traditional methods in prediction accuracy and fitting capability.Furthermore, the robustness of the framework was successfully validated on heterogeneous matrices including SWRI Diesel, IDRC Tablets, and Soil, demonstrating robust generalizability across diverse liquid and solid physical states. Experimental results indicate that prioritizing high-quality feature selection over variable quantity significantly enhances model performance. The proposed systematic framework demonstrates robust analytical capabilities for high-dimensional spectral data across diverse and complex molecular systems.