[A novel approach to NIR spectral quantitative analysis: semi-supervised least-squares support vector regression...

Area of Science:

• Near-infrared spectroscopy
• Chemometrics
• Machine learning

Context:

• Quantitative analysis of chemical compositions using near-infrared (NIR) spectral data is limited by the availability of samples with accurately measured chemical values.
• Traditional mathematical models often exclude samples lacking chemical data, potentially reducing model performance.
• Developing robust models that leverage all available data, including unlabeled samples, is crucial for advancing NIR spectral quantitative analysis.

Purpose:

• To propose and evaluate a novel semi-supervised least squares support vector regression (S2 LS-SVR) model for near-infrared spectral quantitative analysis.
• To enhance the utilization of both labeled (with chemical values) and unlabeled (without chemical values) samples in chemometric modeling.
• To improve the accuracy and efficiency of predicting chemical compositions in complex samples.

Summary:

• A semi-supervised LS-SVR (S2 LS-SVR) model was developed, building upon the LS-SVR framework to incorporate samples lacking chemical values.
• The S2 LS-SVR model training is computationally efficient, equivalent to solving a linear system.
• Models were constructed for predicting total sugar, reducing sugar, total nitrogen, and nicotine content in flue-cured tobacco using PLS regression, LS-SVR, and the proposed S2 LS-SVR.

Impact:

• The S2 LS-SVR model demonstrated superior performance compared to PLS regression and standard LS-SVR, achieving average relative errors as low as 6.11% and high correlation coefficients (up to 0.9741) for tobacco composition analysis.
• This approach significantly improves the feasibility and efficiency of quantitative analysis in scenarios with limited labeled data.
• The findings highlight the potential of semi-supervised learning in chemometrics for more accurate and comprehensive spectral data analysis.