A near-infrared spectroscopy method for detecting corn starch content based on UVE-LightGBM feature selection

Abstract

This study introduces a feature selection methodology that uses Near-infrared spectroscopy (NIRS), combining Uniform Variable Elimination (UVE) with the LightGBM algorithm in Gradient Boosting Machines (GBMs) for the swift, non-destructive evaluation of maize starch content. The research initially employed various preprocessing methods on the original spectral data. It assessed their effectiveness by Partial Least Squares Regression (PLSR). The findings demonstrated that first derivative (1D) preprocessing was the most efficacious, yielding an R²C of 0.9994, RMSEC of 0.0187, R²P of 0.9375, RMSEP of 0.2498, and RPD of 3.9985. UVE was subsequently utilized to identify essential wavelengths, while LightGBM further optimized the selection, markedly enhancing modeling efficiency and precision. Multiple feature selection techniques were employed for the comparison of regression models, including Ridge Regression (RR), Gaussian Process Regression (GPR), Multilayer Perceptron Regression (MLPR), and Random Forest (RF). The findings indicated that UVE-LightGBM modeling had a superior coefficient of determination and reduced root mean square error, with an R²P of 0.9972 and an RMSEP of 0.0470. The physical and chemical significance of specific wavelengths was clarified by SHapley Additive exPlanation (SHAP), validating their role in enhancing the model's predicted accuracy and interpretability.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.