Classification of Impurity Levels in Contaminated Sugarcane Leaf Pellets using NIR Spectroscopy

Abstract

This research investigates the feasibility of classifying impurity levels in sugarcane leaf pellets. Samples were prepared by mixing sugarcane leaves with varying proportions of sand. NIR spectroscopy combined with linear discriminant analysis (LDA) was employed to categorize samples across discrete contamination tiers (5% to 25% sand content). Because an absolute 0% clean baseline was not evaluated, the framework is interpreted strictly as a system for grading impurity severity levels. Experimental results revealed that centering and 1^st derivative preprocessing yielded the highest accuracy. These techniques effectively mitigated interference from light scattering and baseline shifts, while enhancing specific spectral features related to biomass chemical constituents, such as cellulose, lignin, and hemicellulose. The LDA model, optimized with preprocessing, achieved classification accuracy exceeding 97%, with minimal errors observed in samples containing low to moderate sand content. However, accuracy slightly declined when sand content exceeded 20% due to reflective interference from silica (SiO₂). The findings confirm that selecting appropriate preprocessing methods is a critical factor for model performance. These laboratory-scale results demonstrate the strong potential of the method, serving as a foundational step toward developing future on-line monitoring pipelines for real-time soil contamination tracking.

This research investigates the feasibility of classifying impurity levels in sugarcane leaf pellets. Samples were prepared by mixing sugarcane leaves with varying proportions of sand. NIR spectroscopy combined with linear discriminant analysis (LDA) was employed to categorize samples across discrete contamination tiers (5% to 25% sand content). Because an absolute 0% clean baseline was not evaluated, the framework is interpreted strictly as a system for grading impurity severity levels. Experimental results revealed that centering and 1^st derivative preprocessing yielded the highest accuracy. These techniques effectively mitigated interference from light scattering and baseline shifts, while enhancing specific spectral features related to biomass chemical constituents, such as cellulose, lignin, and hemicellulose. The LDA model, optimized with preprocessing, achieved classification accuracy exceeding 97%, with minimal errors observed in samples containing low to moderate sand content. However, accuracy slightly declined when sand content exceeded 20% due to reflective interference from silica (SiO₂). The findings confirm that selecting appropriate preprocessing methods is a critical factor for model performance. These laboratory-scale results demonstrate the strong potential of the method, serving as a foundational step toward developing future on-line monitoring pipelines for real-time soil contamination tracking.