Quantification of individual solid wastes from a mixture using hyperspectral imaging and machine learning

The quantification of components in mixed solid wastes is fundamental to waste-quality assessment and to informed decision-making for downstream utilization. Here, we develop a non-destructive quantitative framework that couples near-infrared hyperspectral imaging with machine learning for a five-component system comprising HDPE, PP, PS, pine shavings, and corn stalks. Standard normal variate (SNV) preprocessing and successive projections algorithm (SPA) feature selection were used to construct feature spectral inputs, and five machine learning models (PLSR, XGBoost, SVR, RF, and 1D-CNN) were benchmarked using sample-level grouped cross-validation. Results demonstrated that the 1D-CNN delivered consistently strong performance across all components (R² = 0.977-0.989, MAE = 1.966-2.662 wt%). Guided by analyses of component crosstalk, agreement bias assessment, and error patterns, we further identified confusing component pairs and low-concentration regimes, and introduced targeted augmentation using decoupling and low-content datasets to reinforce the model. Shapley additive explanations (SHAP) analysis established interpretable links between key bands and component competition. Furthermore, the reinforced model demonstrated excellent stability in independent tests across days and batches. In leave-combination-out extrapolation test using five-component formulations, the model maintained high accuracy, with R² ranging from 0.898 to 0.990 and MAE from 1.036 to 3.119 wt%, demonstrating transferability to unknown recipes and practical engineering potential. This study provides an interpretable and scalable route to rapid quantitative analysis for precision blending of mixed solid waste.