Classification of polypropylene types using NIR hyperspectral imaging and machine learning for improved recycling

During mechanical recycling of polypropylene (PP), different PP types and grades are inevitably mixed, leading to the loss of their tailored material properties. To enable the preservation of those properties, improved sorting technologies are required. This study investigates the separability of PP types (homopolymer, block copolymer and random copolymer) using near-infrared hyperspectral imaging (NIR HSI) – a state-of-the-art method for automatic plastics sorting. A proof of concept was first carried out using 45 different virgin PP samples. To better reflect the variability in real PP waste streams, the dataset was subsequently extended by 75 waste samples whose PP types were independently identified using Differential Scanning Calorimetry (DSC) and Fourier Transform Infrared Spectroscopy (FTIR) measurements. For model development and initial evaluation, 20 representative NIR spectra were selected from each sample and used to assess several classification algorithms (PLS-DA, LD, Random Forest, ExtraTrees, and SVM). On the virgin dataset, the models achieved excellent classification performance, reaching a sample-level F1 score of 1.0. The increased variability introduced by including waste samples reduced classification performance. However, optimized models in combination with hyperparameter tuning and genetic-algorithm-based feature selection achieved promising performance (F1 score > 0.89) on an independent waste sample test set. Notably, the best-performing SVM model correctly classified all test samples when applied to the hyperspectral images using pixel-wise classification and majority voting. These results demonstrate the potential of NIR-HSI classification approaches for distinguishing PP types in waste streams, providing a basis for improved sorting and higher-quality PP recyclates.