Plastics detection and sorting using hyperspectral sensing and machine learning algorithms

Plastic waste second life management requires effective detection (and sorting if necessary) techniques to tackle the environmental challenge it poses. This research explores the application of hyperspectral imaging in the spectral range 900-1700 nm and machine learning for plastic waste processing. Results are of interest for setting up sorting processes in recycling plants and for aerial remote sensing for plastic litter detection. The experimental investigation considered the most diffused polymers according to plastic converter demand. Both virgin polymers and plastic litter, also collected from the beach, were investigated. Experiments were conducted in an indoor laboratory environment with artificial lighting and an outdoor environment with sunlight. Feature selection and classification were performed using MATLAB® codes employing the minimum-Redundancy Maximum-Relevance method and Principal Component Analysis along with Linear Discriminant Analysis and k-Nearest Neighbors algorithms. For all the cases examined, the combination of minimum-Redundancy Maximum-Relevance and Linear Discriminant Analysis algorithms proved to be the most effective approach in terms of performance and processing time without the need for complicated calibrations. In both the indoor and outdoor scenarios, the Matthew's Correlation Coefficient is higher than 0.94. The application on outdoor data of classifiers learned on indoor ones also appears quite successful (Matthew's Correlation Coefficient greater than 0.90). Realistic plastic litter is properly detected (Matthew's Correlation Coefficient ranging between 0.48 and 0.96 depending on the sample shape).