Plastic litter sorting using hyperspectral sensing and Linear Discriminant Analysis

Plastic litter causes damage to the environment. Thus, it is mandatory to provide mechanical recycling plants with fast, accurate, and reliable tools and equipment specifically addressed to separate and recover single polymer streams, eliminating polluting elements (i.e., other polymers or other materials) present in the feed. Technologies for plastic waste separation are expected to be cost-effective and suitable to guarantee high quality of the products, mainly in terms of purity. This should guarantee the production of competitive secondary. The adoption of hyperspectral sensors working in the wavelength range 1000-1700 nm to perform sorting requires the utilization and the implementation of online analytical tools and robotic units to perform the separation. Materials in fact, have to be first detected, then identified and topologically assessed in the stream; after these steps automated devices realize the sorting. Many different polymers can be separated by near infrared sensors as they are characterized by different spectral signature in such wavelength range (i.e., PP, PE, PVC, PET, PS, etc.). This paper will describe the experimental campaign carried out to characterize plastic samples belonging to the most diffuse polymer typologies both in controlled light conditions and with natural light. Our detection methodology is based on linear classifiers, obtained by Linear Discriminant Analysis applied on a subset of spectral bands selected by minimum-redundancy-maximum-relevance criterion or by Principal Component Analysis. The classifier is obtained by optimization on a wide set of manually labeled examples. Results confirm how sorting with hyperspectral sensors may guarantee the achievement of the required standard for recycled plastics reuse.