Hyperspectral imaging applied to the waste recycling sector

The need to develop and deploy increasingly effective, fast and robust sensing techniques capable of detecting, characterising and sorting solid waste products and materials represents one of the most challenging aspects in the industrial recycling sector. Most waste is of low value, so innovative technologies must be costeffective and, at the same time, achieve high performance in terms of materials identification, both for sorting and quality control in recycling plants, so as to end up with high-quality secondary raw materials that are competitive in the market with the corresponding primary raw materials. In this perspective, the use of hyperspectral imaging (HSI) technology to carry out products/materials characterisation, through fast and reliable handling/processing, is becoming more and more important. In this article, some HSI-based applications in the waste recycling sector, originally developed by the authors, are presented and discussed. All the procedures have been designed, implemented and set up with the aim of providing sensing/inspection tools able to perform non-invasive, contactless and real-time analyses at both laboratory and/or industrial scale. Hyperspectral imaging Hyperspectral imaging is an innovative technique that combines the properties of digital imaging with those of spectroscopy. 1 Using this approach, it is possible to detect the spectral signature of each pixel of the acquired image in different wavelength regions (visible, near infrared, short-wave infrared etc.) according to the characteristics of the selected sensing device. A hyperspectral image can thus be considered as a three-dimensional dataset with two spatial dimensions and one spectral dimension, the so-called “hypercube”. HSI can be considered one of the best and most powerful nondestructive technologies for accurate and detailed information extraction from the acquired images, with a high level of flexibility. The large amount of spectral information collected by HSI from the sample surfaces must be processed in order to extract the information of interest. Furthermore, as a preliminary step in any inspection or quality control logic development, hyperspectral libraries of reference spectra, to be utilised for unknown sample recognition, must be built. To reach the previously mentioned goals, algorithms and procedures for spectral data pre-processing, exploration and classification are usually implemented through chemometric strategies. Different pre-preprocessing algorithms can be applied to hyperspectral data, finalised to linearise relationships among variables and remove external sources of variation that are not of particular interest for the analysis. Principal Component Analysis (PCA) is applied for exploratory purposes, providing an overview of the complex multivariate data.2 PCA decomposes spectral data into several principal components (PCs), linear combinations of the original data, embedding the spectral variations of each collected spectral data set. The first few PCs, resulting from PCA, are used to analyse the common features among samples: in fact, samples characterised by similar spectral signatures tend to aggregate in the score plot as a cluster. Finally, the recognition of different products and/or materials is obtained utilising classification methods, such as Partial Least-Squares Discriminant Analysis (PLS-DA). PLS-DA is a supervised classification technique requiring prior knowledge of the data and allowing the classification of samples into predefined groups.3 In order to do that, starting from reference samples, a discriminant function is built and then this is later applied to classify samples belonging to an unknown set. Once the model is built, it can be applied to validation images. An interesting and powerful classification method is hierarchical modelling. Adopting this kind of classification logic,4 objects are divided into subsets and then they are split again into further subsets, until each of them contains only a single object. During each step, objects that are different from the others are selected, isolated and compared through successive PLS-DA classification models.