(713b) Geometrical Sampling, Gaussian Mixture Modeling, and Spectral Feature Extraction for Classifying Municipal Solid Waste Using Hyperspectral Data

Accurate classification of solid waste (MSW) mixtures into distinct material subcategories is critical for recycling or waste valorization pathways, such as gasification [1]. While various industrial processes and configurations exist for performing this task, they often demand significant human oversight or involve complex and costly machinery thus limiting their wide-spread applicability. Furthermore, the type and accuracy of classification is often designed/adapted based on the target usage of the material subcategories. Hyperspectral imaging using near-infrared (NIR) and mid-infrared (MIR) cameras, particularly when fused with machine learning algorithms can be a potential solution for non-contact material identification [2]. However, several challenges remain toward construction of effective sortation models including production of high volume data sets from relevant, realistic, and controlled materials for model training, processing large volume of data within a few seconds, pre-processing of data to remove noise and outliers including background, identification of characteristic feature(s) in the data for classification, selection of representative data for classification, and training and calibration of the classification approach while still preserving essential spectral features [3].

Although several classification approaches have been highly successful for hyperspectral image classification, many of those rely on exhaustive pixel-wise analyses and extensive feature sets, leading to computational bottlenecks, especially for the spectral data, making them incapable for real-time classification for MSW on the belt. The high dimensionality not only adversely affects training duration and memory requirements but also can lead to overfitting. Furthermore, subtle variations in properties like thickness or color frequently necessitate intensive recalibration. In response to these limitations, this work develops an algorithm for classification that seeks to maximize data reduction while minimizing loss of information. A geometrical sampling approach that adaptively identifies the most relevant spatial-spectral subsets is developed by using computer vision [4]. An approach is developed based on spectral characteristics for automatic removal of background spectra from the data. While various clustering algorithms are developed and applied, the Gaussian mixture model-based clustering approach with optimal covariance structure type (selected out of types such as tied, diagonal, spherical, full, etc.) is found to yield the best results as measured in terms of clustering metrics such as Calinski-Harabasz, Davies-Bouldin, and Silhouette scores. A variance-based information measure is developed for representative feature set. The processed feature set are then feed into an advanced, convolutional neural network (CNN) enabling rapid and accurate classification in real time. The efficacy of the proposed methodology is demonstrated by using data from two NIR cameras (0.95 µm – 1.7 µm and 1.62 µm – 2.19 µm) and one MIR camera (2.7 - 5.3 µm), where it is desired to classify 22 distinct MSW material subcategories. It is observed that the algorithm leads to large reduction in data volume for training while still yielding more than 95% classification accuracy for all sub-categories.

References

[1] L. Milios, L. Holm Christensen, D. McKinnon, C. Christensen, M. K. Rasch, and M. Hallstrøm Eriksen, “Plastic recycling in the Nordics: A value chain market analysis,” Waste Manag., vol. 76, pp. 180–189, Jun. 2018, doi: 10.1016/j.wasman.2018.03.034.

[2] S. Jiang, Z. Xu, M. Kamran, S. Zinchik, S. Paheding, A. G. McDonald, E. Bar-Ziv, V. M. Zavala, “Using ATR-FTIR spectra and convolutional neural networks for characterizing mixed plastic waste,” Comput. Chem. Eng., vol. 155, p. 107547, Dec. 2021, doi: 10.1016/j.compchemeng.2021.107547.

[3] J. M. Amigo, I. Martí, and A. Gowen, “Hyperspectral Imaging and Chemometrics,” in Data Handling in Science and Technology, vol. 28, Elsevier, 2013, pp. 343–370. doi: 10.1016/B978-0-444-59528-7.00009-0.

[4] J. Xu and P. Mishra, “Combining deep learning with chemometrics when it is really needed: A case of real time object detection and spectral model application for spectral image processing,” Anal. Chim. Acta, vol. 1202, p. 339668, Apr. 2022, doi: 10.1016/j.aca.2022.339668.