Robust model development for HSI-based characterization of post-consumer plastics

The need for effective recycling solutions has led to the investigation of advanced technologies for material characterization. This study explores chemometric methods including artificial neural networks for characterizing post-consumer plastics using near-infrared hyperspectral data. We analyze different classification methods in terms of overall performance, their ability to generalize onto another dataset as well as outlier detection. The investigated methods are Linear Discriminant Analysis (LDA), Artificial Neural Networks (ANN), Spectral Angle Mapper (SAM), and Partial Least Squares Discriminant Analysis (PLS-DA). We show that the performance on the main dataset is similar for all methods, with an accuracy of approximately 98%, whereas the models varied in terms of generalization and outlier detection. SAM performed best in terms of outlier detection, outscoring other methods by 8 percentage points.