Joint Tensor-Train Decomposition And Stacking Ensembles For Enhanced Classification Of Microplastics Via Raman Spectroscopy

Abstract

The growing problem of ecological pollution by micro-plastics requires fast and accurate identification techniques. The current Raman spectroscopy technique is severely limited by surface enhanced baseline interferences, peak overlaps, and instrumental noise. State-of-the-art deep learning architectures like 1D Convolutional Neural Networks (CNNs) are designed to address these spectroscopic challenges, but are prohibitively expensive and vulnerable to over-fitting owing to the small sample sizes that are usually encountered in the environmental sciences. To address these problems, we introduce the Joint Tensor-Train Fractional Scattering Network (JT-FSN), which is a mathematically intuitive feature extraction network alternative to opaque deep neural networks. We first build a pseudo-multimodal tensor from the raw spectral signal and its first derivative, then deterministically separate the low frequency macro baseline using Tensor-Train (TT) decomposition, and extract high frequency, sub-pixel chemical variance using a Fractional Wavelet scattering cascade. Representations are classified with a heterogeneous Stacking Ensemble of SVM, Random Forest, XGBoost and Gradient Boosting machines, after compressing the representations into an 4,364 dimensional invariant feature space via Principal Component Analysis (PCA). Empirical accuracy of the proposed framework is 79.00% on a highly augmented Microplastics Raman Spectra dataset, with perfect Recall (100%) for the various polymer blends (PE_PTFES1, PE_PTFES2, etc.). The framework drastically reduces the computational load and is a highly efficient and deterministic approach for chemometric classification in high interference environments, such as microplastics.