Comparison of data augmentation and classification algorithms based on plastic spectroscopy.

Area of Science:

• Environmental Science
• Analytical Chemistry
• Computer Science

Background:

• Plastic waste management is a critical global environmental issue.
• Spectroscopy and deep learning offer effective plastic classification methods.
• Limited data and algorithm comparisons hinder progress.

Purpose of the Study:

• To propose a plastic spectroscopy generation model for data augmentation.
• To systematically analyze and compare various classification algorithms.
• To improve the accuracy and efficiency of plastic waste classification.

Main Methods:

• Preprocessing spectral data using cubic interpolation, normalization, S-G filtering, linear detrending, and SNV.
• Developing a C-GAN based model for multi-class plastic spectroscopy generation.
• Comparing machine learning (SVM, RF) and deep learning (GoogleNet, ResNet) algorithms.

Main Results:

• Preprocessing improved classification accuracy, visualized via PCA.
• The C-GAN model generated consistent and valid plastic spectra.
• Data augmentation increased classification accuracy by at least 3%.
• 1D-ResNet achieved peak accuracy (0.991) for FTIR data with twofold augmentation.
• 1D input models generally outperformed 2D models.
• Deep learning excelled on large datasets, while traditional ML showed stability on small datasets.

Conclusions:

• Data augmentation via spectroscopy generation models is crucial for improving plastic classification.
• 1D-ResNet demonstrates superior performance in identifying spectral features for accurate classification.
• Both traditional ML and deep learning have roles depending on dataset size.