Spectral Classification of Large-Scale Blended (Micro)Plastics Using FT-IR Raw Spectra and Image-Based Machine...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Microplastics (MPs) are pervasive emerging pollutants requiring effective identification methods.
• Current MP identification often relies on small datasets and library searches, limiting scalability.
• Accurate classification is crucial for monitoring and managing microplastic pollution.

Purpose of the Study:

• To develop and compare machine learning classifiers for microplastic identification.
• To evaluate model performance using large-scale blended plastic datasets.
• To provide an open-source tool for rapid and accurate microplastic spectroscopic analysis.

Main Methods:

• Four machine learning classifiers were developed: 1D CNN, 2D CNN, Decision Tree, and Random Forest (RF).
• Models were trained and tested on two large-scale blended plastic datasets.
• Fourier transform infrared (FTIR) spectroscopy raw spectral data and spectral images were used as inputs.

Main Results:

• The 1D CNN achieved the highest overall accuracy (96.43% on small, 97.44% on large datasets).
• 1D CNN demonstrated superior performance in predicting environmental microplastic samples.
• Random Forest (RF) showed robustness with limited spectral data; 2D CNN also suitable for limited data scenarios.

Conclusions:

• 1D CNN is the most effective classifier for microplastic identification with sufficient spectral data.
• RF and 2D CNNs offer viable alternatives when spectral data is limited.
• The developed open-source tool enhances the speed and accuracy of microplastic analysis.