Adaptable microplastic classification using similarity learning on µFTIR spectra collected from µFTIR focal plane...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Microplastic identification is crucial for environmental monitoring.
• Deep learning models for microplastic classification face challenges like large dataset needs and overfitting.
• Current methods require retraining for new microplastic types or significantly different data.

Purpose of the Study:

• To explore a similarity learning approach for training deep learning models for microplastic classification.
• To address limitations of traditional deep learning methods in microplastic identification.
• To improve accuracy and adaptability in classifying microplastics from µFTIR spectra.

Main Methods:

• A one-dimensional convolutional neural network (CNN) was trained using similarity learning on µFTIR spectra from 45 microplastic samples (11 compositions).
• The similarity learning CNN was compared against cross-entropy trained CNNs and classical machine learning algorithms.
• Performance was evaluated on both pristine and 'noisy' datasets (microplastics on filters with background material).

Main Results:

• The similarity learning CNN achieved the highest accuracies, with an F1-score up to 0.973.
• Even on noisy datasets, the similarity learning CNN maintained high accuracy (up to 0.905 F1-score).
• Similarity learning enabled detection of microplastic classes not included in the initial training set.

Conclusions:

• Similarity learning offers a robust and accurate method for microplastic classification using µFTIR spectra.
• This approach overcomes challenges of diverse polymer compositions, limited data, and background noise.
• It provides a more adaptable and efficient deep learning solution for microplastic identification.