Machine learning outperforms humans in microplastic characterization and reveals human labelling errors in FTIR data.

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Microplastics are widespread environmental contaminants with largely unknown harmful effects.
• Analyzing complex environmental samples for microplastics is challenging, often requiring time-consuming or unreliable methods.
• Fourier Transform Infrared (FTIR) spectroscopy is a key technique for microplastic identification.

Purpose of the Study:

• To develop an efficient and accurate method for classifying microplastic types using FTIR data.
• To overcome the limitations of current manual and automated analysis of environmental samples.
• To explore the potential of deep learning models for high-throughput microplastic identification.

Main Methods:

• A dense feed-forward neural network (DNN) was designed and trained to classify FTIR spectra.
• The DNN was trained to identify 16 distinct microplastic categories, exceeding the scope of previous studies.
• Model performance was benchmarked against human classification and other contemporary models.

Main Results:

• The developed DNN achieved high accuracy in classifying microplastic categories from FTIR data.
• The DNN model demonstrated superior performance compared to existing models and human annotators.
• Analysis revealed that the DNN's correct classifications in disagreement cases indicate generalizable decision-making.

Conclusions:

• A small, efficient DNN can enable high-throughput analysis of challenging FTIR spectroscopic data.
• The DNN's predictions match or exceed the reliability of traditional low-throughput methods.
• This approach offers a scalable solution for microplastic identification in environmental monitoring.