Convolutional neural network for soil microplastic contamination screening using infrared spectroscopy.

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Microplastics, synthetic polymers <5mm, are pervasive environmental pollutants.
• No established limits for soil microplastic concentration necessitate baseline data.
• Visible-near-infrared (vis-NIR) spectroscopy shows potential for microplastic detection in soil.

Purpose of the Study:

• To develop a rapid screening method for microplastic contamination in soil.
• To utilize a convolutional neural network (CNN) for classifying soil samples based on microplastic concentration.
• To investigate the performance of CNN models in classifying spectral data for microplastic detection.

Main Methods:

• Soil samples were collected from industrial areas around metropolitan Sydney to establish a baseline.
• Visible-near-infrared (vis-NIR) spectra were obtained from soil samples.
• A CNN model was trained to classify soil samples into different degrees of microplastic contamination.

Main Results:

• The CNN model achieved higher accuracy in classifying uncontaminated samples than contaminated ones.
• Increasing the number of contamination classes improved classification accuracy for highly contaminated samples.
• Transfer learning enhanced CNN performance at the extremes of contamination levels but not for intermediate classes.

Conclusions:

• CNN-based classification of vis-NIR spectral data offers a feasible approach for rapid soil microplastic screening.
• The model's accuracy varies with contamination levels, suggesting further refinement for intermediate concentrations.
• This method provides a foundation for developing standardized protocols for soil microplastic assessment.