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

Abstract

Microplastics are ubiquitous and appear to be harmful, however, the full extent to which these inflict harm has not been fully elucidated. Analysing environmental sample data is challenging, as the complexity in real data makes both automated and manual analysis either unreliable or time-consuming. To address challenges, we explored a dense feed-forward neural network (DNN) for classifying Fourier transform infrared (FTIR) spectroscopic data. The DNN provides conditional class distributions over 16 microplastic categories given an FTIR spectrum, exceeding number of categories in other works. Our results indicate that this DNN, which is significantly smaller than contemporary models, outperforms other models and even human classification performance. Specifically, while the model broadly reproduces the decisions of human annotators, in cases of disagreement either both were incorrect or the human annotation was incorrect. The errors not being reproduced indicate that the DNN is making informed generalisable decisions. Additionally, this work indicates that there exists an upper limit on metrics measuring performance, where metrics measure agreement between human and model predictions. This work indicates that a small and efficient DNN can making high throughput analysis of difficult FTIR data possible, where predictions match or exceed the reliability typical to low-throughput methods. Environmental samples are collected and separated into size classes, followed by chemical digestion to remove biological matter. Optical microscopy identifies particles with morphologies consistent with microplastics. These particles are chemically validated using Attenuated Total Reflectance-Fourier Transform Infrared Spectroscopy (ATR-FTIR), where degradation processes introduce significant variability in polymer spectra (example shown for Polypropylene), making manual analysis time-consuming and traditional library matching unreliable. To address this challenge, we explored the use of an dense artificial neural network (DNN) that learns to recognize fundamental traits of classes instead of matching to known examples. A dataset compiled from diverse sources environmental data, and was used to train the DNN. The model’s performance was evaluated on holdout data using conventional metrics like ROC curves and confusion matrices. Additionally, we applied dimensional reduction techniques to investigate the model’s structuring of the sample space, providing important insights into its reliability and capability. • Our method trains small dense networks that outperform complex models. • We graph and interpret dense network outputs to understand model behaviour. • We share all code, a trained model, and results in a repository for community use. • Our model finds systematic human errors in public microplastic datasets. • Our model learns to make decisions that do not replicate errors in human decision making

Topics & Concepts