Explainable Detection of Microplastics Using Transformer Neural Networks

Abstract

Microplastics are environmental contaminants that put marine and aquatic ecosystems at serious risk. Monitoring microplastics is necessary to understand the level of microplastic pollution in our environment. However, the lack of a standard protocol for quantifying and classifying microplastics causes problems in the reliability and comparability of results. Previous literature has employed deep learning models to classify and quantify microplastic polymers with great success, but the ability of these models to classify microplastics from new domains is unanswered. This paper presents an innovative approach to microplastic classification that employs a deep learning approach using a transformer neural network. Our specific contributions are: (1) A novel way to pre-process FTIR spectral data to dramatically increase classification accuracy. (2) Developed a transformer neural network for classifying microplastic polymer FTIR spectra. With the inclusion of a wider range of data, future deep learning approaches will improve the classification and quantification of microplastic polymers, subsequently reducing the costs and labour involved.