Holographic imaging and machine learning for microplastic size and shape analysis in water

Microplastics are a growing global concern, particularly in drinking water, due to their potential negative impacts on human health. To effectively monitor, quantify and understand the sources and implications of microplastics in water, it is critical to identify their physical and chemical properties. However, existing laboratory-based methods popularly used for characterising microplastics have several limitations. Using a novel method, this study explored the feasibility of quantifying the physical properties of microplastics in water. Specifically, we utilised a portable holographic camera to record digital holograms of commercial microplastics floating in water. Furthermore, we developed a simple Python algorithm to determine the size of the microplastics from the particle images. This study also evaluated and compared the performance of two deep-learning architectures, MobileNetV2 and ResNet101, in classifying the shapes of the microplastic particles into spherical and hemispherical shapes. Findings from this study demonstrate the capability of the proposed holographic system to rapidly and automatically produce particle images of microplastics while simultaneously measuring their sizes. Performance metrics, including accuracy, precision, recall, F1 score, confusion matrix and training time, showed that MobileNetV2 achieved the best performance despite being a more lightweight model with fewer parameters than ResNet101. Therefore, MobileNetV2 was recommended for classifying the shapes of microplastics from particle images. The time and cost-effectiveness of the proposed digital holographic method make it suitable for large-scale monitoring of microplastics in water. This will be significant in identifying the sources, understanding their behaviour and reducing the associated health risks to humans. • A portable holographic camera recorded holographic images of microplastics floating in water. • The holographic images were reconstructed using Field-Programmable Gate Array Software to produce microplastic particle images. • The sizes of the microplastics were calculated by developing a Python-coded algorithm. • Deep-learning models using MobileNetV2 and ResNet101 were developed to classify the shapes of the microplastics. • MobileNetV2 consistently outperformed ResNet101 in terms of classification accuracy and computational efficiency.