An Investigation into Methods to Advance Microplastic Retrieval, Detection and Characterisation Using Forensic Science and Machine Learning Approaches

The analysis of microplastics is a quickly growing area of interest as more is learnt about sources and their potential effect on the environment are discovered. As the field of microplastic research is very interdisciplinary, there are various diverse approaches to microplastic research, with different methods being deployed and different priorities being investigated depending on what is being studied. As a result of this, there is a lack of standardisation in the techniques used and even what is considered a microplastic can differ between different studies. This makes it difficult to fully compare different studies and get a clear interpretation of the presence and effects of microplastics. Machine learning has the potential to help streamline and speed up microplastic research by automating the detection of microplastics, allowing more research to be conducted in a shorter period of time.

The aims of this thesis were to investigate currently used methods of extraction and analysis and propose a new standardised approach to the retrieval and characterisation of microplastics. These methods were tested in a laboratory simulation experiment, the results of which will inform a field study conducted on samples taken from the Hudson River. In addition, the use of machine learning to automate the detection of microplastics was investigated.

These aims were met by a variety of methods. Firstly, the currently used methods of extraction and analysis were investigated by completing a comprehensive literature review to identify gaps in knowledge and areas that could be standardised and improved to help streamline microplastic research. The use of Easylift® was investigated as a method of retrieval for microplastics on filter papers to aid in the maximisation of microplastics collected and facilitate subsequent analysis. The tape was tested with two different filter papers (glass and cellulose filter papers) and two different filtration methods (Buchner filter and glass frit) to determine if one method resulted in a greater percentage of microplastics retrieved. The funnel and filter combination found to have the highest retrieval rate with Easylift® was cellulose filter paper with the glass frit, as it had a mean recovery rate of 99.16% with a standard deviation of 0.96% points. The glass filter paper with the Buchner filter combination resulted in the lowest retrieval rate (91.21% with a standard deviation of 2.03%).

From the literature review conducted, it was found that the samples are often searched for using a stereomicroscope. Then, the polymer type is confirmed by either Fourier transform infrared spectroscopy (FTIR) or Raman spectroscopy. The inclusion of Polarised light microscopy (PLM) as a method of searching and further characterisation was employed to determine if potential microplastics were missed by stereomicroscopy, and if the advantages of PLM were beneficial to the characterisation and interpretation process and if certain features meant that microplastics were more likely to be missed. A study was conducted where the sample slide was searched first by stereomicroscopy and then with PLM. The inclusion of PLM found an additional 549 particulates over the 244 sample slides subsection used in this study. The PLM found a significant number of colourless particulates that were missed by the stereomicroscope search, finding 371 (67% of those found by PLM) more colourless particulates compared to the 285 (21% of those found by stereomicroscopy) colourless particulates found with the stereomicroscope. The use of PLM also allowed for the samples to be characterised further than it was possible with the stereomicroscope, including birefringence, sign of elongation, cross-sectional shape and presence of delusterant. These features allowed for contamination to be identified with a greater level of discrimination. Finally, the use of machine learning was investigated as a method of automating the detection process, using 3102 images taken from the ‘Hudson River samples’, which were comprised of water and air samples. The images were split into training, validation, and test set with a 70% (2171), 20% (621), and 10% (310) split. The training was undertaken with YOLOv5, YOLOv7 and YOLOv8 three times for each algorithm, once with 25 epochs, once with 100 epochs and once with the augmented training set with 100 epochs. It was found that YOLOv5 performed the worst out of all algorithms investigated with an F1 score of 0.278; YOLOv5 also had the lowest mAP50 with a mean of 0.191, meaning that it had inaccurate bounding boxes. YOLOv7, with 100 epochs, had the best all-round performance, with the correct identifications for most classes, including microfibres and fragments.

This research has contributed to the knowledge of microplastic analysis by applying a forensic science approach to the methods used in the collection and characterisation of microplastics. A new novel method of collecting microplastics from the surface of filter papers was proposed, and the inclusion of PLM was recommended to provide greater characterisation of microplastics. Certain features were recommended to be recorded, including colour, presence of delusterant and birefringence. These features allowed for potential sources of contamination to be identified and removed and potential sources of microplastics to be investigated. Furthermore, the implementation of machine learning to automate the detection of microplastics would allow the research process to be greatly sped up, allowing more research to be conducted in a shorter period of time, helping to increase knowledge of the problem and potential solutions to be identified.