From microplastics to pixels: testing the robustness of two machine learning approaches for automated, Nile red-based marine microplastic identification
Authors:
- 24 October 2024
- journal article
- research article
- Published by Springer Nature in Environmental Science and Pollution Research
- Vol. 31 (52) , 61860-61875
Abstract
Despite the urgent need for accurate and robust observations of microplastics in the marine environment to assess current and future environmental risks, existing procedures remain labour-intensive, especially for smaller-sized microplastics. In addition to this, microplastic analysis faces challenges due to environmental weathering, impacting the reliability of research relying on pristine plastics. This study addresses these knowledge gaps by testing the robustness of two automated analysis techniques which combine machine learning algorithms with fluorescent colouration of Nile red (NR)-stained particles. Heterogeneously shaped uncoloured microplastics of various polymers—polyethylene (PE), polyethylene terephthalate (PET), polypropylene (PP), polystyrene (PS), and polyvinyl chloride (PVC)—ranging from 100 to 1000 µm in size and weathered under semi-controlled surface and deep-sea conditions, were stained with NR and imaged using fluorescence stereomicroscopy. This study assessed and compared the accuracy of decision tree (DT) and random forest (RF) models in detecting and identifying these weathered plastics. Additionally, their analysis time and model complexity were evaluated, as well as the lower size limit (2–4 µm) and the interoperability of the approach. Decision tree and RF models were comparably accurate in detecting and identifying pristine plastic polymers (both > 90%). For the detection of weathered microplastics, both yielded sufficiently high accuracies (> 77%), although only RF models were reliable for polymer identification (> 70%), except for PET particles. The RF models showed an accuracy > 90% for particle predictions based on 12–30 pixels, which translated to microplastics sized < 10 µm. Although the RF classifier did not produce consistent results across different labs, the inherent flexibility of the method allows for its swift adaptation and optimisation, ensuring the possibility to fine-tune the method to specific research goals through customised datasets, thereby strengthening its robustness. The developed method is particularly relevant due to its ability to accurately analyse microplastics weathered under various marine conditions, as well as ecotoxicologically relevant microplastic sizes, making it highly applicable to real-world environmental samples.Keywords
Funding Information
- Belgian Federal Science Policy Office (B2/20E/P1/Andromeda)
- Agence Nationale de la Recherche (ANR-19–383 JOCE-0002–01)
- Norges Forskningsråd (311313/E40)
This publication has 58 references indexed in Scilit:
- Rapid detection of nanoplastics and small microplastics by Nile-Red staining and flow cytometryEnvironmental Chemistry Letters, 2022
- Recent advances in the application of machine learning methods to improve identification of the microplastics in environmentChemosphere, 2022
- Aging of microplastics increases their adsorption affinity towards organic contaminantsChemosphere, 2022
- Effect of weathering on environmental behavior of microplastics: Properties, sorption and potential risksChemosphere, 2020
- Accumulation of different shapes of microplastics initiates intestinal injury and gut microbiota dysbiosis in the gut of zebrafishChemosphere, 2019
- Marine microplastic: Preparation of relevant test materials for laboratory assessment of ecosystem impactsChemosphere, 2018
- Nanoplastics in the Aquatic Environment2018
- Microplastic pollution in deep-sea sedimentsEnvironmental Pollution, 2013
- Biological degradation of plastics: A comprehensive reviewBiotechnology Advances, 2008
- Bagging predictorsMachine Learning, 1996