Replacing human judgment in fluorescence-based microplastic detection with machine learning: toward harmonization with spectroscopic methods

• Identification of an optimal machine learning model for microplastic detection • “NRmachine”, an automatic microplastic identification application, was developed • Detection probability was 50% for 40-µm particles and 90% for 100-µm particles • Machine learning can replace manual identification in microplastic detection • The accuracy of the proposed method was compared with FTIR Fluorescence based microplastic identification, despite being widely used as a rapid and cost-effective approach, relies heavily on analyst experience and therefore has intrinsic limitations in objectivity and reproducibility, particularly when compared with spectroscopic reference methods such as Fourier transform infrared spectroscopy (FTIR). If the identification workflow can break its dependence on the operator’s empirical judgment and achieve consistency with chemically validated reference data, the reproducibility and reliability of monitoring results can be substantially improved. This study presents the systematic evaluation of whether machine learning (ML) based fluorescence identification of microplastics can be harmonized with FTIR based identification under actual analytical conditions using wildlife samples, which represent highly complex analytical matrices. To identify the most suitable model, six machine learning models were trained using color information from 33,876 pixels of Nile red–stained microplastic fluorescence images obtained from wild marine species. An automatic microplastic detection application (NRmachine) was then developed using the best performance model, and a new microplastic detection workflow without human judgment was constructed. The proposed approach achieved the same level of accuracy as human-based identification (F1 score = 0.74), indicating that manual work can be replaced by a machine learning model. The error of the proposed approach to FTIR based identification of microplastic counts was stable within ±15 particles per run. Nineteen percent of microplastics in the evaluation samples did not exhibit detectable fluorescence, indicating that both the ability to stain microplastics of Nile red itself and particle size were key limiting factors of detection. These results suggests that ML based Nile red fluorescence analysis can achieve practical consistency with FTIR based identification while eliminating bias that depends on the operator. Uniform identification in fluorescence-based method by machine learning models, combined with rigorous QA/QC of the analytical workflow, provides reliable baseline information for addressing the global threat posed by plastic pollution.