mnnnnn b

Abstract

\section{} Microplastics (plastic particles smaller than 5 mm) are among the most widespread environmental pollutants, and monitoring them requires both counting particles and identifying their polymer type. Spectroscopic methods such as FTIR and Raman are accurate but slow and costly. Nile Red staining offers a fast, inexpensive alternative, as stained plastics fluoresce under blue light, but turning fluorescence images into quantitative data requires reliable automated detection and classification. We used the public Nile Red dataset of Rermborirak et al. (2025), covering six polymer classes (ABS, Nylon, PE, PET, PS and PVC). A systematic audit found that 1,012 of the 2,566 published images (39.4%) were near-duplicates, many with bounding-box labels rotated by 180°. After de-duplication, 1,554 unique images (23,832 labelled particles) were split 70/15/15 into training, validation and test sets, stratified by polymer class. Five YOLO detectors (YOLOv8s, YOLOv9s, YOLOv10s, YOLO11s and YOLO26s) were trained for 60 epochs at 640 pixels from COCO-pretrained weights, all with identical hyperparameters. The best architecture, YOLO26s, was then fine-tuned for 100 further epochs at 640 and 960 pixels, with model selection based on validation performance only. Evaluation combined standard detection metrics with an independent error analysis separating located, misclassified, missed and spurious particles. YOLO26s, YOLO11s and YOLOv10s performed equivalently (test mAP50-95 0.822–0.828), about 3 points above YOLOv8s (0.795), the architecture used in the source study. The final YOLO26s model, fine-tuned at 960 pixels, achieved test precision 0.909, recall 0.904, mAP50 0.962 and mAP50-95 0.876, a gain of 8.1 points over YOLOv8s. At a confidence threshold of 0.25 it located 96.0% of labelled particles, and only 0.5\ a threshold of 0.50 it counted particles with a mean absolute error of 1.46 per image. The accuracy gain splits roughly equally between newer architecture (+2.8 points), longer training (+2.4) and higher input resolution (+2.9). Polymer classification of located particles is effectively solved; the main remaining challenge is spurious detection of faint, unannotated particles. The model, code and methodology are released for reproducibility.