[Paper Review] Morphological Detection and Classification of Microplastics and Nanoplastics Emerged from Consumer Products by Deep Learning

Hadi Rezvani, Navid Zarrabi+6 more|arXiv (Cornell University)|Sep 20, 2024

Microplastics and Plastic Pollution4 citations

TL;DR

This paper introduces MiNa, an open-source dataset of micro- and nanoplastics (MiNa) derived from consumer plastics under simulated aquatic conditions, using scanning electron microscopy (SEM) images for deep learning-based detection and classification. It benchmarks YOLOv10, Faster R-CNN, and Mask R-CNN models, finding two-stage detectors (Faster R-CNN, Mask R-CNN) outperform one-stage YOLOv10 in recall and F1 score, with false negatives most prevalent in polypropylene (PP) and polystyrene (PS) due to high particle density and size variability.

ABSTRACT

Plastic pollution presents an escalating global issue, impacting health and environmental systems, with micro- and nanoplastics found across mediums from potable water to air. Traditional methods for studying these contaminants are labor-intensive and time-consuming, necessitating a shift towards more efficient technologies. In response, this paper introduces micro- and nanoplastics (MiNa), a novel and open-source dataset engineered for the automatic detection and classification of micro and nanoplastics using object detection algorithms. The dataset, comprising scanning electron microscopy images simulated under realistic aquatic conditions, categorizes plastics by polymer type across a broad size spectrum. We demonstrate the application of state-of-the-art detection algorithms on MiNa, assessing their effectiveness and identifying the unique challenges and potential of each method. The dataset not only fills a critical gap in available resources for microplastic research but also provides a robust foundation for future advancements in the field.

Motivation & Objective

• To address the critical gap in open, diverse, and realistic datasets for automated micro- and nanoplastics (MiNa) detection and classification using deep learning.
• To develop and evaluate state-of-the-art object detection models (YOLOv10, Faster R-CNN, Mask R-CNN) on a novel dataset of MiNa particles across polymer types and sizes.
• To identify and analyze the primary challenges in MiNa detection, including particle crowding, overlapping particles, and size variability, particularly for PP and PS.
• To provide a foundation for future research by enabling automated, accurate, and cost-effective quantification and identification of MiNa in environmental and consumer product samples.

Proposed method

• The MiNa dataset comprises 1,000 SEM images of micro- and nanoplastics (1 µm – 1 mm) generated from consumer plastics (PE, PS, PP, PET) after 12-week degradation under environmentally relevant conditions.
• Images are annotated with bounding boxes and class labels for polymer type (PE, PS, PP, PET), enabling training of object detection models.
• Three deep learning models—YOLOv10 (one-stage), Faster R-CNN, and Mask R-CNN (two-stage)—are trained and evaluated using metrics including precision, recall, F1 score, and AP50.
• Evaluation includes analysis of false negatives per polymer class, with focus on challenges like particle overlap, small size, and high density, particularly in PP and PS.
• The study uses intersection over union (IoU) for bounding box evaluation and applies statistical sampling to estimate total particle counts from detected particles.
• The dataset is publicly released to support reproducibility and future development of automated MiNa detection systems.

Experimental results

Research questions

• RQ1 How effective are state-of-the-art deep learning object detection models (YOLOv10, Faster R-CNN, Mask R-CNN) in detecting and classifying micro- and nanoplastics from SEM images of degraded consumer plastics?
• RQ2 What are the primary failure modes in MiNa detection, and how do they vary across different polymer types (e.g., PP, PS, PE, PET)?
• RQ3 To what extent do particle size, density, and overlapping morphology affect detection performance, especially for two-stage vs. one-stage models?
• RQ4 Can a publicly available, realistic, and diverse SEM-based dataset significantly improve the automation and accuracy of MiNa detection compared to existing datasets?

Key findings

• Two-stage models (Faster R-CNN and Mask R-CNN) outperform the one-stage YOLOv10 model in both recall and F1 score, with Faster R-CNN achieving the highest AP50 (mean average precision at IoU threshold of 0.5).
• False negatives are most prevalent in polypropylene (PP) particles (22.1% false negative rate), followed by polystyrene (PS) (20.3%), primarily due to high particle density and overlapping morphology.
• Polyethylene (PE) particles have the lowest false negative rate (8.5%), attributed to their larger average size and lower image crowding.
• Polyethylene terephthalate (PET) particles have a 12.7% false negative rate, linked to their larger average area and lower particle count per image.
• The models struggle significantly with overlapping and densely packed particles, indicating a persistent challenge in real-world environmental samples.
• The MiNa dataset enables accurate estimation of total particle counts via random sampling and statistical inference, accelerating lab-based identification and reducing costs.