A Machine Learning Approach To Microplastic Detection And Quantification In Aquatic Environments

Abstract

The pervasive contamination of aquatic ecosystems by microplastics (MPs), defined as plastic particles <5 mm, poses a significant threat to marine life and human health. Current methods for their analysis, primarily involving visual counting under microscopes followed by spectroscopic validation, are labor-intensive, time-consuming, and prone to human error. This study presents a robust, automated machine learning (ML) framework for the detection and quantification of microplastics from digital microscopy images of water samples. We developed a pipeline that utilizes a deep learning object detection model, YOLOv7, to accurately identify and classify MPs based on size and shape (e.g., fibers, fragments, beads). Subsequently, a pixel-wise segmentation model, U-Net, is employed for precise quantification of particle dimensions. We curated a novel dataset of over 5,000 annotated microscope images from water samples collected from various aquatic sources. The YOLOv7 model achieved a mean Average Precision (mAP@0.5) of 96.8% in detecting MPs, while the U-Net model achieved a Dice coefficient of 0.94 for particle segmentation. Our system significantly reduces analysis time from hours per sample to minutes, with a high degree of accuracy and reproducibility. This approach provides a scalable, efficient, and accessible tool for environmental monitoring agencies and researchers, enabling large-scale mapping and monitoring of microplastic pollution.