Automated detection of morphologically diverse microplastics using ensemble deep learning

Abstract

Microplastics in water emerge as hazardous materials that contaminate the environment and adversely impact public health. Current approaches for detecting and characterizing microplastics are manual and inefficient in time, cost, and labor. Recent advances in deep learning-based computer vision approaches have enhanced efficiency, while they suffer from low accuracy for various types of microplastics. This paper presents a size-adaptive ensemble deep learning approach, aiming to overcome the challenge of morphological diversity and automate the detection of microplastics. The approach employs two specialized YOLOv8 models, including a high-resolution model for small particles (e.g., pellets) and a standard-resolution model for large particles (e.g., fibers, films, fragments). The presented approach was implemented and evaluated using a dataset including 1,600 images. The results show that the approach improved detection accuracy (17.5% increase in mAP50, 32.1% increase in mAP50-95) compared with traditional unified deep learning models. This research advances the ability to automatically detect various microplastics based on computer vision and deep learning.