Identification of microplastics in natural environment using hyperspectral microscope and machine learning

Abstract

Mass production of plastic began in the 1950s, and production has increased approximately 260 times by 2021. Due to the increase of production, it has also led to an increase of plastic waste which recycled rate was only 9%. Rest of the microplastics were discharged into nature through incineration, landfill, and ocean discharge. Waste plastic exposed to the environment is broken into small pieces by exposure to ultraviolet rays or weathering. Small plastics less than 5 mm in size are called microplastics. Microplastics exposed to the environment not only cause damage due to leakage of additives and adsorption of other pollutants, and threaten marine life, but can also pose a major threat by being found in human bodies, so regulation and monitoring of microplastics is necessary. The devices mainly used for microplastic analysis methods include FT-IR spectroscopy, Raman spectroscopy, and pyrolysis GC-MS. However, all three reporter analysis devices require analyzing particles individually one by one, which takes a long time to analyze, and known for the expensive analysis cost. Therefore, recently, analysis using hyperspectral imaging, which is also used for monitoring and overcomes the limitations of existing equipment, is being actively conducted. Therefore, in this study, we attempted to exceed the existing detection limit (100 µm) by combining hyperspectral imaging with microscopy and developed a technique that improved analysis accuracy for each type of microplastic using machine learning By combining a hyperspectral camera with a microscope, conditions were sought to maximize the detection signal of microplastics. Afterwards, a large amount of data was collected and an image classification model was first trained to obtain only the plastic spectrum by separating the plastic and background filters. Approximately 12,000 images were annotated for plastic using Roboflow software and then trained. As a result of learning, a model with a precision of 89% was learned, and it was confirmed that particle size classification up to 10 µm was possible. Next, 800,000 hyperspectral signals were collected for the classified plastics, and six other models were trained using H2O.ai software. As a result, the accuracy was highest when 400,000 signals were used, and the Stacked ensemble and GBM model have shown the highest accuracy which were 84%. Afterwards, to evaluate the image classification model and the signal learning model, an experiment was conducted under conditions where four plastics were mixed, and the Stacked ensemble and GBM models showed precision of 86% and 80%. As a result of testing the size detection limit using this new technology and the models, it was confirmed that analysis was possible up to 20 µm for PP and 10 µm for PA, PE, and PET. Ultimately, the hyperspectral camera-based high-speed/high-efficiency microplastic detection technique developed through this study is expected to be used to understand the environmental impact of micrometer-sized microplastics.