Microplastic Identification in Seawater using Generative Adversarial Networks

In recent years, microplastics-plastic particles smaller than five millimeters-have become a significant environmental contaminant. These tiny plastic pieces have been found permeating aquatic ecosystems globally, posing a threat to marine life upon ingestion. However, identification and quantification of microplastics in seawater samples remains an arduous task. The small size and translucent nature of microplastics makes them difficult to distinguish from natural particulates in seawater under a microscope. This has hindered efforts to accurately monitor microplastic pollution levels in the marine environment. Our goal in this study was to use cutting edge deep learning techniques to create an automated image based system for detecting microplastics in seawater. We chose to employ a generative adversarial network (GAN) architecture for its proven capability in generating highly realistic synthetic images. The GAN model was trained on a dataset of microscope images of microplastics and natural particulates extracted from seawater samples. Coastal seawater samples were first filtered to isolate particulates of a size range containing microplastics. Density separation was then used to separate the plastic microparticles from denser natural particulates. We implemented a deep convolutional GAN with a generator network to produce synthetic microplastic images, and a discriminator network to differentiate real from synthetic images. Through iterative adversarial training, the two networks evolved to produce and accurately classify images of microplastics with a high degree of realism. Advanced image processing techniques were used to enhance the training data. After training, the GAN model was evaluated on an annotated testing dataset of images containing real-world microplastics and natural particulates extracted from seawater. Our proposed GAN-based technique achieved a microplastic classification accuracy of 92.5% on the test set, demonstrating its effectiveness. The development of an automated visual detection system for microplastics can accelerate the analysis of seawater samples to monitor plastic pollution. Our study provides a proof-of-concept for the viability of using GANs to achieve rapid and accurate identification of microscopic plastic debris from environmental samples. With further development, this approach could become an invaluable tool for marine pollution sensing.