Machine Learning Identification of Pollutants and Other Debris in Canadian Waterways

Abstract Plastic pollution is a global challenge, necessitating innovative monitoring solutions. This paper explores the use of Artificial Intelligence (AI)-enhanced methods for litter detection by leveraging state-of-the-art convolutional neural network (CNN) architectures. A novel dataset was created, using stationary camera data from Canada. This dataset enabled a comparative analysis of model performance, identifying DenseNet121 and MobileNetV2 as optimal architectures for large and small model categories, respectively. The study also investigates the impact of training models on specific environments versus unified datasets that encompass varied conditions. Results show that a unified dataset leads to only a minor decrease in predictive performance while retaining the ability to distinguish between classes unique to specific datasets. This capability may be attributed to the distinct image properties resulting from differences in imaging equipment and techniques, which needs further investigation. To enhance future model performance, the importance of consistent imaging properties is therefore emphasized. The next steps involve integrating the developed models into a hardware setup for long-term monitoring of water sections, as well as exploring advanced AI techniques for identifying individual and sparsely distributed litter items.