An integrated framework of optimized learning networks for classifying oil-mixed microplastics

Abstract

Neural network modeling (e.g., deep neural networks) has been growingly applied in environmental data and imaging analysis but is significantly limited by its large dataset size and quality demand. To help address the challenge, this study developed an integrated framework by combining the convolutional neural networks (CNNs) and the improved deep convolutional generative adversarial networks (i-DCGANs). Optimization algorithms introduced to DCGANs for enhancing data augmentation. The employment of local interpretable model-agnostic explanations benefited the modeling prediction and reliability. The framework was tested by an environmental case study on characterizing the scanning electron microscopy (SEM) images of the weathered microplastics and oil-dispersants agglomerates (WMODAs) under various weathering conditions. The optimized model achieved a high score on the F-test (0.9192) with an accuracy of 0.9986, indicating a robust prediction. The classified results effectively differentiated WMODAs at different weathering degrees to facilitate a better understanding of the impact of microplastics on oil fate and transport in the oceans where plastic and oil pollution coexist in a growing trend. The proposed approach could be useful for image-related classification in other fields along with the findings to help fill some knowledge gaps of oil-plastic co-contamination.