GoogLeNet-Based Deep Learning Framework for Underwater Microplastic Classification in Marine Environments

The growing pollution of the marine ecosystem due to microplastic waste is a dangerous threat to the environment and human health. Most existing research has focused on shallow-water detection of macroplastic debris, but little has been done on underwater microplastic classification because of the light scattering caused by turbidity. This paper suggests a deep learning model that is founded on the GoogLeNet convolutional neural network (CNN) which is applicable in classifying the underwater image into multiple classes. The model classifies images into four groups, primary microplastics, secondary microplastics, non-microplastic rubbish, and marine biota. The data were captured with a submersible drone in coastal waters of Padang, Indonesia and further supplemented with rotation, flipping, contrasting, and addition of Gaussian noise as resulting in a dataset of 2,400 images. The preprocessing tools including median filtering, contrast stretching, and HSV color transformation were incorporated to enhance clarity of the image. The optimized GoogLeNet had an accuracy, precision, recall, and F1-score of 95.2, 96, 96, and 95.8 respectively, which was better than baseline CNNs such as ResNet-18 and MobileNet and was computationally efficient. The proposed setup demonstrates scalability to real-time underwater tracking and leads to the mitigation of marine pollution in accordance with Sustainable Development Goal (SDG) 14. The next steps to be undertaken in the future involve the expansion of data gathering to various regions along the coastlines and research on transformer-based architectures to extract contextual features.