GoogLeNetMP: A Development of GoogLeNet Architecture for Multi-Class Microplastic Classification in Subsurface Water Image

Abstract

Microplastic pollution has become a major environmental concern due to its persistence in marine ecosystems and its potential impact on aquatic organisms and human health. Automatic detection of microplastic particles in underwater environments remains challenging because of turbidity, low contrast, light distortion, and the visual similarity between microplastics and natural marine objects. This study proposes GoogLeNetMP, an enhanced GoogLeNet-based deep learning architecture for multi-class classification of subsurface marine images into four categories: primary microplastics, secondary microplastics, non-microplastics, and marine biota. The proposed framework integrates basic image preprocessing (resizing and noise reduction) with a modified GoogLeNetMP architecture designed to intrinsically handle fine-grained feature extraction under degraded conditions, thereby minimizing the reliance on complex external enhancement pipelines. A dataset of underwater images acquired from the coastal waters of Padang, Indonesia, was used for model development and evaluation. Experimental results show that GoogLeNetMP outperformed the standard GoogLeNet model, achieving 95.75% accuracy, 92.80% sensitivity, 97.00% specificity, and an F1-score of 92.06%. The proposed model also demonstrated more stable training convergence and better discrimination of visually challenging classes. The architecture is designed to internalize the robust feature extraction process, thereby minimizing the reliance on extensive external enhancement pipelines while maintaining standard normalization steps for input consistency. These findings indicate that GoogLeNetMP is a promising approach for AI-based marine pollution monitoring and decision support in sustainable coastal management.