AI-Driven Framework Development for Predictive Classification of Microplastic Concentration of Aquatic Systems in the United States

Microplastics present a critical environmental and public health issue, accumulating in ecosystems and human tissues with harmful consequences. Despite their significance, there is a lack of precise predictive models for microplastic density, hindering effective monitoring and mitigation efforts. This paper systematically studied machine learning approaches by comparing four models: three classical algorithms (Logistic Regression, Random Forest, and Support Vector Machine) and one deep learning model (PyTorch Neural Network). The analysis focused on three geographically and environmentally distinct major United States coastal regions-the West Coast, East Coast, and Gulf of Mexico-using three diverse datasets to ensure comprehensive training and validation. The results indicate that SVM was the most accurate model (Average accuracy across 3 regions: 93.94 %), followed by Random Forest $(81.25 \%)$, and then finally Logistic Regression (48 %). Key environmental factors identified included daylight duration, sunlight intensity, and wind direction. SHapley Additive explanations (SHAP) analysis provided critical insights into these findings by quantifying the impact of each factor, with region-specific influences such as precipitation and river runoff in the Gulf of Mexico or urbanization metrics like vehicle miles traveled on the East Coast. These factors reflect the complex interactions between climatic, geographical, and human-driven elements in determining microplastic density. The PyTorch Neural Network frequently overfitted the data, even when layers were reduced, due to the relatively small and less complex dataset used. This study highlights the potential of machine learning in identifying and predicting microplastic hotspots, offering a foundation for targeted mitigation strategies and improved environmental management.