Random Forest-Based Prediction of Coastal Microplastic Concentration Using High-Dimensional Environmental Data: A Comparative Study with Deep Learning and Machine Learning

Abstract

Microplastic contamination in coastal ecosystems has emerged as a critical environmental issue with significant ecological, economic, and public health consequences. Conventional monitoring approaches rely heavily on field sampling and laboratory-based analysis, which are time-consuming, resource-intensive, and lack scalability. Recent advances in artificial intelligence offer opportunities for automated, data-driven pollution forecasting. This study presents a comprehensive comparative evaluation of classical machine learning and deep learning models for predicting coastal microplastic concentration using environmental and anthropogenic indicators. A dataset containing 1000 cleaned observations was analyzed using Random Forest regression and a Deep Neural Network trained on a log-transformed target variable. Experimental results demonstrate that Random Forest significantly outperforms deep learning, achieving an R² score of 0.9673 and RMSE of 17.86, while the neural network achieved 0.5350 and 79.18 respectively. These findings reveal that classical ensemble learning methods are more suitable than deep learning for moderate-sized tabular environmental datasets. The proposed framework provides a scalable, low-cost, and accurate solution for automated coastal pollution prediction and decision support.