Monitoring Marine Microplastic Distributions in Data-Scarce Ocean Regions Using Environmental Representation Learning and Explainable Machine Learning

Monitoring pollution in the ocean from microplastic and making decisions using this data is challenging, especially in ocean areas with little data and geographically limited observations. The ability to accurately predict microplastic distributions is crucial for the identification of pollution hotspots and for understanding environmental risk and for taking decisions on marine conservation. This study introduces a framework for learning a representation of the environment which is explainable and aware of uncertainty that combines NOAA marine microplastic observation data with oceanographic and geospatial data provided by Copernicus Marine and GEBCO datasets. The proposed Contrastive-XGBoost model outperformed the other models with the highest predictive capability with R2 = 0.651, RMSE = 1.008 and MAE = 0.619. Predictors identified by SHAP analysis included depth, bathymetry, geographical location, temperature, ocean circulation, and proximity to coast. Ablation analysis revealed the advantage of using combined (heterogeneous) environmental variables. Global prediction maps indicated that coastal and semi-enclosed seas, such as the Bay of Bengal, South China Sea, Gulf of Mexico and Mediterranean Sea are potentially accumulation hotspots. Conventional XGBoost, however, showed statistically comparable predictive power and geographically independent validation showed a limited transferability to environmentally different regions. The results show the possibility of using interpretable environmental machine learning for supplying additional data for environmental pollution monitoring, hotspot detection and uncertainty aware environmental decision-making.