AI-driven prediction of marine microplastics from space: a case study of the Indian Ocean

Abstract

Microplastics (MPs) are considered enormous threats to the environment on a global scale, yet their monitoring in oceanic regions remains challenging due to limited sampling and complex circulation patterns. This study leverages 12 remotely sensed oceanographic variables to predict microplastic abundance (MPA) in the Indian Ocean by developing machine learning (ML) models. Random forest, gradient boosting regression (GBR), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), support vector regression (SVR), k-nearest neighbors (KNN), and multiple linear regression (MLR) were the ML models used to train by utilizing 346 georeferenced data. Among the evaluated models, the GBR model demonstrated superior performance during cross-validation, achieving the highest predictive accuracy and model stability (CV R2 = 0.8434, CV RMSE = 0.2424). SHAP analysis revealed that sea surface salinity, surface roughness, dissolved oxygen, distance from the nearest coast, nitrate concentration, and zonal wind stress were among the most influential predictors associated with microplastic distribution, highlighting the importance of hydrodynamic transport, coastal influence, and biogeochemical processes in regulating marine microplastic accumulation. The findings demonstrate that remotely sensed and reanalysis-derived oceanographic variables can provide environmentally informed estimates of MPA across data-limited marine environments. Therefore, this integrated ML-remote sensing approach provides a scalable, cost-effective pollution monitoring system, supporting focused management and policy changes in areas with limited data. GBR demonstrated superior performance and stability during cross-validation. Sea surface salinity, surface roughness, dissolved oxygen, and distance from the nearest coast are the key influencing factors. Adaptive transformation of highly skewed environmental variables may provide incremental predictive benefits. Leakage-free preprocessing, 10-fold cross-validation, Optuna hyperparameter optimization, were the strategies to control overfitting. Nexus of machine learning and remote sensing exploration. GBR demonstrated superior performance and stability during cross-validation. Sea surface salinity, surface roughness, dissolved oxygen, and distance from the nearest coast are the key influencing factors. Adaptive transformation of highly skewed environmental variables may provide incremental predictive benefits. Leakage-free preprocessing, 10-fold cross-validation, Optuna hyperparameter optimization, were the strategies to control overfitting. Nexus of machine learning and remote sensing exploration.