We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Interpretable deep-learning assessment of environmental indicators of global marine plastic pollution across ocean basins and climate zones
Summary
Scientists used artificial intelligence to predict where ocean plastic pollution is likely to accumulate, based on factors like water oxygen levels, distance from coastlines, and the amount of tiny plant life (phytoplankton) in the water. The key finding: what drives plastic pollution differs significantly by region, meaning a one-size-fits-all cleanup or prevention strategy won't work as well as targeted, location-specific approaches. This matters because better prediction of plastic hotspots could help protect seafood supplies and coastal communities from the microplastics that eventually make their way into the food we eat.
Marine plastic pollution has emerged as a pressing global environmental issue and poses a significant threat to both oceanic ecosystems and human well-being. Nevertheless, the patterns of its dispersion and the environmental associations underlying its distribution remain inadequately characterized on a worldwide scale, primarily due to the limitations of traditional monitoring and modeling techniques. In this study, we introduced an interpretable deep learning (DL) framework to predict the plastic pollution level across global oceans and to elucidate the model-attributed environmental associations. Multiple global datasets of plastic pollution concentrations and relevant environmental variables from 2018 to 2022 were integrated. A total of 1193 surface-ocean records were initially compiled. After harmonization and quality control, 1155 records were retained for model development, including 916 training samples and 239 samples in a held-out internal test set. The predictive performance of the FCNN was benchmarked against several widely used machine-learning models. Although XGBoost achieved the highest predictive accuracy and was identified as the best-performing benchmark model, the FCNN demonstrated competitive predictive performance and was retained as the focal deep-learning architecture for evaluating the proposed interpretable neural-network framework. The FCNN model demonstrated moderate predictive performance under random internal validation based on 13 environmental predictor variables spanning geographic, biological, biogeochemical, physicochemical, physical-dynamic, and transport-related categories, achieving a mean five-fold cross-validation R 2 of 0.608 and a held-out internal test-set R 2 of 0.612. Shapley Additive Explanations (SHAP) analysis identified DTC, dissolved oxygen, net primary productivity, and phytoplankton concentration as the main environmental indicators contributing to model predictions. Stratification analysis further revealed distinct regional differences in model-attribution patterns and predictor importance across the four major ocean basins (Atlantic, Pacific, Indian, and Arctic) and climate zones (tropical, temperate, and polar), thereby emphasizing the necessity for region-specific management strategies. This interpretable DL framework provides a data-driven approach for large-scale marine plastic pollution assessment and regional monitoring support within the range of environmental conditions represented by the available data.