Multi-scenario simulation of future marine microplastic distribution under data scarcity: A deep learning approach

Area of Science:

• Environmental Science
• Marine Biology
• Data Science

Background:

• Marine microplastic (MP) pollution poses a significant environmental threat.
• Accurate prediction of MP abundance is hindered by data scarcity and complex spatiotemporal variations.
• Understanding MP distribution drivers is crucial for effective mitigation strategies.

Purpose of the Study:

• To develop a novel deep learning framework (CGMAT) for predicting marine microplastic pollution.
• To identify key drivers influencing MP spatiotemporal distribution.
• To forecast future MP trends in diverse marine ecosystems.

Main Methods:

• Integration of Few-Shot Learning (FSL) with a Transformer-based architecture in the Cross-domain Multi-Graph Attention Network (CGMAT).
• Enhancement of heterogeneous datasets from the Taiwan Strait and Norwegian coastal waters.
• Multi-scenario simulations and multi-scale feature fusion analysis.

Main Results:

• CGMAT demonstrated high performance with an explained variance score (EVS) of 0.91 and a mean absolute percentage error (MAPE) of 0.18%.
• MP concentrations are projected to increase sharply in the Taiwan Strait (312-376 particles/m³ by 2030) and gradually in Norwegian waters (15-53 particles/m³ by 2031).
• MP dynamics are governed by economic interventions, environmental response times, and geographical factors.

Conclusions:

• The CGMAT framework effectively addresses data scarcity in marine pollution modeling.
• Future MP pollution trends exhibit significant regional disparities.
• The study provides a broadly applicable DL approach for marine ecosystem management and pollution mitigation.