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

Assessing future trends in marine microplastic (MP) abundance is a crucial step toward mitigating MP pollution. However, this task is challenged by the scarcity of observational data and the pronounced spatiotemporal heterogeneity of MPs driven by multiple interacting factors. In this study, we introduce CGMAT, a novel deep learning (DL) framework that integrates Few-Shot Learning (FSL) with a Transformer-based architecture. CGMAT enhances heterogeneous datasets from the Taiwan Strait and the Norwegian coastal waters to identify key drivers of MP pollution and to predict the future spatiotemporal distribution of MPs. Multi-scenario simulations demonstrate that Cross-domain Multi-Graph Attention Network (CGMAT) framework achieves excellent performance on the source domain validation data (explained variance score (EVS) = 0.91, mean absolute percentage error (MAPE) = 0.18 %). Nevertheless, forecast results reveal significant regional variations in MP pollution trends. Specifically, MP concentrations in the Taiwan Strait are projected to increase sharply, reaching 312-376 particles/m³ around 2030, whereas concentrations along the Norwegian coast waters are expected to rise more gradually, peaking at 15-53 particles/m³ around 2031. Following the peak, pollution levels are anticipated to stabilize under the combined influence of environmental dynamics and mitigation measures. The multi-scale feature fusion architecture of CGMAT further reveals that the spatiotemporal dynamics of MP distribution are governed by the interplay of three principal mechanisms: the intensity of economic interventions, delayed environmental responses, and geographical barriers. These findings highlight the significant potential of combining FSL with Transformer-based DL models to address data scarcity challenges and provide a broadly applicable framework for different marine ecosystems.