Enhanced spatiotemporal mapping of urban wetland microplastics: An interpretable CNN-GRU approach using satellite imagery and limited samples

Comprehensive investigation of microplastics (MPs) in urban surface waters aids in managing and understanding this emerging contaminant that threatens ecosystems and human health. This work constructed an interpretable CNN-GRU model integrating Satellite Remote Sensing (SRS) data with limited in-situ samples to analyze spatiotemporal distribution characteristics of MPs pollution in wetlands of the Chebei drainage basin, Guangzhou, China. The proposed model exhibited superior predictive performance for MP concentrations, achieving a mean predicted concentration of 44.71 items/L with excellent accuracy (R² = 0.9526) and low errors (MAE = 2.93, RMSE = 4.17) on independent test dataset. Through SHapley Additive exPlanations (SHAP) analysis, the near-infrared band was identified as the most notable predictor, showing consistent positive contributions to model outputs. Furthermore, Remote Sensing Index (RSI) analysis established a notable correlation between MPs abundance and algal content in urban wetlands. The strong sensitivity of near-infrared bands to algal content suggests that algal biomass may serve as a reliable proxy for assessing MPs accumulation levels in these ecosystems. Spatiotemporal analysis revealed notably greater MPs accumulation in wetlands than reservoirs, with seasonal peaks observed during summer months. This work demonstrates the feasibility of SRS-based MPs spatiotemporal mapping with limited samples in urban wetlands, and the SHAP method enhanced the credibility and interpretability of predictions. Notably, the model's reliability depends on the correlation between MPs and algal content, which may vary across different aquatic ecosystems. These findings provide a practical case study for applying interpretable deep learning to urban wetland MPs monitoring, with insights that can be referenced for similar regions.