Mapping the Plastic Legacy: Geospatial Predictions of a Microplastic Inventory in a Complex Estuarine System Using Machine Learning

The persistence of microplastics (MP) in aquatic environments presents a pressing concern, with sediments serving as substantial repositories for these anthropogenic particles. Estuarine depositional systems are sensitive indicator environments for MP monitoring as they: (1) represent aquatic-terrestrial interfaces acting as a bottleneck for MP accumulation coming from land (capable of capturing riverine run-off as well as the often highly populated coastal areas), (2) comprise a confined area, favourable for sampling initiatives as compared to other systems such as long river courses or disperse inland waters. Hence, understanding MP distribution in intricate estuarine systems is essential, yet current models often falter in capturing complexities to sufficiently describe the present pollution patterns for an entire geomorphological region.

We address this gap by employing machine learning techniques to predict spatial MP inventories deposited in the Schlei, Northern Germany. Notably, the complex hydrodynamic regime, influenced by narrowings and braided embayments, freshwater tributaries, wind-driven mixing, and brackish inflows, creates an interplay of fluvial and marine sedimentary processes, which poses non-trivial challenges for reliable modelling of sedimentary MP transportation. Our approach, termed NIXVEGS (Nested Iterative X-Validation-to-Ensemble-modelling through Grid Searches), leverages machine learning, integrating model selection, rigorous validation, and ensemble techniques tailored for small datasets.

We estimated ~20 trillion MP particles or ~14.5 tonnes (50-5000 µm) residing in upper sediments of the Schlei proper, emphasizing the pivotal role of sediments as primary MP reservoirs. Our modelling concept is founded on the idea of applying granulometric proxies to account for the hydrodynamic regime bias in observed MP concentrations. We found that the high complexity of the geomorphology and extreme input events – both are predominant conditions in our study system – produce major spatio-temporal discontinuities in MP data which are not alleviated by a granulometric normalisation. Here we use hydrodynamic tracer simulations to derive variables which incorporate these discontinuities in empirical predictive modelling, but discuss simpler possibilities to enable modelling studies in systems for which such simulation data might not be feasible to acquire.

This study provides a novel framework for geospatial prediction of MP inventories in complex aquatic systems. The integration of granulometric proxies and hydrodynamic discontinuities elucidates MP distribution patterns, offering a pathway for robust predictions and informed mitigation strategies. Our findings underscore the critical role of sediments in storing and reflecting the contemporary plastic legacy, crucial for comprehensive environmental management.