TL;DR
This paper uses machine learning to create the first global maps of lead (Pb) concentration and its isotopes in seawater, filling a critical data gap. By leveraging oceanographic and atmospheric data, the models reveal pollution hotspots and sinking pollution plumes, highlighting the Southern Ocean as a priority for future sampling.
What Problem Does This Paper Address?
Lead (Pb) and its isotopes are invaluable tracers for understanding both pollution pathways from land to sea and the ocean's own biogeochemical processes. However, direct measurements of Pb in seawater are notoriously difficult due to its low concentrations and the risk of contamination during sampling. This scarcity of data, especially in vast regions like the Southern Hemisphere, has prevented a comprehensive global understanding of Pb pollution and its oceanic fate. This study aims to bridge this knowledge gap by creating detailed, global maps of Pb distribution and isotopic signatures.
Methods
The researchers employed machine learning, specifically the XGBoost (eXtreme Gradient Boosting) algorithm, to build three predictive models: one for Pb concentration and two for Pb isotope ratios (²⁰⁶Pb/²⁰⁷Pb and ²⁰⁸Pb/²⁰⁷Pb). These models were trained using a combination of existing oceanographic and atmospheric data as "features" to predict the Pb measurements. To understand why the models made their predictions, they utilized SHapley Additive exPlanations (SHAP) values. SHAP is a powerful technique that quantifies the contribution of each feature to individual predictions, allowing for an interpretable understanding of the model's learned relationships. The models were rigorously validated using cross-validation and separate test sets, including a "geographic test set" to mimic real-world scenarios where data is absent.
Datasets
The study primarily utilized the GEOTRACES Intermediate Data Product 2021v2, a comprehensive dataset of global oceanographic measurements, including Pb concentrations and isotope compositions. This was supplemented with additional published and unpublished Pb data.
For the predictive features, the researchers used global gridded climatologies from:
- World Ocean Atlas 2018 (WOA18): Providing oceanographic variables like temperature, salinity, density, and nutrient concentrations.
- Copernicus Marine Service Global Ocean Biogeochemistry Hindcast: For surface chlorophyll-a concentrations.
- Copernicus Atmosphere Monitoring Service (CAMS): For atmospheric dust aerosol optical depth (AOD) and black carbon AOD.
The resulting global maps of Pb concentrations and isotope compositions are made freely available.
Key Insights and Ideas
- Machine Learning as a Data Gap Filler: The study demonstrates the power of ML in generating valuable scientific insights even when direct observational data is sparse, a common challenge in oceanography.
- Interpretable AI for Ocean Science: The use of SHAP values is crucial, moving beyond "black box" models to reveal the underlying oceanographic and atmospheric drivers of Pb distribution.
- Southern Ocean as a Pollution Hotspot & Priority: The models consistently highlight the Southern Ocean as an area with high uncertainty and significant pollution signals, making it a prime candidate for future targeted sampling campaigns.
- Sinking Pollution: The models suggest that pollution from previous decades is not just at the surface but is actively sinking in major ocean basins like the North Atlantic and Pacific.
- Anthropogenic Fingerprints: Highly anthropogenic Pb isotope signatures, originating from the Southern Ocean, are shown to be spreading at intermediate depths, carried by water masses like Antarctic Intermediate Water (AAIW).
Key Results
- Global Maps Generated: The study produced the first global, three-dimensional maps of Pb concentrations, ²⁰⁶Pb/²⁰⁷Pb, and ²⁰⁸Pb/²⁰⁷Pb ratios in seawater.
- Key Predictors Identified:
- For Pb concentration: Seawater temperature, atmospheric dust, black carbon AOD, and salinity were most influential.
- For ²⁰⁶Pb/²⁰⁷Pb: Dissolved oxygen, salinity, temperature, and atmospheric dust were key.
- For ²⁰⁸Pb/²⁰⁷Pb: Black carbon AOD, dust AOD, seawater temperature, and surface chlorophyll-a were most important.
- Pollution Hotspots: The surface Indian Ocean shows the highest Pb pollution levels.
- Model Performance: The models achieved reasonable performance, with MAPE values for Pb concentration around 18-20% and RMSE values for isotopes around 0.005-0.008. Performance varied by ocean basin, with the Arctic and Southern Oceans showing lower accuracy, likely due to data sparsity and outlier presence.
- Uncertainty Mapping: Coefficients of variation highlight the Southern Ocean and Arctic as areas with higher model uncertainty, reinforcing them as targets for future research.
Limitations and Future Work
- Data Sparsity: Despite using ML, the models are still limited by the inherent sparsity of Pb data, particularly in under-sampled regions like the Southern Ocean.
- Climatological Features: The use of climatological data for atmospheric inputs (dust and black carbon AOD) might not fully capture the regional and temporal variability of Pb sources, leading to some counter-intuitive SHAP explanations.
- Coastal vs. Open Ocean: The pooling of dissolved and total dissolvable Pb data, while justified for the open ocean, might introduce uncertainty in estuarine and coastal areas where isotopic re-equilibration is more significant.
- Future Work: The authors emphasize the need for new observations, especially from coastal areas and the Southern Ocean, to validate model outputs and improve understanding. They also suggest that these models can serve as a benchmark for future pollution levels and a valuable comparison for process-based models aiming to simulate marine Pb cycling.
Summary updated