Introduction
Mapping trace metals like lead (Pb) in the global ocean is a fundamental task for understanding human pollution and ocean circulation. However, data is sparse and heavily clustered in specific regions, leaving vast "blind spots" in the Southern and Indian Oceans. Traditional interpolation methods often fail to capture the complex geochemical relationships governing lead distribution. This study addresses these challenges by utilizing explainable machine learning—specifically XGBoost and SHAP values—to predict global seawater lead and isotope distributions while mitigating spatial bias.
Methodology
The research evaluates two ensemble learning algorithms: XGBoost and Random Forest (RF). XGBoost was selected for the final models due to its superior performance in predicting Pb concentrations and isotope ratios ($^{206}Pb/^{207}Pb$ and $^{208}Pb/^{207}Pb$).
Feature Engineering and Coordinate Exclusion
Models were trained on environmental features including temperature, salinity, dissolved oxygen, and nutrient concentrations, alongside atmospheric data such as Black Carbon and Dust Aerosol Optical Depth (AOD). A key methodological innovation was the intentional exclusion of latitude and longitude as features. Although coordinates initially improved local accuracy, they caused the model to overfit to sampled regions, creating "spatial artifacts" in unsampled areas. Removing coordinates forced the model to rely on physical and chemical oceanographic features, ensuring more scientifically robust global reconstructions.
Explainability
The team employed SHAP (SHapley Additive exPlanations) values to quantify the contribution of each environmental feature. This approach revealed that seawater density is a powerful predictor for lead isotopes, likely because density surfaces act as conduits for water masses carrying distinct isotopic signatures.
Datasets and Validation
The models were trained and evaluated using target variables for Pb concentration and isotope ratios. Data was split into a "Random test set" and a "Geographic test set" to evaluate the model's ability to generalize to entirely different ocean basins.
As illustrated in the sample distribution, the dataset is spatially biased. Furthermore, depth frequency analysis reveals that the majority of samples are concentrated in the upper ocean, presenting a common challenge in marine data science.
Results
The Pb concentration model achieved a high $R^2$ of 0.91. The isotope models showed remarkably low error rates, with a Mean Absolute Percentage Error (MAPE) of 0.2% for $^{206}Pb/^{207}Pb$ and 0.1% for $^{208}Pb/^{207}Pb$. By testing on geographic hold-outs, the authors demonstrated that the model could successfully predict Pb distributions in data-poor regions like the South Pacific.
Limitations and Future Work
Despite the model's success, data scarcity in the deep Indian and Southern Oceans remains a constraint. Future iterations could incorporate temporal features to track how lead levels change in response to shifting industrial regulations. Additionally, expanding the feature set to include more biological parameters could refine the model's understanding of lead's "scavenging" behavior.
Summary updated