Leakage-controlled machine learning for European Cd, Hg, and Pb emission inventories

Timely and spatially resolved inventories of cadmium (Cd), mercury (Hg), and lead (Pb) are needed for transboundary pollution assessment and inventory quality control, yet annual grid products may be delayed, incomplete, or internally inconsistent across releases. We developed an auditable, leakage-controlled framework that converts heterogeneous European emission products and contextual covariates into metal-specific, machine-learning-ready datasets. Its central contribution is the integration of data harmonization, target-family auditing, split-before-processing validation, operational baselines, model-agnostic benchmarking, and interpretation with explicit causal limits; short-horizon spatial updating serves as a downstream demonstration. Spatial blocks were assigned before screening, and all imputation, scaling, and feature selection were fitted using training data only. Compatible 2018-2019 targets were evaluated as absolute log10 emissions and baseline-relative log10 change using ten regression algorithms. LightGBM was selected for five of six tasks and XGBoost for Pb change. On held-out spatial blocks, selected models achieved $R^2=0.973$ -0.985 for absolute emissions and 0.712-0.861 for changes, reducing root mean squared error by 43.9-73.5% relative to persistence or no-change baselines. Co-emission inventories carried more information than broad population covariates, while the weaker Hg change task and the contrast between absolute and relative-change performance exposed metal- and task-specific limits. The framework is intended for provisional inventory updating, anomaly screening, and prioritization of grid cells for expert review, not as a replacement for bottom-up compilation or as evidence of unrestricted forecasting to entirely unseen years.