Tianyi Guan

About

This project developed an auditable, leakage-controlled machine-learning framework for spatially resolved European cadmium, mercury, and lead emission inventories. The study integrated heterogeneous gridded emissions, contextual covariates, target-family auditing, spatially structured validation, train-only preprocessing, model benchmarking, and constrained interpretation to support provisional inventory updating and quality-control screening. Rather than treating high predictive accuracy as sufficient evidence, the framework explicitly separated absolute-emission prediction from baseline-relative change and evaluated both against operational persistence and no-change baselines.

Research Focus

To determine whether spatially resolved heavy-metal emission inventories can be updated using machine learning without contaminating model development through spatial, temporal, target, or preprocessing leakage, while distinguishing persistence-driven accuracy from genuine change-prediction skill.

Study Design

European EMEP gridded Cd, Hg, and Pb inventories were harmonized with co-emission and contextual predictors into metal-specific datasets. Spatial blocks were assigned before feature screening, and all imputation, scaling, and feature selection were fitted using training data only. Compatible 2018–2019 national-total targets were modeled as both absolute log10 emissions and baseline-relative log10 changes using ten regression algorithms, including ElasticNet, random forests, gradient boosting, neural networks, XGBoost, LightGBM, and CatBoost. Model selection was controlled by validation performance, followed by held-out spatial-block testing, operational baseline comparison, spatial cross-validation, residual diagnostics, and permutation-based interpretation.

Key Findings

Selected LightGBM and XGBoost models achieved held-out spatial-block R² values of 0.973–0.985 for absolute emissions and 0.712–0.861 for relative-change tasks. Across metals and target horizons, selected models reduced RMSE by 43.9–73.5% relative to persistence or no-change baselines. Co-emission inventories consistently carried more predictive information than broad population covariates, while the weaker change tasks highlighted the limits of interpreting spatial persistence as temporal updating skill.

Keywords

Environmental informatics; heavy-metal emissions; emission inventory; spatial machine learning; data leakage; spatial validation; LightGBM; XGBoost; inventory updating; target-family auditing; model benchmarking;