コークスサイト周辺土壌における多環芳香族炭化水素（PAHs）の地域スケール暴露リスク評価のためのスタッキングアンサンブルモデルに基づく評価

Coking industrial activities contribute to approximately one-quarter of soil polycyclic aromatic hydrocarbons (PAHs) pollution in the Beijing-Tianjin-Hebei region, posing risks to soil quality and human health. High-resolution, multi-site prediction remains challenging because conventional models seldom incorporate source-explicit drivers. Here, we developed a regional risk-mapping framework that integrates coking enterprise source strength (SBI) and modeled atmospheric deposition flux (CPES) into a stacked-ensemble learning pipeline. A total of 320 soil samples collected around 20 representative plants were used for model training and validation; the optimized model was then applied to predict ∑PAHs within 3-km buffers of 130 coking sites across the region. We constructed three progressively enhanced covariate sets-environmental factors (EF), EF+SBI, and EF+SBI+CPES-to test performance gains and driver attribution. Incorporating SBI and CPES raised R² from 0.41 to 0.63 and reduced RMSE from 0.42 to 0.33, while identifying distance to site, atmospheric deposition, and NDVI as dominant predictors. Operational plants had higher mean ∑PAHs than closed plants (1028 vs 644 ng/g). Closed plants still surpassed background levels by about 1.3 times. Scenario analysis showed a clear greening effect. In the priority management zone, increasing NDVI by 0.025 decreased the predicted mean from 541.7 ng/g to 489.5 ng/g, halving moderately-heavily polluted farmland, and reducing the potentially exposed population by over 40 %. The source-aware framework improves accuracy and transferability of regional PAH exposure risk assessment and supports tiered, site-adjacent zoning and mitigation around industrial point sources.