Geospatial-based estimation of NMHC concentrations through an ensemble stacking Geo-AI algorithm to advance air...

Non-methane hydrocarbons (NMHCs) are volatile organic compounds that play a key role in atmospheric chemistry, serving as precursors to ozone and secondary organic aerosols (SOAs). Despite their environmental and health significance, geospatial-based estimation of NMHC concentrations remains limited due to sparse monitoring networks. To address this gap, this study presents an enhanced geospatial framework for estimating NMHC concentrations in Taiwan, by integrating daily data from 34 monitoring stations between June 2015 and 2021, with the final year reserved for external validation. Four major modeling approaches were evaluated, including a statistical model (land-use regression, LUR), deep learning (Deep Neural Networks, DNN), single machine learning models (Extreme Gradient Boosting (XGBR), Gradient Boosting Regressor (GBR), LightGBM Regressor (LGBMR), CatBoost Regressor (CBR), and Random Forest Regressor (RFR)), and an ensemble stacking model. The best-performing approach was the ensemble stacking model, which integrates five single machine learning algorithms. This ensemble framework achieved strong predictive performance (R² = 0.820), outperforming the LUR model (R² = 0.441) and the DNN model (R² = 0.708). Key predictors included pure residential areas, NO₂, O₃, water bodies, road networks, wind speed, and other land-use and meteorological variables, reflecting the combined influence of anthropogenic emissions and atmospheric processes. Geospatial-based estimation maps revealed distinct spatial patterns and identifiable NMHC hotspots, with elevated concentrations in urban residential regions and selected mountainous areas, while lower levels were observed in coastal and rural regions, likely due to enhanced ventilation and fewer emission sources. Despite temporary emission reductions during the COVID-19 pandemic, the selected model consistently captured underlying temporal trends, demonstrating robustness. Overall, this study demonstrates that ensemble stacking provides an effective framework for improving NMHC geospatial estimation, offering valuable insights for air quality management, environmental monitoring, and exposure assessment.