A feature-level ensemble framework for improving daily PM2.5 estimation across the contiguous United States (2000–2023)

Abstract

Accurate PM2.5 surface concentration estimates are vital for air quality management and exposure assessment. This study introduces a novel feature-level ensemble framework to enhance daily PM2.5 estimation across the contiguous United States from 2000 to 2023. The framework integrates multiple XGBoost models trained with diverse temporal features, including calendar encodings and physically derived indicators like rolling averages and change rates from reanalysis PM2.5. By capturing complementary pollution dynamics, the ensemble outperforms models using only calendar features. Under spatial cross-validation, R2 increases from 0.64 to 0.70 and RMSE drops from 7.87 to 7.31 μg/m3. Temporal extrapolation also improves, with △R2 gains of 0.08 (historical backcasting) and 0.07 (future forecasting). These results demonstrate the framework's robustness, generalizability, and value for long-term PM2.5 monitoring, epidemiological research, and data-driven air quality policymaking.