A Long-Term (2002-2025) PM2.5 Concentration Dataset over China Based on an Adaptive Machine Learning-Based Dynamic (AMLB-Dynamic) Model

This dataset provides long-term PM2.5 concentration estimates over mainland China from June 26, 2002 to October 31, 2025 at a spatial resolution of 0.1 degrees. Day, month, season, and annual scales are available. The dataset was generated using the AMLB-Dynamic framework, an automated multi-stage ensemble model integrating AutoML and LightGBM with adaptive clustering strategies. The model incorporates multi-source data including satellite-derived aerosol optical depth (AOD), meteorological reanalysis data (ERA5), vegetation index (NDVI), and topographic data (DEM). The AMLB-Dynamic model employs a two-stage stacked ensemble approach. In the first stage, AutoML generates preliminary PM2.5 estimates based on feature-augmented data. In the second stage, a multi-seed weighted LightGBM ensemble refines predictions. Additionally, a concentration-aware modeling strategy based on K-means clustering is introduced to improve estimation performance across different pollution levels. The dataset demonstrates high accuracy when validated against ground observations (R² > 0.91, RMSE < 9.51 μg/m³) and provides continuous spatial coverage across China. It supports multi-scale analysis including daily, monthly, and annual variations. This dataset can be used for:- Air pollution analysis- Climate and environmental studies- Public health research- Policy evaluation For detailed methodology, please refer to the associated publication.