Estimation of black carbon concentration in China based on a dynamic weighted ensemble learning model

Accurate estimation of historical black carbon (BC) concentrations is essential for assessing long-term environmental health risks and public health impacts. Machine learning has shown high predictive power in environmental studies. To address the lack of accuracy and generalizability of traditional basic machine learning models and previous integrated models for estimating BC concentration, a dynamic weighted ensemble learning (DWEL) model with Bayesian optimization is proposed in this study. The model is based on ground-based monitoring data in China from 2013 to 2022, and is used to estimate the black carbon concentration for the whole coverage of China from 2013 to 2022 through a novel three-stage architecture that integrates multi-source satellite observations (tracking of air pollution in China (TAP) of BC) data, meteorological parameters, and atmospheric pollution data by (1) using four different machine learning algorithms to generate the base predictions; (2) fusing the original features with these preliminary predictions through four metamodels to further extract the complex nonlinear relationships in the data; and (3) using the Bayesian hyperparameter tuning technique, dynamically adjusting the contribution of each metamodel, and adding bootstrap sampling to reduce the risk of overfitting. The principal results are as follows: (1) The predictive performance of the DWEL model on the test set (R² = 0.84; RMSE = 0.77 µg/m³) was better than that of the baseline model (R² between 0.79 and 0.82) by evaluating 1,740 monthly observations from 114 monitoring stations, also showed a significant advantage in the seasonal estimation results. (2) Spatio-temporal validation further demonstrated a high degree of agreement with independent TAP BC (r = 0.90) and modern-era retrospective analysis for research and applications, Version 2 (MERRA-2) reanalysis data (r = 0.89), as well as closer agreement with ground-based measurements than with existing reanalysis products. (3) This study also reveals the spatial and temporal evolution of BC concentrations in China from 2013 to 2022: spatially, the eastern and coastal economically developed regions are generally higher, while the western regions are relatively lower, and temporally, the overall trend shows a decreasing trend from year to year, with the low-concentration regions ( <1 µg/m³) covering half of the Chinese region in 2022, with higher concentrations in autumn and winter and lower concentrations in spring and summer. The model proposed in this study has advantages in terms of accuracy and precision in estimating black carbon concentrations compared to traditional integrated learning models and basic machine learning models. In conclusion, the DWEL framework provides a powerful tool for black carbon exposure assessment and supports data-driven environmental decision-making.