Emission-Based Machine Learning Approach for Large-Scale Estimates of Black Carbon in China

Abstract

Tremendous efforts have been made to construct large-scale estimates of aerosol components. However, Black Carbon (BC) estimates over large spatiotemporal scales are still limited. We proposed a novel approach utilizing machine-learning techniques to estimate BC on a large scale. We leveraged a comprehensive gridded BC emission database and auxiliary variables as inputs to train various machine learning (ML) models, specifically a Random Forest (RF) algorithm, to estimate high spatiotemporal BC concentration over China. Different ML algorithms have been applied to a large number of potential datasets and detailed variable importance and sensitivity analysis have also been carried out to explore the physical relevance of variables on the BC estimation model. RF algorithm showed the best performance compared with other ML models. Good predictive performance was observed for the training cases (R2 = 0.78, RMSE = 1.37 μgm−3) and test case databases (R2 = 0.77, RMSE = 1.35 μgm−3) on a daily time scale, illustrating a significant improvement compared to previous studies with remote sensing and chemical transport models. The seasonal variation of BC distributions was also evaluated, with the best performance observed in spring and summer (R2 ≈ 0.7–0.76, RMSE ≈ 0.98–1.26 μgm−3), followed by autumn and winter (R2 ≈ 0.7–0.72, RMSE ≈ 1.37–1.63 μgm−3). Variable importance and sensitivity analysis illustrated that the BC emission inventories and meteorology showed the highest importance in estimating BC concentration (R2 = 0.73, RMSE = 1.88 μgm−3). At the same time, albedo data and some land cover type variables were also helpful in improving the model performance. We demonstrated that the emission-based ML model with an appropriate auxiliary database (e.g., satellite and reanalysis datasets) could effectively estimate the spatiotemporal BC concentrations at a large scale. In addition, the promising results obtained through this approach highlight its potential to be utilized for the assessment of other primary pollutants in the future.