Estimating Black Carbon Concentrations in China by Combining Machine Learning Models and Optimization of Satellite-Based Input Schemes

Previous efforts to estimate ground-level Black Carbon (BC) via remote sensing have been hindered by substantial inaccuracies, largely due to the scarcity of ground-based observations and challenges associated with the applicability of selected methods and interpretability of input parameters. This study sought to evaluate the performance of different machine learning models in estimating BC concentrations by employing a range of satellite-based relevant input variables. Specifically, we utilized 885 monthly ground observations collected from 48 monitoring stations across China between 2013 and 2020. These observations were integrated with satellite products like Tracking Air Pollution in China (TAP) BC, China High Air Pollution (CHAP) atmospheric pollutants (PM2.5, CO, SO2, NO2), and other parameters to create seven distinct input schemes. Four machine learning models coupled with the seven input schemes were employed to estimate BC concentrations. By validating against ground-truth data, eXtreme Gradient Boosting (XGBoost) demonstrated the best performance on the test set, with an R² of 0.76 and RMSE of 1.27 μg/m³, surpassing the other models (Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Light Gradient Boosting Machine (LGBM)) as well as the TAP BC estimate. Furthermore, the Shapley Additive Explanations (SHAP) analysis revealed that CO, TAP BC, atmospheric pressure, and NO2 are the four most important predictors of BC concentrations. Utilizing the optimal input scheme and the best-performing model, monthly BC concentrations were estimated for each 0.1°×0.1° grid cell in China from 2013 to 2020. The results indicated that the annual average BC concentration in China decreased from 2.4 μg/m3in 2013 to 1.29 μg/m3in 2020, with the high concentration area (>3μg/m3) decreasing from 14.6% in 2013 to 1.6% in 2020.