An interpretable deep forest model for estimating hourly PM10 concentration in China using Himawari-8 data

Rapid urbanization and industrialization in China had led to increased pollutants emission. PM10 is one of the main components of air pollutants, which significantly impacts human health, environment, and regional or global climate. In this study, a new machine learning deep forest (DF) model was used to construct the aerosol optical depth (AOD) and near-ground PM10 concentration (AOD-PM10) model. The DF model combines the advantages of deep neural networks and tree models, which can provide model interpretability. Combined with the Himawari-8 AOD, meteorological, and auxiliary factors, the hourly PM10 concentration in China (spatial resolution: 0.05 x 0.05) was obtained. The results show that AOD has the highest contribution to the importance of features in the AOD-PM10 model, accounting for approximately 13.5%, and the contributions of boundary layer height, temperature, and relative humidity to the importance of features were 11%, 8.6%, and 7%, respectively. A 10-fold cross-validation was used to evaluate the performance of the model. The hourly cross validation results from 09:00 to 16:00 (Beijing time) show that the R-2 range was 0.82-0.88, and the root mean square error and absolute mean error were 18.55-23.12 mu g/m(3) and 11.54-16.82 mu g/m(3), respectively. The R-2 values of daily, monthly, seasonal, and annual average PM10 estimated by the model were 0.87, 0.91, 0.94, and 0.94, respectively. The areas with high PM10 concentrations are mainly in northern China, especially in the North China Plain, and the peak value of daily average PM10 can reach 91 mu g/m(3); the Intraday variation of PM10 in southern China ranges from 67 mu g/m(3) to 72 mu g/m(3). A large-scale dust weather process was analyzed. Based on the AOD-PM10 model, the contribution of long-range transport dust to PM10 in China and Northern China were 25.6% and 38.1%, respectively. The PM10 measured by the station and estimated by the DF model indicated good consistency.