PM 2.5 is directly related to the air quality and is harmful to human health, thus high-precision monitoring of PM 2.5 is necessary. The dispersion and accumulation of PM 2.5 are influenced not only by near-ground meteorological elements but also by the upper-air meteorological elements. Nevertheless, the current PM 2.5 inversion models based on deep neural network only consider ground elements. To further optimize the model and improve the inversion accuracy, a PM 2.5 hourly inversion model integrating upperair meteorological data was herein proposed, and parametric rectified linear unit was used as the activation function of the model. The results showed that among the input elements, PM 2.5 had the highest correlation with AOD, reaching 0.33. The proposed model achieved the highest accuracy on the test set, with RMSE, MAE, and R 2 of 14.39 μg/m 3 , 9.67 μg/m 3 , and 0.83, respectively. Compared to surface meteorological data and surface+850hPa meteorological data, the RMSE of the proposed model on the test set was reduced by 23.13% and 17.05%, respectively. Meanwhile, the RMSE of the proposed model on the test set was reduced by 56.15%, 39.99%, 14.60% and 5.76%, respectively, compared with adaptive boosting, gradient boosting regression, random forest, and the integrated model of these three models. During the heating season in Shanxi Province, the high-value areas of PM 2.5 were mainly distributed in the basin area, the PM 2.5 reached the highest in November, and peaked at 11 a.m. during the day.INDEX TERMS Aerosol optical depth, air pollution, deep neural network, PM 2.5 , remote sensing.