Air pollution has emerged as one of the most urgent problems worldwide over the past few years. Especially, fine particulate matter with a diameter smaller than 2.5 micrometres (PM_2.5) has seriously endangered the human life and health worldwide. Thus, monitoring accurate and spatio-temporal continuously PM_2.5 concentrations is essential for controlling air pollution. The key to accurately estimating PM_2.5 concentrations is to build a reliable aerosol optical depth (AOD)–PM_2.5 model, which is mainly determined by the AOD remotely sensed products, meteorological parameters, and estimating methods. However, no systematic evaluation and analysis of the influence of these three variables on AOD–PM_2.5 model has yet been reported. Based on Terra and Aqua Moderate Resolution Imaging Spectroradiometer AOD products, reanalysis meteorological data, and ground-based PM_2.5 data during 2014–2015 in eastern China, we analyzed the influence of above mentioned variables on the accuracy of the AOD–PM_2.5 model. The four modeling methods investigated are stepwise multiple linear regression (SMLR), backpropagation neural network (BPNN), classification and regression tree (CART), and random forest (RF). Adjusted R², mean absolute error (MAE), product index (PI), spatial resolution index (SRI), and Taylor diagrams are used to assess the accuracy of models. The results show that adding meteorological elements that vary with time and height can improve the accuracy by 4.97% over eastern China. The average PI of Terra AOD products increased by 0.15 compared with that of Aqua AOD products. Most SRIs have positive values, suggesting that AOD products with a spatial resolution of 10 km can lead to a more accurate AOD–PM_2.5 model. The Taylor diagram indicates that RF model > BPNN model > CART model > SMLR model in terms of performance. Therefore, Terra AOD products with spatial resolution of 10 km and RF are more suitable for use with the AOD–PM_2.5 model in eastern China.