Abstract Numerous previous studies have revealed that statistical models which combine satellite-derived aerosol optical depth (AOD) and PM2.5 measurements acquired at scattered monitoring sites provide an effective method for deriving continuous spatial distributions of ground-level PM2.5 concentrations. Using the national monitoring networks that have recently been established by central and local governments in China, we developed linear mixed-effects (LMEs) models that integrate Moderate Resolution Imaging Spectroradiometer (MODIS) AOD measurements, meteorological parameters, and satellite-derived tropospheric NO2 column density measurements as predictors to estimate PM2.5 concentrations over three major industrialized regions in China, namely, the Beijing–Tianjin–Hebei region (BTH), the Yangtze River Delta region (YRD), and the Pearl River Delta region (PRD). The models developed for these three regions exploited different predictors to account for their varying topographies and meteorological conditions. Considering the importance of unbiased PM2.5 predictions for epidemiological studies, the correction factors calculated from the surface PM2.5 measurements were applied to correct biases in the predicted annual average PM2.5 concentrations introduced by non-stochastic missing AOD measurements. Leave-one-out cross-validation (LOOCV) was used to quantify the accuracy of our models. Cross-validation of the daily predictions yielded R2 values of 0.77, 0.8 and 0.8 and normalized mean error (NME) values of 22.4%, 17.8% and 15.2% for BTH, YRD and PRD, respectively. For the annual average PM2.5 concentrations, the LOOCV R2 values were 0.85, 0.76 and 0.71 for the three regions, respectively, whereas the LOOCV NME values were 8.0%, 6.9% and 8.4%, respectively. We found that the incorporation of satellite-based NO2 column density into the LMEs model contribute to considerable improvements in annual prediction accuracy for both BTH and YRD. The satisfactory performance of our models indicates that constructing LMEs models using various combinations of predictors for different regions would be helpful for predicting PM2.5 concentrations with high accuracy.