PM2.5 monitoring is an important means of air pollution control. Limited by the number of ground observation points, PM2.5 estimation from remote sensing data is an effective complement to conventional ground observation. The key idea of remote sensing estimation of PM2.5 is to retrieve aerosol optical depth firstly, and subsequently to reverse PM2.5 by aerosol optical depth based on the statistical relationship. This approach however is highly possible to cause error transmission, leading to instability of the inversion model. In this paper, we propose a PM2.5 remote sensing estimation method based on random forest algorithm to directly establish the relationship between moderate resolution imaging spectroradiometer (MODIS) image and ground measured PM2.5, so as to avoid the inversion error of atmospheric aerosol optical depth, finally obtain the PM2.5 estimation result with high precision. The method first uses random forest to train and test the MODIS image and ground monitoring station PM2.5 data after kriging interploation, and then selects the best model from multiple models according to the root mean square error (RMSE) of test index. Finally, the approach uses this model in the whole MODIS image to obtain the PM2.5 estimation result of the whole area. This experiment selects many MODIS image data from four seasons in Guangdong province to verify and compare the two performance indicators of R2 and RMSE. The results show that the proposed approach outperforms other approaches significantly.