Abstract: This study extracted the distribution data of 12 dominant tree species from the forest management inventory data of Kunming, and the remote sensing variables from Landsat 8 OLI remote sensing images in 2016. Then, liner stepwise regression model(LSR), back propagation neural network model(BPNN), and random forest model(RF) were used to estimate the forest aboveground biomass(AGB), then selected the optimal model for each forest type to have the AGB inversion in Kunming. The results showed that the RF has the highest accuracy in 3 models, which the adjusted determination coefficient( R_\mathrma\mathrmd\mathrmj^2 ) is 0.683 and root mean square error(RMSE) is 12.68 t/hm² in the Chinese fir forests. And the fitting accuracy of LSR is lowest. When the AGB lower than 50 t/hm² and larger than 100 t/hm², the overestimation and underestimation can be found in all 3 models, but the RF often has the lower absolute mean error values. Moreover, it often has the lowest estimation errors in the different biomass segments. The inversion precision of random forest model is 85.31%. The RF would provide a good tool for invert the forest AGB in Kunming.