In order to improve the accuracy of forest above-ground biomass estimation， constructed from modeling factor selection and modeling aspects, a PSO-LSSVM biomass estimation method was proposed by considering comprehensive of the image texture features, topographical features, spectral features. Selecting Songshan Nature Reserve as study area, with the data sources from ZY-3 satellite remote sensing image, the measured data of 194 survey plots, forest resource inventory data, and the digital elevation model data, the Pearson correlation relationship was analyzed between 46 feature variables and forest above-ground biomass. With the optimal feature extraction variables chosen, the PSO-LSSVM model was established in Matlab 2014a. The determination coefficient (R) and root mean square error (RMSE) were taken for comparative analysis of the accuracy of PSO-LSSVM model and multiple linear regression model. The results showed that the prediction accuracies (R) of PSO-LSSVM model in coniferous forest, broadleaf forest and shrub were 0.867, 0.853 and 0.842, which were improved by 23.15%, 19.13% and 14.40% compared with the multiple linear regression model, respectively. The PSO-LSSVM model had self-study ability and adaptive capability, it can replace the traditional traversal optimization method, and it had great advantages on global optimization and convergence rate with small sample volume requirement and high precision accuracy.