Forests are essential in maintaining carbon balance, and accurate detection of forest biomass plays a crucial role in promoting environmental improvement and related policy formulation. The comprehensive analysis of multi-source data and abundance information was explored to achieve forest biomass inversion. Firstly, MOPSOSCD was used to obtain the endmember bundle of the study area, and the abundance information of each group of tree endmembers was obtained. Totally 46 indicators, including single band factor, vegetation index, terrain factor, texture feature, etc., were extracted from Landsat 8 OLI and ASTGTM DEM to test the model fitting effect before and after fusion abundance through biomass inversion experiments using multiple linear regression and BP neural network models. It was found that when using the multiple linear regression model, among the seven optimized models extracted for abundance, two groups had better RMSE than the multiple linear regression model without added abundance, and seven groups had better R2 than the multiple linear regression model without added abundance. The RMSE and R2 before optimization were 41.09 mg/hm2 and 0.40, and the optimal RMSE and R2 were 38.66 mg/hm2 and 0.44, respectively. When using the BP neural network model, all BP models with added abundance showed an improvement in RMSE, with six groups having better R2 than those without added abundance. The RMSE and R2 before optimization were 32.73 mg/hm2 and 0.56, and the optimal RMSE and R2 were 32.07 mg/hm2 and 0.57, respectively. The BP neural network model with added abundance had the best inversion effect. The MOPSOCCD algorithm was used to extract the abundance corresponding to the endmember bundle, and abundance was used as the inversion factor for biomass inversion. The experiment demonstrated the effectiveness of abundance in improving biomass inversion accuracy. Meanwhile, the higher the accuracy of extracting endmembers was, the better the corresponding model biomass inversion effect was.