Abstract: Using ICESat–2/ATLAS data as data source, combined with 54 measured plots, a machine learning model was built and the AGB of the spot footprint was predicted. Moran's I and semi-variance function were used to study the spatial autocorrelation and heterogeneity of inverse forest AGB. The results showed that the Gradient Boost Regression Tree (GBRT) model had a great prediction accuracy (R²=0.90, RMSE=11.08 t/hm²). The best-fitting semi-variogram function model of forest biomass was exponential model in Shangri-La (C₀=0.12, C₀ + C=0.87, A₀=10200 m). Compared with ordinary Kriging, the spatial distribution of AGB obtained by the Sequential Gaussian Conditional Simulation had better consistency (r=0.59^**, d=0.70). The spatial differentiation of AGB could be explained by topographic factors. In terms of explanatory power, elevation was the largest, slope was the second, and slope was the least. The inversion accuracy of forest AGB based on spaceborne LiDAR ICESat-2/ATLAS data was high (Pp=81.43%), which provided a reliable data source for geostatistical analysis. Therefore, the method based on spaceborne LiDAR and geostatistics can greatly analyze the spatial heterogeneity of forest AGB.