There are various problems concerning the acquisition of forest digital terrain using satellite-based LiDAR (LiDAR). In this paper, a method for extracting the classified photon point cloud is summarised based on the photon point cloud acquired by ICESat-2/ATLAS strong and weak beams for forest digital terrain inversion. In this study, ATLAS was used as the data source to group canopy heights, and airborne LiDAR G-LiHT was used as the validation data to verify the effect of different canopy heights on the inversion of the understorey terrain. The results show that, when the canopy heights were not grouped, the inversion accuracy of both strong and weak ATLAS beams was higher, with an R² greater than 0.99, an RMSE of less than 0.7 m, and an MAE of less than 0.4 m. Overall, the strong beams were superior. Moreover, when the canopy heights were grouped, the R² was greater than 0.99 and the accuracy was best at the canopy height of 5-10 m. The R² decreased when the height exceeded 35 m. The inversion accuracy of both the strong and weak ATLAS beams was higher than 0.99. The accuracy of both strong and weak beam inversions was extremely high, with the strong beam being superior overall. However, both strong and weak beams can provide a scientific basis for the inversion of understory DTMs. Canopy height affected the inversion accuracy, which was inversely proportional to the canopy height, with an increase in canopy height causing a decrease in the accuracy. The inversion of the understorey topography was better than that of the dwarf tree and shrub layers.