The space-borne photon counting laser altimetry can detect continuous elevations of vegetation canopy and earth surface for its high along-orbit resolution. However, the relatively low point cloud density and low signal-to-noise ratio (SNR) of space-borne vegetation point clouds put forward new requirements for estimating vegetation canopy heights. In this paper, an adaptive directional model for estimating vegetation canopy heights using space-borne vegetation point clouds was proposed to meet the new requirements. Firstly, the range of signal elevation was roughly obtained by searching two extremums that represent the crown and ground in the statistical histogram of point cloud elevation. The land slope and average densities of crown, ground and noise were estimated as well. Then, the roughly denoised point clouds were further fine denoised by adaptive directional density-based clustering where the direction of neighborhood is along the land surface, and the thresholds related to density are adjusted adaptively according to the estimated point cloud densities. After filtering, the elevations of ground and canopy were estimated respectively by applying triangular irregular networks (TIN) where the initial points of ground and canopy in TIN were found by the densities and elevation percentage of point clouds. Vegetation point clouds of ATLAS space-borne laser altimeter are used to validate the filtering method. The experimental results show that the adaptive directional model can correctly estimate vegetation canopy heights and is fit for areas with large slope and low leaf area index. The determination coefficients R² of canopy and ground elevation between processed ATLAS data and airborne LIDAR data are 0.99 and 0.77 respectively, and RMSE are 0.28 M and 2.6 m.