Hyperspectral lidar (HSL) can obtain spatial and spectral information simultaneously, which provides more possibilities for three-dimention (3D) reconstruction of wood-leaf. The multi-view tree point clouds were collected using a self-developed HSL system for wood-leaf separation and 3D reconstruction. In order to meet the strict condition of traditional iterative closest point (ICP) algorithm, an improved point cloud registration method was proposed in this work. Firstly, the local feature of every point was described using fast point feature histogram (FPFH). Then, rough and accurate registration of point clouds were successively conducted based on random sample consensus (RANSAC) algorithm and ICP algorithm respectively. Finally, the wood and leaf components were separated based on the spectral data of the registered point clouds, the feature bands were extracted for wood and leaf separation on the basis of the spectral feature channel selection with random forest (RF) algorithm and support vector machine (SVM) algorithm, and 3D reconstruction of wood-leaf components was completed based on separation under different view. The experiment results show that the proposed improved point cloud registration method can achieve excellent registration accuracy both under 15° and 30° view angle difference, and the corresponding separation accuracy using RF algorithm based on feature channel selection reaches 98.17% and 98.87%, respectively.