Herein, light detection and ranging data were collected as remoting data sources by terrestrial laser scanning (TLS). Metasequoia, palm, sapindus, bamboo, and rubber trees were selected as research objects. Three effective features are proposed, which are relative clustering features of trees, features of point cloud distribution of trees, and apparent features of trees. 68 feature parameters are listed. A support vector machine (SVM) classifier was then used to verify and calculate the training dataset and to determine the optimal feature parameters in cross-validation. Finally, the tree species is classified in the test dataset. The research results show that the average classification accuracy of tree classification based on the optimal parameters of relative clustering features of trees is low (45%), that based on the optimal feature parameters of point cloud distribution slightly increases (58.8%), that based on the optimal parameters of tree appearance features is relatively high (63.8%), and that based on the 13 optimal parameters of three types of features is the highest (87.5%). In addition, due to the difference between metasequoia and other tree species is obvious, the metasequoia is outstanding in classification and its misjudgement rate is the lowest (6.5%). The proposed method has high feasibility and provides a powerful tool for obtaining a more accurate distribution of forest species.