The diversity of tree species is an important content of ecological research. The information on tree species and spatial distribution can effectively serve sustainable forest management. However, it is difficult to obtain detailed information on the spatial distribution of tree species in traditional field-based forest inventory. Tree species classification based on remote sensing costs less and has high spatial accuracy, which has become an effective method. Although satellite remote sensing data has been successfully applied in the study of tree species distribution, it is difficult to obtain high-precision classification results limited by its spatial resolution and spectral resolution, especially in complex stand conditions. The UAV remote sensing can obtain local ultra-fine data, which provides the possibility to improve the classification accuracy of tree species. Therefore, this research based on the method of machine learning and the concept of feature fusion to explore the potential for tree species classification under subtropical forest conditions. The multi-source data used in this study includes visible light, hyperspectral, Light Detection, respectively based on which texture features, Minimum Noise Fraction Rotation (MNF), as well as vegetation indexes, And Ranging (LiDAR) structural parameters are extracted and calculated. The impact of classification processes and methods such as classifiers, different data sources, and different classification features on classification accuracy are studied in this research to provide experience and examples for high-precision classification and mapping of subtropical forests. The study found that: (1) The random forests classifier has the highest overall accuracy and F1 score of each tree species, which is more suitable for subtropical multi-tree species classification and mapping. The overall accuracy is 95.63%, and the Kappa coefficient is 0.948 when distinguishing 13 categories (8 trees, 4 herbs). (2) As far as the single data source is concerned, the order from high to low of the model accuracy is hyperspectral, LiDAR and visible light data. Multi-source data can significantly improve the classification accuracy. The full-feature model has the highest accuracy, and hyperspectral and LiDAR data significantly affect the classification accuracy of the full-feature model. The visible light texture data has less effect. (3) The importance of classification features is sorted from largest to smallest into structural information, vegetation index, texture information, and minimum noise transformation component. In addition, texture and MNF characteristics cannot effectively distinguish tree species in subtropical forests. And the data after MNF dimensionality reduction will lose lots of information so that original band information is more important when hyperspectral data are used.