Abstract: The Nianzhu experimental forest farm of subtropical forestry experimental center was selected as the research area, and the visible spectral and LiDAR data of UAV were used as data sources for tree species identification in this study. Based on CHM and visible spectral data, single-wood was firstly carried out, and then extracted features of visible spectral data and LiDAR data to build more characteristic collection. Based on object single-wood to choose RF and SVM classifier to classify recognition, and using confusion matrix to evaluate the accuracy of 12 schemes with different data sources and different feature combinations. To compare different characteristics of the combination and classifier to the influence of tree species classification accuracy. The results show that the single tree segmentation based on CHM segmentation and multiresolution segmentation is effective, which can meet the needs of single tree species identification; and the results show that the precision of SVM is higher than that of random forest classifier, and the precision after RF feature selection is better than those without feature selection. The overall accuracy is improved by 1.45% on average, and the average precision of visible spectral and LDAR data integration has improved the average accuracy by 6.01% compared with using only a single date source. It can be seen that feature selection can reduce dimension disaster, effectively avoid redundancy caused by too many features, and the performance and efficiency of classifiers are further improved. In contest to the RF classifier, SVM performs better in the case of limited multidimensional sample sets and training samples. The combination of multi-source data can effectively combine the advantages of different data sources and improve the classification accuracy.