The window size of texture feature extraction has a significant impact on the accuracy of tree species classification. The forms of all texture features share an optimal extraction window, and different types of texture features use their independent optimal extraction windows, which is conducive to tree species classification. In this study, we used a RedEdge-MX image as the data source and a random forest to determine two forms of the best texture extraction windows and construct their own best texture feature set. Then, we combined the best texture feature sets with spectral bands and the digital surface model (DSM) to analyze the difference between the two best texture extraction forms in tree species classification. The results show that the classification accuracy of the best texture feature set was significantly different between the two extraction forms. The overall accuracy of the first extraction form was 79.6365% and that of the second extraction form was 81.8915%. When they are combined with a spectral band and the DSM, the classification accuracy of the latter was higher than that of the former (between 0.4295% and 2.2248%). Hence, in the classification of tree species, the construction of the best texture feature set should be determined by the best extraction window for each feature type.