This study proposes a method based on machine learning algorithms to improve the accuracy of hyperspectral tree species classification by changing spatial resolution at the regional scale， providing a new approach for tree species classification research in terrestrial surveys. This study used drones to obtain hyperspectral images of the entire Chengdu Botanical Garden， and collected 1 249 samples of 140 tree species in the garden. By constructing 32 vegetation indices and 176 original bands for variable screening， a classification model was established using two algorithms： random forest and support vector machine. Based on the forest stand types and canopy sizes of typical tree species in the study area， 10， 15， and 20 tree species were selected at 9 different spatial resolutions to explore the accuracy of tree species classification. The results showed that when the spatial resolution gradually decreased from 0.12 m to 4 m， the classification accuracy of the models for 10， 15， and 20 tree species reached the highest level at a resolution of 3 m， and the overall accuracy of the support vector machine classification results was relatively high. This indicates that methods based on support vector machine algorithm， feature variable extraction and selection， and determining the optimal observation scale can effectively capture canopy information of different tree species and improve tree classification accuracy.