Abstract: Rapid and accurate identification of tree species and their distribution patterns is the basis and premise for forest resource management and biodiversity conservation. Compared with the traditional field investigation methods, the rapid development of the surface of remote sensing technology in recent years can obtain high resolution hyperspectral remote sensing images flexibly, efficiently, and conveniently. However, how to select features with large amount of information and low redundancy from many features containing rich information for automatic tree species identification is an urgent problem. We used spectral sensors carried by unmanned aerial vehicles to take hyperspectral images, which covered an area of 6 hm² of the 25 hm² temperate mixed conifer-broadleaf plot in Changbai Mountain. Six canopy tree species, including Pinus koraiensis, Ulmus davidiana, Quercus mongolica, Fraxinus mandshurica,Populus ussuriensis, and Tilia amurensis, were selected as the field labeled tree species. Realtime Kinematic Phase Difference (RTK) technology was used to accurately capture the position of those target tree species. In addition, visual interpretation of the image of the study area was performed using ArcGIS based on the forest re-inventory results in 2019. Three classification methods, including convolutional neural network, maximum likelihood and Mahalanobis distance, were used to analyze the automatic classification of canopy tree species. Our results showed that: (1) The overall accuracy and Kappa coefficient of tree species classification of convolutional neural network (99.85%, 0.998) were better than maximum likelihood (89.11%, 0.86) and Maharanobis distance method (79.65%, 0.75); (2) Among the three classification methods, the classification accuracy of single dominant tree species was the highest when using the convolve neural network, and the highest classification accuracy of P. koraiensis, U. davidiana, Q. mongolica, F. mandshurica, P. ussuriensis and T. amurensis were 100%, 99.9%, 99.9%, 99.8%, 99.8% and 99.5%, respectively. (3) Overall, convolutional neural network had the lowest degree of mixing, while Mahalanobis distance method had the most serious degree of mixing problem. This study indicated that the convolutional neural network model based on deep learning approach can obtain the accurate and efficient classification of canopy species in natural temperate forests, which could provide a great step forward into the species diversity monitoring and forestry resource survey.

Key words: natural forest, hyperspectal image, convolutional neural network method, maximum likelihood method, Mahalanobis distance method.