Object-based Tree Species Classification And Identification Using Unmanned Aerial Vehicle Remote Sensing Imageries In A Subtropical Evergreen Deciduous Broad-leaved Mixed Forest

With the recent development of remote sensing technology,multi-source remote sensing data have been widely used in forest tree species classification,making it possible to survey and monitor plant diversity at broad scales.Due to the limitation in the collection and data processing of high-resolution remote sensing imageries,most studies are focused on temperate or boreal forests in Europe and North America.Few studies have been conducted on tropical or subtropical forests with high biodiversity and complex topography.In addition,phenological features of tree species are often ignored.There are many subtropical mountain forests in China,with high plant species diversity,seasonal changed obviously,and is the home of many relict species and endemic species.Therefore,accurately classifying tree species in subtropical forests is of great significance to forest monitoring,protection and management.In this paper,high-resolution(~ 5 cm)UAV(Unmanned Aircraft Vehicle)RGB imageries were used to identify six common tree species,snag and canopy gaps in an evergreen and deciduous broad-leaved mixed forest covering around 100 ha in area in Mt.Tianmu,Zhejiang province.Using the object-based multi-resolution segmentation,I established a feature space containing spectral,spatial geometry and texture features.Four classification algorithms,including K-nearest neighbor(KNN),classification and regression tree(CART),support vector machine(SVM)and random forest(RF),were selected for tree species classification.Then,I analyzed the potential of UAV RGB remote sensing imageries with high spatial resolution and low cost in the classification and identification of subtropical forest species in a large area,and discussed the contributions of phenological information and texture features in improving the accuracy of species classification.Meanwhile,we divided the study area into three different habitat types based topographic data,to evaluate the influence of complex topography in tree species identification.The results show that:(1)The UAV RGB remote sensing imageries with high spatial resolution and low cost has a large potential in subtropical forest species classification and identification in a large extent,and the combination of KNN and spectral,spatial geometry features achieved the highest classification accuracy(overall accuracy = 83.30%,Kappa coefficient = 0.799),followed by the combination of RF and spectral,spatial geometry,texture features(overall accuracy = 83.13%,Kappa coefficient = 0.798).Overall,CART performed the worst.(2)Spectral and phenological features contributed the most to tree species classification,followed by spatial geometric features,while texture features contributed different among classification algorithms.Texture features improved the overall accuracy by about 4%in CART and RF,while reduced the accuracy by about 5% in KNN and SVM.One reason is that the spectral/phenological features of the UAV RGB imageries were rich,while the texture features were fuzzy.The drone imageries I used were acquired in a special phenological period when the colors of tree species was greatly different,and the flying distance for the UAV to the forest surface was high during the survey.(3)The quality of tree classification for conifer species(except for Pinus Taiwanensis)were better than broad-leaved tree species.Liquidambar acalycina,Cryptomeria fortunei,snag and canopy gap can be accurately identified in all algorithms.The differences of species classification were mainly due to the biological characteristics of these species.The differences in leaf colors and tree morphology at different phenological stages were one main reason.The difference in the quantity and quality of ground georeferenced sampling points of different tree species was another reason.(4)The complex topography in the study area had a significant impact on tree species classification and identification,and different tree species had different responses to topographical variables.The performances of the same tree species in three habitat types were different,and the species classification accuracy was generally higher in the upper slope,especially for Cryptomeria fortunei,Ginkgo biloba and Pseudolarixamabilis.In conclusion,in subtropical forests with complex terrain and high biodiversity,using UAV RGB imageries for tree species classification is a low-cost and effective method,and suitable for biodiversity surveys in large geographic coverages,especially in mountainous areas with complex terrain.This method can be greatly useful for forest resource survey and the studies of species spatial distributions,and provides critical data basis for biodiversity monitoring and conservation.