Research On Urban Forest Tree Species Classification From UAV Image Based On Deep Learning

Forest resources are an important part of the earth's ecological resources.With the rapid development of the city,the urbanization process is gradually accelerating.The concept of building an ecological city and urban green development has become a hot spot of global concern.Urban forests are indispensable in urban ecosystems.They play an important role in purifying urban air,regulating urban climate,improving residents' living environment,and afforesting and beautifying urban environment.Reasonable tree species are the basis for building a stable urban forest system.Therefore,studying tree species and extracting tree attribute information is of great significance to urban forest construction.The traditional urban forest resource survey method is time-consuming and labor-intensive.The application of remote sensing technology saves the time of forest resource survey and reduces the cost of survey.As a high-resolution remote sensing image acquisition platform,UVAs provide more possibilities for the study of urban forests.The canopy is an important part of discriminating tree species.It is essential in the study of tree species.The canopy image is obtained through the UAV remote sensing platform,and the classification of tree species is gradually emerging.However,the lack of urban forest canopy image data sets limits the development of new methods,especially deep learning methods.In order to realize the classification and recognition of the deep learning method on the canopy image,this paper mainly carried out the following aspects:(1)Using the UAV to obtain the images of the urban forest canopy,a tree canopy classification data set TCC-10(Tree Canopy Classification)is proposed.The data set contains two types of image data: a simple background canopy image and a complex background canopy image,with a total of 19302 canopy images.In addition,the data set covers different canopy images of different seasons and has certain research reference standards.The experiment uses two methods of adjusting brightness and rotation to enhance the data.(2)Based on AlexNet,VGG-16 and ResNet-50,the two types of image data in the data set were tested,and the direct training method was used to prove the effectiveness of the deep learning method on the TCC-10 data set.It indicates that the TCC-10 data set can be used as a data benchmark for future urban forest tree species research.Compared to AlexNet and VGG-16,ResNet-50 exhibits better classification accuracy.(3)Based on the structure of ResNet-50,an effective residual network model is designed.Under the same training steps,the model achieved a total classification accuracy of 94.4%,and the classification accuracy of deep learning methods is significantly higher than that of traditional image classification algorithms.Among the classification accuracy of each category,improved residual network also has better performance.Among the 10 types of tree species,7 categories have higher classification accuracy than AlexNet,VGG-16 and ResNet-50.It shows that improved residual network plays an important role in improving the accuracy.