Identification Of Urban Tree Species In UAV RGB Images Based On FC DenseNet

Urban greening is an important elements in a city.In the construction of a city,a full understanding of the spatial distribution of greening information is conducive to scientific decision-making and management of green space,maximizing service provision,and reducing negative impacts.Currently,greening extraction is usually obtained through field survey of individual trees,which is relatively intuitive,but requires significant human and material resources,as well as time costs.Therefore,combining remote sensing images,in-depth learning,and other means with the purpose of obtaining urban greening may provide efficient means for obtaining spatial distribution of greening information.This article hopes to identify urban greening in Changsha through a relatively complex deep learning algorithm FC-DenseNet model based on UAV near-ground remote sensing images.Through three precision images(three flight heights)and three segmentation accuracy(landscape elements,trees,shrubs,and trees),it is intended to detect the recognition accuracy of greening trees under complex algorithms,comparing with previous research papers,Provide flight altitude reference and experimental method considerations for future near-ground remote sensing collection,and add data to the orthometric projection dataset for tree species identification,providing data set support for subsequent research.In response to the above issues,this article first obtained near-Earth remote sensing data from a total of 13 regions in Yuhua District of Changsha through drone aerial photography of four types of green space-park green space,residential land ancillary green space,green space within roads and transportation facilities,green space for public management and public service facilities-including 5 types of landscape elements and 21 types of tree species as research samples,and then through data synthesis,labeling,clipping,partitioning,and enhancement,Build training and validation data sets for landscape elements,trees,shrubs,and trees.After that,the FCDenseNet algorithm framework was built,and the dataset was input for recognition experiments.The experimental results show that the overall val/m_iou based on scene element recognition is 79.1%for 500m images,80.5%for 250m images and 81.7%for 120m images;The overall val/m_iou based on tree,shrub and grass recognition is 76.5%for 500m images,82.7%for 250m images and 84.1%for 120m images.After performing variance analysis using SPSS,it can be concluded that there is no significant difference in the recognition results of landscape element recognition at three heights,while there is no significant difference between the 250 m image recognition results of trees,shrubs,and grasses and the 120 m image recognition results.The results based on the overall identification of tree species show that the three flight altitudes do not have the ability to identify tree species.When the number of tree species identification tags is reduced to 3,the 120m aerial image can identify 9 kinds of plants such as Cinnamomum camphora,Osmanthus fragrans,Platanus orientalis,Magnolia grandiflora,Koelreuteria paniculata,Elaeocarpus glabripetalus,Photinia serrulata,Sophora,Cherry and so on with limited interference of other tree species.Among them,the overall val/m_iou of Cinnamomum camphora,Osmanthus fragrans and Platanus orientalis groups is 63.7%,the overall val/m_iou of Magnolia grandiflora,Koelreuteria paniculata and Elaeocarpus glabripetalus groups is 69.8%,and the overall val/m_iou of Photinia serrulata,Sophora and Cherry groups is 58.7%.ased on the comparative analysis of FC-DenseNet generated forecast map and manual labeling,it can be concluded that after reducing the label grouping recognition,the 120m aerial image can be used to identify 9 plants,including camphor,osmanthus,magnolia,sycamore,lupine,eucalyptus,photinia,sophora,cherry,and so on,with a limited amount of other tree species interference.Other landscape tree species may expect to further learn recognition at lower flight altitude or improved algorithms for landscape tree species.Compared with previous literature,this article can use drones as a data collection tool compared to other landscape element papers.Although the recognition accuracy is slightly lower,compared to identifying buildings and water bodies separately,it can analyze five types of landscape elements simultaneously,and has good application value;Compared to other papers on trees,shrubs,grasses,and tree species,this article can collect and recognize images with relatively complex backgrounds at near ground heights,perform multi label recognition,and obtain relatively good recognition results,which has certain reference significance for future research.