Dominant Tree Species Classification In A Natural Forest Of Northwest Yunnan Province Based On Deep Learning Networks And Multi-source Remote Sensing Data

Accurate information of dominant tree species and their spatial distribution are key ecological monitoring factors for accurately characterizing forest biodiversity,depicting tree competition mechanism and quantitatively evaluating forest ecosystem stability.Unmanned aerial vehicles(UAV)remote sensing has the advantages of high efficiency,flexibility and high spatial resolution,and satellite remote sensing can provide a wide range of forest spectral information.They provide great potential for the identification of dominant tree species types and fine spatial distribution in natural forests.In this paper,Gehuaqing area in the south of Baima Snow Mountain National Nature Reserve was taken as the research area,high resolution images(spatial resolution=0.23m)were obtained by RGB camera(Sony ILCE-6000)carried by fixed wing UAV(Mini Crosswind 1600),and sub-meter(spatial resolution=0.5m)multispectral images were obtained by using Super View-1 1A products.Firstly,the object-oriented multi-resolution segmentation(MRS)algorithm was used to segment individual tree crown from the UAV high-resolution RGB image and satellite multispectral image in the forests with different densities(low(547 n/ha),middle(753n/ha)and high(1040 n/ha)),and parameters of the MRS algorithm were tested and optimized for accurately extracting the tree crown and position information of the individual tree.Secondly,the texture metrics of the UAV high-resolution RGB image and the spectral metrics of the satellite multispectral image within the individual tree crown were extracted,and the random forest algorithm,three deep learning networks(light weight network Mobile Net V2,residual network Res Net34 and dense network Dense Net121)were utilized to classify the five dominant tree species(Nepal alder,Oriental white oak,maple,aspen and Yunnan pine)in the three typical plots(100×100m~2).And the performances were validated by using the field measurement data,the influence of the number of training samples on the accuracy of dominant tree species classification using three deep learning networks was also investigated.Finally,the tree species types classified by the three typical plots(100×100m~2)were used as the amplification data of the sample set of individual dominant tree species,so as to classify and map the dominant tree species of the Gehuaqing area(3×3km~2).The results showed that:(1)The MRS algorithm was used to segment the high-resolution RGB image and satellite spectral image in the forests with different densities,and the individual tree segmentation results of the dominant tree species in Gehuaqing area were shown:In the forests with different stand densities,the F1-score of individual tree segmentation based on UAV high-resolution RGB image were 75.4–79.2%,and that based on satellite multispectral image were 71.3–74.7%.The F1-score of individual tree segmentation accuracy of the dominant tree species in the Gehuaqing area based on high-resolution RGB image was 72.4%,and that based on multispectral image was67.1%.(2)Used the best individual tree segmentation parameters,combined with the spectral and texture metrics extracted from the crown width,then random forest algorithm and three deep learning networks were used to classify the dominant tree species in the three typical plots(100×100m).The results showed that:The overall accuracy of dominant tree species classification using light-weight network Mobile Net V2(OA=71.11–82.22%),residual network Res Net34(OA=78.89–91.11%)and dense network Dense Net121(OA=81.11–94.44%)were higher than that of random forest algorithm(OA=60.00–64.44%).The addition of texture metrics improved the classification accuracy of dominant tree species in typical plots,among which the overall accuracy of random forest algorithm was improved by 4.44%,and that of the deep learning networks were improved by 11.11–13.33%.By dividing the individual tree sample sets of each dominant tree species in proportion(20%,40%,60%and 80%),the three deep learning networks were trained for the classification of dominant tree species.The result showed that the changes of overall accuracy of dominant tree species classification of the three deep learning networks influenced by the number of training samples were 2.69–4.28%.The overall accuracy of the dominant tree species classification using three deep learning networks was gradually improved with the sample size increased.(3)The tree species types obtained from the classification of three typical plots(100×100m~2)were used as the amplification data of the sample set of individual trees,and the three deep learning networks were trained to classify the dominant tree species of Gehuaqing area(3×3km~2).The results showed that:dense network Dense Net121had the highest overall classification accuracy(OA=97.25%,Kappa accuracy=96.52%),indicated that the deep learning network combined with large sample data could accurately describe the spatial distribution of dominant tree species in the study area.