Tree Species Classification And Biomass Estimation Using Multisource Precise UAV Data In A Subtropical Forest Of Northwest Yunnan Province In China

Accurate forest tree species identification and biomass estimation are important research topics in biodiversity monitoring and forest resource management.The composition of forest tree species is an important index to describe the forest ecosystem,and forest biomass plays an important role in describing the structure and function of the forest.They provide key information for the study of the structure and function of the ecosystem.Through the current advanced UAV technology,it is possible to accurately obtain multisource precise remote sensing data such as point clouds and images.It has rich forest structure and spectral information,and it is of great meaning for identification of main forest dominant tree species,extraction of above-ground biomass（AGB）and the applications of forestry and ecology.This paper took Xiangguqing,a town in the southern part of the Baima Snow Mountain National Nature Reserve in northwestern Yunnan,as the research area.High overlap,high spatial resolution multi-spectral and RGB images acquired by fixed-wing drones.First,a large number of single images obtained by aerial digital photogrammetry were stitched into multi-spectral and RGB orthoimages,and digital aerial photography（DAP）point clouds were generated through structure from motion（Sf M）algorithm;Then use the digital elevation model（DEM）generated by Li DAR point cloud filtering and interpolation to achieve DAP point cloud normalization;Next,compare different individual tree segmentation algorithms（ie point cloud segmentation（PCS）,multi-resolution segmentation based on RGB images（MRS）and multi-resolution segmentation combined with RGB image and canopy height model（CHM）image）on the extraction effect of individual tree canopy;Finally,based on the metrics of spectral,texture and point cloud structural,the random forest classifier was used to classify the tree species（typical dominant trees）and forest types（ie,coniferous forests and broad-leaved forests）and precision evaluated.The multi-scale inversion models were constructed to estimate and map the spatial distribution of the forest AGB in the study area,and explored the effect and application value of the precise canopy spectral,texture and structural metrics of subtropical forests in northwestern Yunnan based on drones on the classification of dominant tree species and estimation of AGB.The research indicates:（1）Comparison of individual tree segmentation methods based on different data combinations and parameter sensitivity analysis results show that:the multi-resolution segmentation combined with RGB orthophoto and CHM extracted the highest crown width accuracy（F₁ value reaches82.5%）,followed by individual tree segmentation based on point cloud（F₁ value reached 79.6%）,and multi-resolution segmentation based only on RGB imagery had the lowest accuracy（F₁ value78.6%）.（2）Combined with the best single tree segmentation algorithm,the spectral,texture and point cloud structural characteristics were comprehensively extracted in the crown,and the random forest model was used to classify the object-oriented tree species.The results showed that the classification accuracy of according to the forest type（ie,coniferous and broad-leaved forest）was higher than the classification accuracy of dominant tree species（ie coniferous:Pinus yunnanensis,Tsuga chinensis（Franch.）Pritz.,Tsuga chinensis（Franch.）Pritz.;broad-leaved trees:Cyclobalanopsis oxyodon.,Quercus aliena Blume,Acer forrestii,Alnus nepalensis D.Don,Populus davidiana）;The random forest model with integrated spectral,texture and structural metrics had the highest classification accuracy（overall accuracy was 80.20%,Kappa coefficient was 60.37%）;In the classification process,the classification accuracy using only the spectral features was low,indicating that the texture features and point cloud structure features could improve the classification of tree species（the overall accuracy and Kappa coefficient were increased by 1.49-4.46%and 2.86-6.84%,respectively）,and the model that combines structural,spectral and texture metrics had the highest classification accuracy（the overall accuracy was69.34%,and the Kappa coefficient was 58.53%）;Among the classified models,spectral features（RGRI,GNDVI,etc.）,texture features（correlation degree of blue band,etc.）,structural features（H95,H50,D7,etc.）were at the forefront of the importance of the model,indicating structure,spectrum and texture features all had a positive effect on training the classification model.（3）Combining the spectral and point cloud structure characteristics,the multi-scale linear regression model was used to estimate the forest AGB in the study area.The results showed that the accuracy of the aboveground biomass model predicted by the normalized DAP point cloud structure characteristics（R²=0.77）,RRMSE=10.78%)was higher than the model that only uses spectral features to predict（R²=0.52,r RMSE=15.46%）;the prediction model combining point cloud structural features and spectral features had the highest accuracy（R²=0.81,r RMSE=9.72%）,The structural characteristics of DAP point cloud played an important role in improving the accuracy of the AGB prediction model;the AGB inversion mapping of the study area effectively obtained the heterogeneity of the spatial distribution of biomass.