Individual Tree Crown Detection And Delineation Based On UAV Images

Individual Tree Crown Detection and delineation(ITCD)is of great significance to forest accumulation,biomass estimation,species identification and the construction of tree growth model.Unmanned Aerial Vehicle Remote Sensing technology(UAVRS)brings new opportunity for forest resource inventory.Based on the UAV images and its dense-matching point cloud,we aimed at extracting individual tree crown and studying a small-scale forest mapping scheme with controllable cost,which can be a supplement(or even partially as a replacement)to satellite,aerial photograph or LiDAR.(1)In high-resolution Unmanned Aerial Vehicle-Digital Orthophoto Map images(UAV-DOM)based ITCD research,spectral heterogeneity within individual tree crown is the biggest obstacle to tree crown identification process.We introduced a new method of image processing called bias field estimation,which can not only maintain lower spectral heterogeneity on the tree canopy image,but also keep its boundary integrity at the same time.Based on Local Intensity Clustering(LIC)algorithm and Marker-controlled watershed algorithm we succeed in extracting individual tree crown from 2 forest plot samples.The study compared different kinds of watershed algorithms(e.g.,A combination of internal and external markers,only internal markers and unmarked conditions)when they were used to extract individual tree crown.The results of accuracy assessment experiment:watershed segmentation algorithm with internal markers achieved a relatively higher precision(F-score?90%).At the same time,we conducted an extended experiment for the algorithm in 5 complex forest scenes(Plot 1 to 5).We further verified the robustness of the algorithm(F-score?65%).(2)UAV images matching point cloud based research can be divided into 4 processes including Rasterized Canopy Height Model(RCHM)construction,individual tree crown detection and delineation(ITCD),3D segmentation of the tree crown and individual tree structure parameter extraction.The construction of RCHM is based on a rasterization process of normalized image matching point cloud.Based on Progressive Morphological Filter,Universal Kriging method,"Normalized point cloud Rasterization method" and Local Maximum-Mean Filter,we obtained the Rasterized Canopy Height Model;Further we used Local Maximum Filtering and an Improved Seeded Region Growing(ISRG)algorithm to detect individual tree's position and delineate the contour of its crown during the ITCD process;At the same time,tree crown can be reshaped by the image point cloud within a reference of the convex hull(contours of the detected tree crowns).From the normalized point cloud which has got rid of the terrain effect we can extract individual tree location,height,crown width and other parameters.Results show that the detection rate of individual tree detect algorithm is more than 75%;Producer accuracy is more than 85%;User accuracy is no less than 75%;The absolute of the relative area error is less than 9%and F-score in both Samples are more than 85%;(3)A multi-level and multi-angle accuracy assessment scheme was proposed.Based on the hand-painted contour as the reference,we proposed a gradually refining framework for accuracy assessment from stand to the individual tree,from the point precision to the precision of the polygon.It means that we used point accuracy in forest plot scale,polygon accuracy in forest plot scale,point accuracy in individual tree scale and polygon accuracy in individual tree scale to achieve the accuracy assessment process of ITCD.At the same time,we conducted a regression analysis and a residual test between the reference tree crown diameter and the estimation results.Individual tree crown diameter extraction accuracy was up to 0.06m(RMSE).