Study On Extraction Of Tree Measurement Factors By Muliti-platform And Multi-source Remote Sensing

Forest resources survey and monitoring provides data support and important evidence for regulating and managing national forest resources.And its work efficiency and data quality are mainly determined by tree measuring insturments and methods.In fact,the present instruments have many drawbacks such as its backward data collecting means,high cost and heavy workload of indoor and field,and the traditional manual investigation method could hardly meet the requirements of today’s foresty development because of its time-consuming and inefficiency.With the progress of scientific technology,especially the photogrammetry technology,three dimensional laser scanning,remote senseing and unmanned aerial vehicle(UAV)technology,the growing number of foresty workers have been focusing on how to obtain the information of forest resources more efficiently by using advanced technologies.In this paper,foucsing on "forest resources survey and technology",we studied the scientific extraction of tree measurement factors using the data obtained from different kinds of equipments which were designed on ground,UAV and remote sensing satellite due to various demends of forest resource investigation and forest mensuration.This study,in the field of forest resources inventroy,not only tries to improve its traditional infromation retrieval technique and efficiency,and drop its cost,but also aimes to provide theory basis and technique support for the implement of digitization,automation,intelligentize and integration of it.The main research content and conslusions in this paper are as following:(1)The extraction of tree measurement factors based on ground platformThe DBH(diameter at breast height),tree height and stand structure parameters were measured using single laser photograhpy dendrometer which was designed by us independently with the application of single-image photogrammetric principle.And the results showed that the MAE(mean absolute error)of DBH and tree height were 0.6cm and 0.37m,respectively,the mean accuracy was 97.45%and 97.18%,the mean accuracy of uniform angle index and neighborhood comparison were 97.50%and 97.14%,respectively,and the measuered mingling degree was basically equal.We measured DHB,tree height and volume of standing tree using CCD smart station single-image photogrammetric and multi-image merging method,and verified the factors that affect the measurement accuracy of the insturments in detail from the aspects of instrumental error,distance,angle and photography mode.And the results showed the accuracy of tree height and volume of standing tree,meausred by the abovementioned method,were 98.41%and 98.01%,the optimalizing distance and angle were 15m and 0°,respectively.Besiedes,we desigened a 3-dimensional ground marking balls,and proposed a continuous photographic observation project using samrt phones’ "annular photography",which could help us restore the 3D scenes of a flat plot and extract the DBH of standing tree and its coordiante position.After the analysis of the restore of different "annulars"with one-loop and composite pattern,we found the composite mode with the radii of 12m/8m was the preferred embodiment.The experiment confirmed that the MAE of DBH was 1.96cm and the accuracy was 90.27%.Furthermore,a low cost hand-hold 2-dimensional Lidar scanning dendrometer was designed using Lidar scanning principle and SLAM(simultaneous localization and mapping)technology.With this device,not only can we measure pith coordinates and trees density in a continuous moving mode,but also measured the DBH through single station mode.The results indicated that the point position error of pith coordinates was constantly less than 5cm,the MAE of DBH was 1.06cm,and the mean accuracy was 96.20%.(2)The extraction of tree measurement factors based on UAV images and Lidar point cloud data.In this part,the ordinary digital camera and Lidar as sensors,based on the UVA(unmanned aerial vehicle)sensing system,to carry out flight experiment at the Wanshan and Xinxing forest farm,the pilot site,adminstered by Weihe Forestry Bureau of Heilongjiang Province.The UAV high overlapping images(the overlap rate in flight direction was 80%,and the rate in lateral direction was 60%)and high density Lidar point cloud(>50 point/m2)in this area were captured.The UAV Lidar point cloud help us extract the single tree factors,such as tree height,crown breadth and position coordinates,and the stand factors,including canopy density,stand mean height,stand density,leaf area index and stand volume,under subcompartment scale.In addition,the Lidar point cloud data was preprocessed by Lidar360 software,which created DME(digital evalution model),DSM(digital surface model)and CHM(canopy height model).Using method based on CHM or point cloud to extract single tree factors and comparing the results obtained from these two methods found the accuracy of individual tree segmentation was basically equal,and the segmentation method based on CHM(the average F-score of low canopy density coniferous forest,medium or high canopy density coniferous forest and medium or high canopy density broad-leaved forest were 0.94,0.76 and 0.57,respectively)was slightly better than point cloud method(the average F-score of low canopy density coniferous forest,medium or high canopy density coniferous forest and medium or high canopy density broad-leaved forest were 0.90,0.70 and 0.55,respectively);the average accuracy of tree height and crown breadth obtained from point cloud extraction(95.38%and 74.09%)was superior than those obtained from CHM extraction(90.42%and 68.01%).A series of methods,including CHM height threshold method,point cloud extraction and Beer-Lambert Law,were applied to obtained gross crown-area,canopy density,stand average height,leaf area index,and point cloud height,density and intensity;then took these parameters as variables in which the canopy density and leaf area index were made independent variables.Finally,the multiple linear regression equation was founded using the relationship between stand volume and variables,which helped us deduce the UAV Lidar optimization model:lnM=0.4901nLAI+1.2261nh30+1.296,where R2 is 0.84,RMSE is 16.78m3/ha,and rRMSE is 11.39%.The intensive SfM point cloud was created by feature point extraction,matching and encryption from high overlapping images,and then the DOM,DSM,DEM and CHM were obtained.Using object-oriented multiresolution segmentation,combined with DOM and CHM,to extract tree height and crown breadth.On the contrary,using CHM generated from SfM point cloud and partial Lidar ground point cloud to extract canopy density,average tree height and mean crown breadth.Similarly,took CHM and canopy density(CC)as independent variables to bulid non-linear average height model:H=2.521CHM0.646 CC-0.997,where R2 is 0.722,and RMSE is 1.497m.At last,the UAV photography stand volume inverting model was constructed in which all four parameters were made independent variables,the equation was M=0.277H2.629C-0.423CC1.655,where R2 is 0.737,and RMSE is 21.72m3/ha.(3)Landsat8 remote sensing image inversion model of stand volumeTaking ground survey stand volume as samples,combined with Landsat8 remote sensing data,to invert stand volume in the scale of foresrty stations.The Landsat8 OLI extracted feature variables and stand volume inverting model was built by traditional multiple linear stepwise regerssion method and machine learning least square method which was optimized from random forest and quantum-behaved particle swarm algorithm.The results demonstrated that machine learning method was superior than the traditional method(R2=0.58,RMSE=37.56m3/ha)in accuacy evaluation,the optimal stand volume model was random forest model with R2=0.75 and RMSE=29.393/ha.Therefore,the evaluation of stand volume and the mapping of its distribution were based on the random forest model.In this work,we study the dendometer factors extracted from differnet sources which was obtained from 3 different sacles of many platforms,which showed great practical significance for the forest resources survey and monitoring.The main techniques proposed in this paper not only had important meaning on improving our level on forest survey and monitoring,and reducing labour intensity and production cost,but also provided guidence and reference significance to meet the requirements of forest resources investigation from different scales,accuacy and budget cost,while obtaining and updating of the forest resources information more precisely and efficiently.