Retrieval And Estimation Research Of Forest Parameters Based On Digital Aerial Photograph Data

As the most important part of terrestrial ecosystem,forest plays an important role in maintaining ecosystem balance,purifying air,carbon sequestration,water and soil conservation.Forest aboveground biomass(AGB)and leaf area index(LAI)are two important parameters for evaluating forest growth and health.Individual tree structure parameters of forest are very important for management of fine forestry and construction of modern precision forestry.Traditional forest resources survey is based on intensive labor field measurements.For field investigation,not only the work load is large,work cycle is long,but also it is very difficult to quickly obtain forest parameters of large area,and not to achieve a wide range of continuous sampling measurements.Remote sensing technology provides a quick and efficient estimation method for forest resources investigation,especially for estimation of forest AGB and LAI,and extraction of individual tree structure parameters.Passive optical remote sensing with a wide range of continuous imaging can provide abundant vegetation spectral indices and textural features information,and active remote sensing technology such as light detection and ranging(LiDAR)has strong detection capability for forest height and canopy vertical structure.Both of them have been successfully applied in quantitative retrieval of forest parameters.In recent years,researchers combined LiDAR with optical remote sensing data to retrieve forest parameters,spectral and textural feature of optical remote sensing data can be used,and the advantage of Li DAR to detect forest canopy vertical structure also can be fully utilized.Combination of both can improve the retrieval ability of forest parameters.Digital aerial photograph(DAP)data is processed based on structure from motion(SfM)algorithm and regional net adjustment method to generate digital surface discrete point clouds similar to LiDAR and digital orthophoto mosaic(DOM)similar to optical remote sensing image.With the construction of high-resolution earth observation space infrastructure in our country,the number of airborne platforms and sensors in China has been increasing.Optical remote sensing based on light,small aircraft and Unmanned Aerial Vehicles(UAV)platform--with low cost,high resolution,fast and efficient data acquisition--is booming in the different fields,and its business application level has been increased rapidly.This study area is located in Mengjiagang forest farm of Jiamusi city,in Heilongjiang Province.Using digital aerial photograph-and LiDAR data acquired by a Y-5 light,small aircraft and ground survey data to carry out retrieval and estimation of forest parameters,and obtain the following research results:(1)To obtain the DOM,canopy height model(CHM)and normalized photography point cloud(PPC)of research area.Aerial optical data was processed using SfM algorithm and regional net adjustment technology to generate digital surface model(DSM),discrete dense point clouds and DOM.The absolute heights of DSM and discrete point clouds were normalized by substracting the terrain heights from digital elevation model(DEM)of LiDAR to obtain the relative heights,two new products were CHM and PPC.By comparing and analyzing,the mean distance deviation between PPC and the LiDAR point cloud was within 0.5m.(2)To constructe the retrieval model of forest AGB and LAI,and making15m-resolution distribution maps of AGB and LAI.Visible vegetation indices and textural features were extracted from DOM data,and forest height and coverage features were extracted from PPC data.AGB and LAI were retrieved by machine-learning regression models(Cubist,KNN,RF,SVR)based on only DOM data,PPC data alone and combination of both.15 m-resolution distribution maps of AGB and LAI in the study area were made using those features from DOM data,respectively.The results showed that combination of DOM and PPC data had the highest retrieval accuracy for AGB and LAI.Forest height features contributed more to AGB retrieval,and the forest canopy coverage and vegetation index feature contributed more to LAI retrieval.The Cubist model in four models was the best one for AGB estimation,and the SVR model was the best for LAI estimation.The best results of AGB and LAI estimation were R~2=0.73 and RMSE=20 t/ha,R~2=0.79 and RMSE=0.48,respectively.(3)To completing the individual tree segmentation.CHM data were segmented using marker control watershed algorithm,and the regional growth algorithm was used to segment the PPC data.The tree positions and heights were extracted from the results of segmentation.Combining results of segmentation with the field measurements of sparse big trees,dense big trees and dense medium-young trees to determine the accuracy.The complete rate of segmentation(r),precision(p)and segmentation quality(q)were used to evaluate the results.The results showed that the segmentation results of PPC data were slightly better than those of CHM data.The accuracy of individual tree segmentation of sparse big tree plots was the highest,and the r,p and q were 90.41%,82.50%,75.85%.Dense medium–young tree plots was the lowest,and the r,p and q were 57.01%,79.22%and 49.59%respectively.(4)To complete the estimation of individual tree height.First,the individual tree heights were extracted according to tree positions from CHM and PPC data.Then,the linear regression relationships between estimated tree heights and true tree heights were established,and the tree heights were divided into different grades.The results showed that most of young trees were over-estimated and most of big trees were under-estimated,with an average estimate value of about±3.5m.The R~2 and RMSE of individual tree heights estimation based on CHM and PPC were 0.667 and 2.83m,0.661 and 2.36m,respectively.In addition,the linear regression formulas between estimated and true tree heights were established according to the flight overlap.The results showed that the estimation accuracy of high overlap area was higher than that of low overlap area.