Inversion Of Forest Parameters Using Airborne LiDAR

The airborne LiDAR pulse can penetrate part of the forest canopy and obtain the three-dimensional structure information of the forest.It is one of the most active remote sensing technologies in forestry remote sensing.Although domestic and foreign scholars have achieved some good scientific research results in the inversion of forest parameters from airborne LiDAR,the theory and practice are far from mature,the core algorithm still needs to be improved,and the technical system is to be improved.Therefore,this paper takes the airborne LiDAR data and sample survey data of the Dayekou River Basin in Zhangye City obtained from the Heihe Integrated Remote Sensing Joint Experiment as the research object,systematically studies the technical process of airborne LiDAR using for forest parameter extraction.Thoroughly compared and analyzed,and the optimal algorithm and parameter settings in the inversion process are found out.A new LiDAR metric for inverting forest parameters is proposed in this paper.The main conclusions of the study are as follows:(1)The Cloth Simulation Filter algorithm(CSF)has better filtering effect in this study area than the commonly used Irregular Triangulation Filter algorithm(TIN)and Progressive Shape Filtering algorithm(PMF).By comparing the DEM generated by the three filtering algorithms with the results of the DGPS measurements in the test area,the CSF algorithm can obtain a DEM with a maximum error of 1.9 m and an average error of0.138 m/m~2,which can meet the needs of forest parameter inversion.(2)The Canopy Height Model(CHM)generated by Normalized Point Cloud(NPC)is better than the traditional CHM method(DSM-DEM)in terms of the performance of the three-dimensional structure of the canopy and the elimination of"hole"in the CHM.At the same time,comparing the three interpolation algorithms often used in the CHM generation process,Distance Weighting Method(IDW),Irregular Triangulation Interpolation(TIN)and Kriging interpolation,it was found that the simplest IDW algorithm is more suitable for CHM interpolation.(3)The CHM-based watershed single-tree segmentation algorithm and the point cloud-based spatial clustering algorithm are compared and analyzed.It is found that the latter cannot be used in this study because of the insufficient point cloud density.By orderly comparison of the parameters of the watershed algorithm,it is determined that the CHM of 0.5m pixel size can obtain the best single-wood segmentation effect,the single-wood segmentation rate reaches 68.55%,and the segmentation accuracy rate reaches57.46%.Although the parameter values of the algorithm need to be changed according to the point cloud density and the stand condition,the method proposed in this paper has certain reference value.(4)Inverting the stand parameters from the idea of single-wood segmentation,in the case that single-wood matching is not possible,this paper proposes a method to fit the measured values of stand and single-segment estimation with simple linear regression model.The weighted mean heighted estimate of the crown area was used to fit the measured weighted mean height(LorH)of the breast height area,the correlation coefficient(R~2)was 0.375,and the root mean square error(RMSE)was 2.16 m.Using the estimated average tree height and average crown amplitude,the natural logarithm transformation was used to linearly fit the breast height area(BA)and aboveground biomass(AGB),and R~2reached 0.571 and 0.639,respectively.(5)Retrieving the stand parameters from the plot-based ideas.This study extracted 43indicators from the point cloud and fitted the measured forest parameters through multiple linear regression,support vector machine and artificial neural network.The results show that the latter two are better than the commonly used multiple linear regression model,mainly because of the collinearity problem between the extracted indicators.After LASSO feature variable selection,for LorH and AGB predictions,the support vector machine is optimal,R~2is 0.784 and 0.849,respectively,and RMSE is 1.256m and 21.298t/ha respectively.For BA prediction,the neural network model is optimal,R~2 is 0.859,and RMSE is 3.134m~2/ha.(6)This paper proposed a plot-level LiDAR extraction index—the top of canopy height(TCH).The existing allometric growth model is used on the premise that the forest stand has the same allometric growth law as the single tree.The parameters of each forest stand are predicted.The fitting coefficients for LorH,BA and AGB are 0.129,0.696 and0.706 respectively.The prediction accuracy is higher than the inversion accuracy based on single-segment segmentation,which is lower than the plot-based forest inversion using machine learning.At the same time,this paper makes a scientific analysis of the CHM high threshold and CHM pixel size that affect the accuracy of TCH prediction.The method has certain reference significance.In a word,this paper systematically studied the method of extracting forest parameters from airborne LiDAR,and made a detailed comparison and analysis of the core algorithms.The conclusions obtained are reliable and effective,which will strongly promote the application of airborne LiDAR technology in forestry.