Research On Forest Aboveground Biomass Estimation Based On Airborne LiDAR Data

Aboveground Biomass(AGB)is an important indicator for measuring the productivity of ecosystems,and is also the basis for studying the material cycle of forest ecosystems.LiDAR technology can not only obtain the vertical structure information of the forest but also has the advantage of no signal saturation.Machine learning methods can improve the prediction accuracy of forest aboveground biomass,and can also overcome the shortcomings of poor generalization ability of traditional regression equations.There are many feature variables when using LiDAR technology and machine learning methods to predict forest above-ground biomass.Different feature variables will lead to differences in prediction accuracy.Most previous studies used only one feature selection method.Therefore,this paper takes the Penobscot Forest in Maine,USA as the research object,selects the percentile density and percentile height of the forest as the characteristics of estimating forest above-ground biomass,and uses Spearman correlation coefficient method and LightGBM(Light Gradient Boosting Machine)algorithm and random forest-recursive feature elimination method to filter features,then use linear regression model,random forest algorithm and support vector machine algorithm to predict the forest above-ground biomass in the plot with Python and Scikit-learn programming platform.The purpose is to compare and analyze the better methods to provide reference for forest resource investigators.The research contents and results of the thesis are as follows:(1)Use Spearman correlation coefficient method,LightGBM algorithm and random forest-recursive feature elimination method to screen a total of 22 features of percentile height and percentile density.In the results of the Spearman correlation coefficient method,the correlation between biomass and height features is above 0.5,and only two of the density features have a correlation with biomass above 0.5,and the rest are weak or negative correlations;In the results of LightGBM algorithm and random forest-recursive feature elimination method on the importance ranking of height features and density features,the importance of height features is obviously greater than that of density features.Studies have shown that:biomass has a strong correlation with the height characteristics of the forest;but has a weak correlation with the density of the forest.(2)Combined with the characteristics of screening,linear regression,random forest algorithm and support vector machine algorithm were used to predict forest aboveground biomass.The prediction accuracy of the three feature selection methods corresponding to the random forest algorithm all get the maximum value.Research tells that the random forest algorithm is superior to the other two methods.And the number of training samples also has an impact on the prediction accuracy.The number of training sets is generally 60%-70%of the sample size.(3)Using linear regression,random forest algorithm and support vector machine algorithm to predict forest biomass in combination with all features,the corresponding prediction accuracy R2 is 0.7119,0.7921 and 0.7595,respectively,and the selected features corresponding prediction accuracy R2 is 0.7679,0.8679 and 0.8594,respectively.Studies have demonstrated that the prediction accuracy of aboveground biomass in forests is not positively correlated with the number of features.Choosing appropriate features can help improve the prediction accuracy of biomass.(4)In the linear regression model,the average precision of the biomass predicted by the Spearman correlation coefficient method is R2=0.5278,the average precision of the biomass predicted by the LightGBM algorithm is R2=0.6974,and the average accuracy of biomass prediction for random forest-recursive feature elimination method is R2=0.6823;in the random forest algorithm,the corresponding average prediction accuracy R2=0.7314,0.8255,0.7877;in the support vector machine algorithm,the corresponding average prediction accuracy R2=0.6818,0.8166,0.7366.Studies have shown that:the features selected by Spearman correlation coefficient method have the worst biomass accuracy,and the features selected by LightGBM have the best biomass accuracy.