Estimation Of Forest Aboveground Biomass Based On Multi-source Remote Sensing Data

Forest aboveground biomass(AGB)is one of the most important basis for estimating forest carbon,which has a significant influencation for assessing ecosystems.To improve the precision of ABG estimation,it is necessary to make effective use of the advantage of multi-source remote sensing data,select appropriate models and realize scale extrapolation.This research was mainly divided into two parts.The fist part is the AGB estimation in flight area.To select the optimal model,we compared six kinds of regression model with selected features based on different remote sensing data(airborne Lidar,combination of airborne LiDAR and airborne hyperspectral data).The second part is the AGB estimation in a larger area.With the AGB estimation results of flight area were taken as the training samples,we conducted a comparative analysis of different feature combinations with SVR based on Landsat8 data.The optimal feature combination was selected to calculate forest aboveground biomass in the up-scaling area.The main conclusions are as follows:(1)The LiDAR structural metrics could provide a high accurate AGB estimates(RMSE values in 16.78?25.98??(6(6,rRMSE values in 20.03?29.82%,R~2 values in 0.42?0.66).The combination of LiDAR and hyperspectral data has a better result(RMSE values in 15.94?28.87??(6(6,rRMSE values in 19.12?32.72%,R~2 values in 0.38?0.70).(2)The Leave-One-Out validation showed that Multiple Linear Regression had the worst result(CV-RMSE=25.98??(6(6,CV-rRMSE=29.82%,CV-R2=0.42),while SVR had the best(CV-RMSE=16.78??(6(6,CV-rRMSE=20.03%,CV-R2=0.66)based on Lidar.After combining LiDAR and hyperspectral data,MLR performed even worse(CV-RMSE=28.87t?ha,CV-rRMSE=32.72%,CV-R2=0.38).The accuracy of other models was improved,and SVR still performed best(CV-RMSE=15.94t?ha,CV-rRMSE=19.12%,CV-R2=0.70).Finally,SVR was selected as the best model in this study.(3)Base the navigation area,forest aboveground biomass estimation results and Landsat8data has is used to determine the optimal model,we compare different feature combinations(texture index+vegetation index,texture index+vegetation index+terrain factor,texture+vegetation index+surface parameters+terrain factor,and Boruta feature set).10 fold cross-validation method showed that compared with the texture+vegetation index feature set(R2=0.508,RMSE=16.87t/ha),the estimation accuracy is slightly improved after terrain and surface parameters are added(R2=0.515,RMSE=16.76t/ha).The best estimation result of feature subset obtained based on Boruta algorithm(R2=0.553,RMSE=16.09).(4)The feature subset selected by Boruta algorithm was used as the input variable,then we validated the results by 18 filed plots and found that SVR result was relatively good(R2=0.55,RMSE=11.36t/ha).