Estimation Aboveground Biomass Of Natural Secondary Forest Using Multi-source Data Based On Bias-corrected Random Forest

Forests are a component of terrestrial ecosystems.The research of ecosystem material cycles is of great significance to the global climate change.As one of the important indicators,the accurate estimation of forest aboveground biomass(AGB)plays an important role in the estimation of forest ecosystem carbon storage,understanding of the global carbon cycle mechanism and response to climate change.Using 142 fixed sample plots of forest resources continuous inventory in the Maoershan Forest Farm of Northeast Forestry University,Airborne Laser Scanning(ALS)and Landsat8 OLI image as the data sources,46 feature variables were extracted and then selected,multiple stepwise regression(MSR),Support Vector Machine(SVM),random forest(RF)and three bias-corrected random forest((Bias Corrected,BC),(Residual Rotation,RR)and(Best Angle Rotation,BAR))were used to construct estimation models of forest AGB.Adjusted coefficient of determination(Radj2),root mean square error(RMSE)and relative root mean squared error(%)(rRMSE)were applied to evaluate the model accuracies.Then the best model was used to estimate forest AGBof Maoershan forest farm area.The results showed that:(1)Applying multi-source remotely sensed data yielded higher accuracy than applying single data source alone,while non-parametric models(RF(Radj=0.68,RMSE=49.71t·hm-2,rRMSE=32.48%)and SVM(Radj2=0.64,RMSE=52.80t·hm-2,rRMSE=35.28%))yielded better model performances than traditional MSR model(Radj2=0.52,RMSE=57.29 t·hm-2,rRMSE=43.26%).(2)The optimal RF estimation model is selected for different bias correction,and the performance of BCRF is rRMSE=21.84%,with the best biomass estimation among these models(rRMSE reduced 10.64%compared to RF model);The performance of RR is rRMSE=28.27%,rRMSE reduced 4.21%compared to RF model;The performance of BAR is rRMSE=25.46%,rRMSE reduced 7.02%compared to RF model.(3)The non-parametric models(SVM,RF,BCRF,RR and BAR)had the greatest improvement for the AGB of 100～200 t·hm-2(the RMSE reduced from 48.87 t·hm-2 to 13.72-23.55 t·hm-2 compared to MSR model,and rRMSE reduced from 28.15%to 8.69%～16.13%).In particular,when the AGB is less than 100 t·hm-2,the BCRF,RR and BAR can improve the estimation saturation of the RF model by more than 11.0%(the RMSE of BCRF decreased by about 27.66t·hm-2 and rRMSE decreased by 10.99%compared to RF model).In general,the BCRF model is the most stable deviation correction model with improved performance,BCRF model yielded the highest accuracy and most stable performance for estimating AGB using multi-source remotely sensed data,which could effectively weaken the saturation phenomenon during biomass estimation.