Inversion Study Of Forest Aboveground Biomass By Coupling Spaceborne LiDAR And Optical Remote Sensing Data

Aboveground biomass(AGB)can directly measure the growth status and carbon sequestration capacity of forests,and quickly and accurately obtaining AGB is a fundamental task of forest resource management and ecosystem dynamic monitoring.The acquisition of regional forest AGB by remote sensing technology can significantly improve survey efficiency,and reduce costs and damage to forests compared with traditional manual surveys.In addition,Ice,Cloud,and Land Elevation Satellite-2(ICESat-2),as one of the representatives of spaceborne Light Laser Detection and Ranging(LiDAR),can penetrate the forest canopy gaps and obtain forest vertical canopy parameters on the earth’s surface and thus has the potential to realize the inversion of forest AGB in large areas.In the study,the plantation forests in the Saihanba forest farm and the natural forests in the Natural Forest Resources Protection Project(NFRPP)area of Tibet were taken as objects.Based on Sentinel-2 optical remote sensing images and ICESat-2 spaceborne LiDAR data,remote sensing variables were extracted.Combined with ground survey forest AGB data,linear stepwise regression(LSR),Boruta,Variable Selection Using Random Forests(VSURF),and Stepwise random forest(SRF)proposed by the study were used to screen feature variables.Multiple Linear Regression(MLR),k Nearest Neighbor(kNN),Support Vector Machine(SVM),random forest(RF),and Stacking model were constructed based on the screening results to retrieve forest AGB,and the optimal inversion results were coupled with optical remote sensing variables to achieve continuous spatial distribution mapping of forest AGB.In addition,the study proposed an optimized regression kriging model based on empirical Bayesian to correct the original inversion results to improve the accuracy of forest AGB inversion.The main research conclusions and innovative points are as follows:(1)Combining Sentinel-2 and ICESat-2 data for modeling significantly improved the accuracy of forest AGB inversion,revealing that the combined data source can combine the advantages of optical and LiDAR variables and has the potential to improve the inversion of forest AGB.The root mean square errors(RMSEs)of the inverse model constructed by the combined variable group were reduced by 17.5%～22.4%in the Saihanba forest farm and 34.0%-45.8%in the NFRPP area of Tibet compared with the Sentinel-2 optical data only.In addition,the light saturation points of the combined variable group in the Saihanba forest farm and the NFRPP area of Tibet reached 269.3 Mg/hm2 and 261.3 Mg/hm2,respectively,which were improved by 16.9%and 15.5%,respectively,compared with the Sentinel-2 variable group.The LiDAR variable can reflect the vertical structure information of the forest,and combining optical and LiDAR variables for modeling can effectively mitigate the light saturation effect,thus improving the accuracy of forest AGB inversion.(2)A novel SRF method for efficient screening of remote sensing feature variables was proposed,which significantly improved the efficiency of feature variable screening and effectively reduced the model estimation error compared with other linear and nonlinear methods.The study optimized the nonlinear feature variable screening method based on importance evaluation,and innovatively proposed the SRF method to screen the feature variables for forest AGB modeling,and the linear method such as LSR and nonlinear methods such as Boruta and VSURF were conducted for comparion.The SRF method achieved the lowest estimation error in all three variable groups set up in the two study areas,with SRF method reducing 5.3%,9.0%and 14.3%compared to LSR method in the Saihanba forest farm,and 6.7%,13.3%and 16.6%in the NFRPP area of Tibet,respectively.The running time of the SRF method is moderate,which can effectively improve the efficiency of variable screening.In addition,none of the modeling variables obtained by the SRF method screening had a significant relationship with the residuals of the optimal model(P>0.05),indicating that the screening method is reasonable and feasible.(3)An integrated learning model of Stacking with heterogeneous integration optimization was constructed,which integrated the advantages of parametric and nonparametric models and effectively improved the accuracy of forest AGB inversion.Among all the inversion models constructed in the Saihanba forest farm and the NFRPP area of Tibet,the Stacking model constructed by heterogeneous integration optimization achieved the highest coefficient of determination and the lowest estimation error,with coefficients of 0.74(RMSE=43.08 Mg/hm2)and 0.88(RMSE=15.09 Mg/hm2),respectively.The overestimation and underestimation of forest AGB are significantly improved compared with other models,and can effectively achieve high-precision inversion of forest AGB.(4)The coupling of Sentinel-2 optical images has broken through the limitations of discrete distribution of spaceborne LiDAR data and achieved spatially continuous mapping of forest aboveground biomass distribution.The optimal inversion results constructed from the combined data sources are used as the derived values to establish a coupling relationship with Sentinel-2 optical remote sensing variables to obtain highly accurate and spatially continuous forest AGB distributions.The higher predicted AGB values in the Saihanba forest farm were mainly distributed in the northeastern,southwestern,and central parts,with smaller forest AGB values in the eastern and southern parts.The western and southern areas contain parts of cultivated and built-up areas,resulting in smaller values of AGB in the surrounding distribution of forests.The higher predicted forest AGB values in the NFRPP area of Tibet were mainly distributed in the northern,eastern,and southwestern regions,while the predicted values in the northwestern region were relatively small.In addition,the forest AGB values distributed in some central and western regions were smaller due to the influence of human activities.The inversion results of forest AGB in the Saihanba forest farm and NFRPP area of Tibet were in good agreement with the actual forest distribution patterns.(5)An empirical Bayesian-based optimal regression kriging forest AGB estimation correction model was constructed,which effectively improved the original inversion results and enhanced the local estimation and spatial distribution mapping of forest AGB.The coefficients of determination of the corrected forest AGB inversion models reached 0.93(RMSE=21.57 Mg/hm2)and 0.95(RMSE=9.83 Mg/hm2)for the Saihanba forest farm and NFRPP area of Tibet,respectively.The spatial distribution of forest AGB in the study area was expanded by 26.4%and 35.6%,respectively,and the spatial distribution of forest AGB was more reasonable,which can analyze the local distribution characteristics and differences of AGB more effectively and provide more scientific and reasonable references for forest resource management and survey design.