Modeling Of Forest Aboveground Biomass Based On Optical And Interferometric Synthetic Aperture Radar

Forest aboveground biomass(AGB),an essential functional parameter of terrestrial ecosystems,plays a vital role in the carbon cycle.By taking charge of the carbon exchange between the land and atmosphere,AGB has the important effect on climate changes.The spatial explicit forest AGB data support the global carbon budget and management.Additionally,as an input of models on the productivity and climate changes,the accuracy of forest AGB definitely determines the reliability of predictions.Thus,monitoring spatio-temporal dynamics of forest AGB is essential to the research on the terrestrial ecosystem function,carbon budget and global climate changes.Satellite remote sensing techniques have rapid developments and wide applications on the retrieval of forest horizontal and vertical parameters due to advantages on the long-time and large-area monitoring in a comparable manner with passive and active sensors.It also results in the fast growth of remote sensing models on forest AGB.Previous models on forest AGB are based on the relationship between field-measured biomass and predictor variables from remote sensing data,which has the mismatching problem of the scale.Parametric models are urgently needed in ecological and environmental applications at large scales,while it is crucial to improve the accuracy and ability to model complex relationships.In order to solve premise problems of parametric models based on field points,this study aims to build an accurate allometric growth equation at the pixel scale based on horizontal and vertical parameters derived from Sentinel-1 synthetic aperture radar(SAR)and Sentinel-2 multi-spectral instrument(MSI)data.This study combined recently launched and applied Advanced Land Observing Satellite-2(ALOS-2)L band and Sentinel-1 C band SAR as well as Sentinel-2 MSI,and explored useful and valid predictors for forest AGB modeling.Then hybrid models were firstly put forward to simulate complex relationships between predictors and AGB,and acquired pixel values of forest biomass at the 10 m pixel size.Finally,the more accurate allometric growthequation was built based on pixel values of AGB and Sentinel-based predictors,i.e.,height and biophysical variables,to map forest biomass at the regional scale.Main objectives and results are follows:(1)By the correlation analysis between field-measured forest AGB and multi-sensor variables from Sentinel-1,Sentinel-1 and ALOS-2,the sensibility to biomass was determined and discussed.After deleting the redundancy of variables,predictors for forest AGB estimation were acquired.Results showed that texture features of backscatters from L band SAR were most sensitive to forest AGB,followed by vegetation indices and biophysical variables from the optical sensor,while backscatters from C band SAR and micro-topographic indictors from C band interferometric SAR(In SAR)were least.As for improving the sensibility of SAR variables to forest AGB,the influence of the wavelength and interferometric processing was greater than that of the texture analysis.Horizontal-vertical(HV)and vertical-vertical(VV)channels were more related to forest AGB than horizontal-horizontal(HH)and vertical-horizontal(VH)channels.About optical variables,red,red-edge and short wave infrared bands were more sensitive to forest AGB.Vegetation indices calculated by red-edge bands had the stronger correlation to AGB.Totally,34 predictors were selected,including gamma HV and HH backscatters and their texture features from L band SAR,gamma VV backscatter and texture features of VV and VH backscatter from C band SAR,and the reflectance,vegetation indices and biophysical variables from MSI,as well as topographic indicators from C band In SAR.(2)The optimal model for predicting pixel values of forest AGB was determined by comparing capabilities of modeling local or global,linear or non-linear relationships among 10 algorithms belonging to parametric,non-parametric and hybrid approaches.According to coefficients from parametric models,connection weights from artificial neural network(ANN),increasing root-mean-square error(RMSE)by excluding predictors from support vector machine for regression(SVR),and the attribute importance from random forest(RF),and contributions of predictor variables wereranked and discussed.Results demonstrated that the accuracy ranking in a decreasing order was RF kriging(RFK),RF,ANN kriging(ANNK),SVR kriging(SVRK),geographically weighted regression kriging(GWRK),stepwise regression kriging(SWRK),ANN,SVR,geographically weighted regression(GWR),and stepwise regression(SWR).Two-step hybrid approaches acquired more accurate predictions than the corresponding individual models,and non-linear models showed better performances than linear models.RFK was the optimal model and achieved RMSE of29.72 Mg/ha.Predictors from C band SAR contributed least on AGB prediction,yet,that from L band SAR influenced most.Texture features from L band HV backscatters were most important,especially,the entropy(HV＿ENT).It was indicated that the randomness of HV backscatters had the greatest effect on forest AGB prediction.Reflectance was more important in parametric models,while vegetation indices were more vital in non-parametric algorithms.(3)As a pioneering study,this thesis retrieved two forest vertical parameters based on Sentinel data,i.e.,digital surface model(DSM)by In SAR,and tree height by backscatters,fractional vegetation cover(FVC)and the simplified water-cloud model.Results revealed that most accurate DSM of 2017 achieved RMSE of 97.11 m(14.28%)and correlation coefficient(r)of 0.86.The attenuation coefficient of VV ranges from0.02 to 0.05,however,that of VH was 0.03 to 0.07,which indicated that VV backscatters had larger saturation values of forest vertical parameters.Final mosaic tree height of the study area in July,2017 had the RMSE of 2.72 m(17.98%)and r of 0.75.(4)An accurate allometric growth equation was built based on pixel values of AGB and Sentinel-based horizontal and vertical parameters to map AGB in a regional scale,which was compared to the traditional point-based parametric model.Results illustrated that,the pixel-based allometric growth equation was more accurate than the point-based.The pixel-based allometric growth equation built by DSM and FVC was the optimal model for AGB mapping with RMSE of 39.45 Mg/ha and r of 0.90.The pixel values of forest AGB predicted by RFK had the great accuracy,and provided training samples for biomass modeling.Sentinel-based horizontal and vertical parameters explained spatialvariations of forest AGB at a large extent.The pixel-based allometric growth equation,which took full advantages of pixel values of remote sensing data and was built in the same scale,lessened the saturation problem of Sentinel data.Based on the optimal equation,forest AGB of 2017 in Dunhua city of Jilin Province was mapped with values from 0 to 898.9 Mg/ha.The forests with large values of AGB were located in the northern,eastern and western marginal areas with the higher altitude,and that with small biomass values were in the north of the downtown with a lower elevation.