Estimation Of Tree Biomass In Hubei Province By Coupling Optical Image Data And Topographic Factors

Global forest ecosystems,as a huge carbon sink,play an important role in the global carbon balance.Therefore,in the background of global climate change,accurate estimation of forest biomass is crucial to measure the spatial distribution of forest biomass and to assess the carbon balance of terrestrial biota.In contrast,remote sensing technology has become the main way to estimate forest biomass because of the large monitoring area,rapidity,and dynamism compared with traditional biomass measurement methods.This paper constructs a forest biomass estimation model based on Sentinel-2 remote sensing images,DEM digital elevation model,and sample plot survey data,and takes forest vegetation in Hubei Province as the research object.Firstly,the sample plot survey data were processed to screen out tree forest sample plots,and the sample plot biomass was calculated based on the allometric growth equation of tree species using the per-tree survey data of the sample plots.Secondly,the obtained Sentinel-2 remote sensing images and elevation data were pre-processed to extract the required texture features,vegetation index,band reflectance,elevation,slope,and aspect.Finally,the factors with high correlation are screened out to form the most feature set through correlation analysis,and then the forest biomass models are constructed by using random forest and support vector institutions to compare the model prediction accuracy and select the optimal model,while the influence of red-edge band and topographic factors in optical remote sensing data on forest biomass estimation is explored,and biomass inversion models are constructed separately for different types of vegetation.The main research results are as follows:(1)There were large differences in the correlations between different variable characteristics and biomass.By comparing the correlations of different variables with biomass,it was found that the topographic factor had the highest importance,with the elevation variable having the highest correlation with biomass,P=0.23,which was significantly higher than the other variable characteristics.Among the spectral variables,each band was significantly correlated with biomass,with the highest correlation in the red-edge band(B7)with P=0.23,while among the vegetation indices,the enhanced vegetation index had the highest correlation.The correlations of texture variables were higher than those of vegetation indices,and the highest correlations with biomass were obtained with the mean and contrast texture features calculated in the NIR and red-edge bands.(2)By using two regression algorithms,random forest and support vector machine,to construct biomass models with different feature combinations,the model optimum of the random forest algorithm was obtained,and it was also found that adding the red-edge band and topographic factor would improve the model estimation accuracy.The biomass inversion model was constructed for tree forest vegetation based on texture features and spectral variables(including the red edge),and the root mean square error(RMSE)of the model was 33.06 Mg/hm~2,which was higher than the accuracy of the biomass model without red edge(RMSE=36.05 Mg/hm~2),combined with the correlation between red edge and biomass,indicating that the red edge band could improve the model accuracy of biomass estimation,but due to the influence of the extent of the study area and the complexity of the forest structure make the improved model accuracy limited.For the large-scale area,the root mean square error(RMSE)of biomass model accuracy was 31.32 Mg/hm~2 by adding topographic factors,which was higher than that of the biomass model without topographic factors,indicating that adding topographic factors could significantly improve the model accuracy.(3)Dividing the vegetation into three forest types can improve the estimation accuracy of the forest biomass model.To construct biomass models for different forest types,the model accuracy of the broad-leaved forest was the highest,followed by coniferous forest,and finally mixed coniferous forest,and their root mean square error(RMSE)were 29.49 Mg/hm~2,30.72 Mg/hm~2,and 32.19 Mg/hm~2,and the biomass model accuracy of broad-leaved forest and the coniferous forest was higher than that of the overall tree forest model,However,the biomass model accuracy of mixed forests was not improved due to more tree species and complex canopy structure.This indicates that classifying vegetation before biomass estimation can improve the biomass estimation accuracy of tree forests.