Optimized Method For Forest Aboveground Biomass Estimation Based On Remote Sensing Data And Its Spatiotemporal Analysis

Forests are an important part of terrestrial ecosystem and are the largest carbon reservoir in terrestrial ecosystem,in which it plays an important role in absorbing CO₂ and other greenhouse gases in the atmosphere,reducing the concentration of greenhouse gases,and mitigating global climate change.As the most basic quantitative characteristic of forest ecosystem,forest biomass is an important indictor to measure the capacity of carbon fixation,to judge the carbon source and sink of forest ecosystem,and is also an important parameter to characterize the function and ecological value and to evaluate the carbon balance of forest ecosystem.Therefore,the study of forest biomass estimation for the regional and global area in the context of climate change can not only provide a theoretical basis for the study of terrestrial ecosystem carbon cycle and global climate change,but also provide strategic guidelines for sustainable forest management and has great significance for appropriately utilizing the forest resources and improving the forest ecology environment.Accurate and rapid estimation of forest biomass in vast areas is attracting increasing attention in global change research.However,while remote sensing-based forest biomass estimation is extensively used and in general well developed,improving the accuracy of biomass estimation remains challenging.In this study,the Xiangjiang River Basin in Hunan Province was selected as the study area.China's National Forest Continuous Inventory data and Landsat images in 1999,2004,2009,and 2014 were used to establish linear regression(LR)model,random forest(RF)model,and extreme gradient boosting(XGBoost)model for the estimation of forest aboveground biomass(AGB)based on forest type,and analyze the spatiotemporal change of AGB and its driving factors.The study included the following tree steps:The first step is to analyze the impact of parameter optimization and variable selection of the models during the modeling process,then obtain the optimal models to estimate the AGB in the study area.The second step is to analyze the spatial autocorrelation of the residuals of the observed and predicted AGB,then correct the original predicted AGB results using the interpolation map of the residual AGB.The third step is to analyze the spatiotemporal change characteristics of AGB and the spatial differentiation of the driving factors.The results indicate the following:(1)XGBoost model has large number of parameters,and the performance of models can vary greatly depending on the selected values of these parameters,and XGBoost models are also more susceptible to overfitting if proper parameter values are not chosen.RF model has only two parameters that need to be optimized,thus the tuning process is relatively simple;it can realistically express the performance of RF model when the model chooses a larger number of decision trees.The stepwise regression approach was used to select the variable for LR,and the variable importance-based method was used for RF and XGBoost.Choosing appropriate variables can effectively improve model performance,and the variable selection was more important for XGBoost than for RF.The 3^rd,4^th,and 6^th bands and their texture variables play an important role in each model.It is different of assigning variable importance between RF and XGBoost,RF models mostly split the importance among the correlated multiple variables,whereas XGBoost models are inclined to centralize the importance at a single variable.In the LR,RF,XGBoost models of the broadleaved,coniferous,and mixed forest,the accuracy of the models of broadleaved forest was the highest,followed by coniferous and mixed forest.The results indicated that the approach of AGB modeling based on forest type is a very advantageous method for improving the performance.For the same forest type,the performance of XGBoost was better than that of RF,and RF was better than that of LR.(2)When using Kriging,it is necessary of conducting the semivariance analysis before the prediction of interpolation.The results of semivariance analysis indicated that the spatial distribution of the residual AGB had strong spatial autocorrelation,and the results of the Global Moran's Index also verified that the residual AGB were significantly spatially autocorrelated;therefore,Kriging interpolation can be used for spatial interpolation.The AGB map corrected by the spatial interpolation map of the residual AGB can achieve relatively low and high AGB values and,thus moderate the under-and overestimation,and significantly improve the accuracy of AGB estimation.The results of matched samples t-test showed that there were significant differences of the predicted AGB between the uncorrected and corrected AGB maps of three forest types.(3)In 1999,2004,2009,and 2014,the AGB was increased with year,and the average AGB was 54.623 Mg/ha,66.242 Mg/ha,68.079 Mg/ha,and 77.579 Mg/ha,respectively.The results of the transfer matrix,landscape metrics,centroid distributions,and hot spot analysis of the AGB in different years showed that the total AGB was growing and the forest quality was improving.The driving factors were selected by using the global ordinary least squares linear regression,and the factors of elevation(ELEV.),normalized difference vegetation index(NDVI),annual mean temperature(TEMP.),annual precipitation(PREC.),populations(POP.),and gross domestic product(GDP)were finally selected;then,these factors were used to construct the geographical weighted regression(GWR)models.The results showed that NDVI was the most important positive factor causing the change of AGB,and TEMP.was another main factor which has an important influence on the growth of vegetation,according to the regression coefficients of each driving factor in the GWR models.In terms of the role of driving factors,ELEV.and NDVI had a positive effect on the change of AGB and had a gradually increasing trend,PREC.had a positive effect but had a gradually weakening trend,TEMP.and POP.had a negative effect and had a gradually increasing trend,the effect of GDP was turned negative to positive and had a gradually increasing trend.In addition,the driving factors worked powerfully in the mountainous area with higher elevation in the southeast and south of study area,while the driving factors were relatively weak in the plain areas with lower elevation in the middle and north of study area.