Biomass Estimation Methods Of Poplar Plantation Based On Multi Source Data

This paper,which is based on the technology of 3S and combined with the 118 plots of the 2016 survey data(including 88 block used for regression modeling,30 for accuracy test model)and takes Dongtai city of poplar in Jiangsu Province as the research object,using Landsa8 and GF-2 image data,uses traditional regression,BP neural network regression,support vector regression(Support Vector Regression,SVR),random forest regression(Random Forest,RF)and other methods to construct artificial forest biomass estimation models of poplar.The results of this study include the following aspects:(1)According to the field investigation in plots about the height of the tree and width of the crown,we can use the data which includes arithmetic mean height,crown area weighted height and basal area weighted high and measured biomass to establish a simple regression analysis.The results show that the correlation between weighted mean height and measured biomass was the highest(R~2=0.917).(2)Import the coordinate points of poplar,which is obtained through field investigation,into Arcmap to extract the remote sensing information of the corresponding plot,then screen factor for modeling by the correlation between the independent variables and the measured sample biomass.The results show that there are 21 factors significantly related to the biomass of the 26 variables,which are based on Landsat8 image data extraction andthe highest correlation factor is vegetation index DVI,whose correlation coefficient is 0.763.At the same time,the near infrared band is the band whose correlation factor is the highest between single factors,whose correlation coefficient is 0.689;based on 22 independent variables GF-2 image data extraction,there are 18 factors significantly related to the field measured biomass,the highest correlation factor is vegetation index NDVI,whose correlation coefficient is 0.776.Similarly,the near infrared band is the highest correlation factor in the single band factor,whose correlation coefficient is 0.651.(3)Combining the remote sensing factors which is extracted by Landsat8 and GF-2 image data with the kind of high thoracic section weighted average,we can construct 8 estimation models of traditional poplar biomass,which are L model based on multiple linear regression(R~2=0.708)model,L-H(R~2=0.838),G model(R~2= 0.711),G-H model(R~2=0.852)and stepwise regression model based on L(R~2=0.764),L-H model(R~2=0.841),G model(R~2=0.796)and G-H model(R~2=0.891).The results show that the multi factor GF-2 remote sensing image data extraction,based on sample weighted average breast height section high building regression model,is the optimal estimation model of traditional regression model of Poplar Plantation biomass.(4)Use BP neural network to estimate the biomass of Poplar Plantation in the study area,and compare with the estimation after the principal component analysis(PCA).The results show that the the model accuracy(R~2=0.900)which uses the independent variables of GF-2 image data as the input variables of BP neural networkis better and the model effect(R~2=0.882)is better as well.Similarly,the prediction effect(R~2=0.839)of the BP neural network model based on the independent variables of Landsat8 image data is better than that of the PCA-BP neural network model(R~2=0.808).(5)The highest accuracy of the biomass estimation model of poplar plantation which is based on SVR,and the best prediction effect can be.For the Landsat8 image data,the training accuracy and prediction accuracy of the SVR model are 0.903 and 0.896,respectively.For the GF-2 image data,the training accuracy and prediction accuracy of the SVR model are respectively 0.927,0.915.It can be used as a preferred method in estimating biomass of poplar plantation.(6)The OOB-RF model based on the importance of OOB is the most accurate estimation model of poplar biomass in RF model.In the three data analysis of random forest model to estimate the biomass of poplar plantation method using R,OOB,correlation coefficient GRA,the estimation accuracy of OOB-RF model is highest,with Landsat8 and GF-2 images as the data source of the estimated R~2 were 0.899,0.922.Comparison and analysis of two kinds of estimation models are based on Landsat8 and GF-2.The results show that the prediction accuracy of GF-2 image data is higher than that of Landsat8,and the prediction accuracy of SVR model is the highest for the two kinds of data.