Study On Forest Biomass Inversion Using Machine Learning Method Based On SAR Image

Forest plays an important role in maintaining the balance of biosphere and realizing the sustainable development of resources et al.Forest biomass characterizes the life activities of forest,reflects the growth condition of vegetation,and especially embodies the forest’s ability to obtain energy and realize carbon sequestration.The estimation of forest biomass is a significant part of researching,monitoring and managing the ecological environment,and also it is one of the hot issues explored by many scholars.Because of its unique advantages,such as all-weather,all-time,strong penetration and so on,SAR technology provides a new means for the study of vegetation related issues,and is increasingly applied to related research of biomass inversion.Aiming at the problem that the nonlinear relationship between SAR image parameters and forest biomass is difficult to achieve an appropriately excellent fitting,the backscattering coefficient of SAR and texture features of SAR image are used to invert forest biomass by the regression modeling method of Machine Learning.Firstly,SAR basics are introduced;and the SAR backscattering coefficient and texture features of the experimental area are extracted;Secondly,the related problems of inverting forest biomass using Machine Learning methods and Ensemble Learning schemes are discussed.The main conclusions of this paper are as follows:(1)The Mean and Variance in SAR texure features have a positive significance to invert forest biomass,and their importance scores are relatively high.Random Forest(short for RF)and Recursive Feature Elimination algorithm(short for RFE)are combined to perform feature selection on a total of 12 predictors such as backscatter coefficients and texture features et al.The optimal combination of perdictors is determined as:Mean,DB,Variance and LIA,the importance scores for biomass regression are 100,90,53 and 39 in order.(2)The performance of Machine Learning methods is superior to multiple linear regression(short for MLR),and also Mean and Variance are beneficial and helpful to improve the accuracy of forest biomass inversion.The four variables(optimal combination)are used to establish the following invertion model of forest biomass regression in study area:RF(R~2=7581)>SVM(R~2=7339)>ANN(R~2=0.6807)>MLR(R~2=0.6790),in which RF is the best.In order to realise thorough analysis of texture features and Machine Learing methods,the futher work using two variables(DB and Variance)and one variable(DB only)to realize forest biomass inversion of above four models continues to be carried out.The overall performance of Machine Learing method prefer better than MLR.After the introduction of texture features(Mean and Variance),the accuracy of forest biomass inversion is over than the two-variable and univariate R~2 increased by 0.0604 and 0.0651,respectively.(3)The ensemble learning schemes which properly combined by Base Learners with greater differences and better performances have improved the accuracy of forest biomass inversion.The correlation calculation and the difference measurement of the combination are determined among common Machine Learning algorithm:RF,SVM,ANN,KNN and CART,which makes six ensemble learning schemes are determined.The integration strategy of“Stacking”is used to establish the models of forest biomass inversion based on optimal combination of predictive variables.The following schemes are presented:C8(R~2=0.7898)>C1(R~2=0.7791)>C12(R~2=0.7607)>C14(R~2=0.7528)>C2(R~2=0.7417)>C6(R~2=0.7067),in which the C8 scheme composed of RF,SVM and ANN is the best,and the accuracy R~2 is increased by 0.0317 compared to the previous optimal model RF.