Effect Of Form Quotient On Individual Tree Volume Of Betula Platyphylla And Model Construction

Individual tree volume plays an important role in forest resource survey and is the primary basis for estimating stock volume and biomass.Therefore,in order to better protect forest natural resources,increase its economic and ecological benefits,and more accurately monitor the dynamic changes in quantity and quality of forest resources,it is necessary to accurately predict the volume of individual tree.Taking Betula platyphylla in the Daxing’an Mountains as the research object,constructs a volume model with form quotient,and compares it with the two-variable white birch volume model（WBVM）and traditional basic volume model of Betula platyphylla in Northeast China.At the same time,the generalized additive model（GAM）and four machine learning algorithm models（random forest,RF;artificial neural network,ANN;support vector regression,SVR;k-nearest neighbor,k-NN.）were constructed to study the influence of form quotient on different types of volume models.The 15 form quotient on the trunk were introduced into the traditional basic volume model to construct the two-variable and three-variable single volume model respectively.The GNLS module of R software was used to fit each model,and the variance function was introduced to eliminate the heteroscedasticity in the fitting process of each volume model.At this time,there were four combinations of independent variables,namely,D,D and H,D and q_0.4,D and H and q_0.4.Based on these four independent variables,GAM model and four machine learning models are constructed respectively.There are 24 volume models of 4*6.Finally,mean percentage of bias,mean absolute bias,root mean square error and determination coefficient were used as evaluation indexes to compare and analyze the fitting and prediction abilities of each model.In the process of fitting the conventional volume model,it has been observed that the inclusion of the form quotient can remarkably enhance the fitting performance of the volume model.When the relative tree height was 40%,the fitting effect of the two-variable and three-variable volume model with form quotient q_0.4 were the best.In the GAM model,six smooth spline functions were found to be able to predict the individual wood volume well.In the RF model,the variance and MSE are calculated to select the optimal mtry,and the optimal ntree is determined by drawing the variation curve of MSE against the number of trees.In the ANN model,the number of neurons is continuously increased by trial and error method,and the changes of MSE with the number of neurons are obtained,and the optimal number of neurons in the hidden layer is obtained.Then,by comparing the fitting results,it is found that the performance of all models is the best when the activation function is logistic,and when the input variable changes,the optimal algorithm changes accordingly.In the SVR model,the 10-fold cross-validation method is used,and it is found that the radial basis kernel function has the best performance for different input variables.When the input variable is only D,the optimal gamma=0.1 and cost=5 are selected,and the optimal gamma=0.1 and cost=10 are selected for other cases.In the k-NN model,the grid method are used to find the optimal k value.The fitting results show that when the independent variable is only D,the optimal k value is 7,and in other cases,the optimal k value is 3.Finally,the optimal parameters are substituted into each model to build the model,and the test data is used for validation.The results showed that only when the input variable was D,the accuracy of GAM model was the highest,followed by ANN and SVR model.When using the remaining three input variable combinations(D and H,or D and q_0.4,or D and H and q_0.4),the test accuracy of ANN model was the highest.The specific validation accuracy is sorted as follows:When the input variable is D,GAM>ANN>SVR>k-NN>RF>GNLS;When the input variables are D and H,ANN>SVR>GAM>RF>GNLS>k-NN>WBVM;When the input variables are D and q_0.4,ANN>SVR>GNLS>RF>GAM>k-NN;When the input variables are D and H and q_0.4,ANN>SVR>GNLS>GAM>RF>k-NN.Therefore,it is recommended to use GAM or ANN models to estimate individual tree volume of Betula platyphylla.