| Standing volume is an important index of forest resources investigation as well as the main basis for calculating forest volume and biomass.For enhancing the effect of preserving the forest resources,prediction of standing volume accurate is particularly important in improving the social values,economic values and ecological values of the forest,grasping the dynamic changes of the quantity and quality of forest resources better and formulating more scientific and reasonable forest management policies namely.This paper taking Pinus sylvestris var.mongolica litv.of Tuqiang Forestry Bureau in Greater Khingan Mountains as the research object and selects the traditional nonlinear model(NLR)and three machine learning algorithms: reverse neural network(BP),ε-Support vector machine regression(ε-SVR),random forest(RF).By using R language and MATLAB2019 b,this paper aims to build the optimal volume model with skin and peeling.Taking Akaike information criterion(AIC)and Bayesian information criterion(BIC)as the evaluation indexes of heteroscedasticity correction of traditional models,the ability of model fitting and prediction is evaluated by determining coefficient(R2),root mean square error(RMSE),relative root mean square error(RMSE%)and mean absolute error(MAE).The mean relative error(MRB)is applied to test the prediction deviation of the model.It can be deduced that the shape ratio of 50%relative tree height have the highest accuracy in the ternary volume equation,and the heteroscedasticity phenomenon in NLR model can be effectively solved by adding power function and constant plus power function to the regression equation;And in the BP model,the problem of initial random weight and threshold are eliminated by 30 times in average.10-fold cross validation is used to select the optimal learning algorithm and number of neurons in the hidden layer of the BP model;In ε-SVR model,in order to reduce training time,the genetic algorithm is selected to globally study the optimal kernel function,as well as the penalty factor C and kernel parameter gamma under the kernel function;In the RF model,the values of the three main RF parameters ntree,minleaf and mtry are screened by the oob error test.Substitute the parameters with the highest fitting accuracy into each model for modeling,obtain the optimal model under different variables,and predict the test sample data.The results show that the prediction accuracy of the three machine learning algorithm models can be stimulated as a higher model than the traditional one,in which the RF model tends to have the highest accuracy under input and output variables,BP and ε-SVR model are less accurate than previous ones and the result made by NLR model tends to have the lowest accuracy.In the binary volume model based on DBH D and tree height H,compared with NLR,the R2 prediction accuracy of RF in skinned and peeled volume is increased by 4.31% and 4.45% while RMSE is decreased by 48.44% and 44.69%,which can be discovered as a significant difference.The overall results of BP and ε-SVR models are mostly similar,but there are slight differences due to the different input and output variables.In the volume model with DBH D,BP is slightly better than ε-SVR;In the volume model it were introduced with DBH D and tree height H and the BP accuracy is slightly higher when predicting the volume with skin,the accuracy of ε-SVR is slightly higher when predicting the volume without skin;In the volume model incorporating DBH D,tree height H and shape ratio q,the ε-SVR is slightly better than the BP model.In the process of comparing different parameters under the same model,it is found that adding parameters can effectively improve the accuracy of the same model.The accuracy of model which only one variable of diameter at breast height D is introduced into RF can be found even higher than that made by NLR model,while two variables of diameter at breast height D and tree height H are introduced.These above can be concluded that predicting standing tree volume by applying RF model is more flexible than traditional models and tend to maintain higher accuracy in predicting height of tree even under the tough condition.It can be seen that the application of machine learning algorithms can provide new solutions for the accurate investigation and management of forest resources. |
