The increasingly mature nonlinear technique can facilitate accurate forecasting of transient sap flow process of plant.In this paper,the dominated tree species,Pinus tabulaeformis and Platycladus orientalis in Beijing...The increasingly mature nonlinear technique can facilitate accurate forecasting of transient sap flow process of plant.In this paper,the dominated tree species,Pinus tabulaeformis and Platycladus orientalis in Beijing mountainous area were chosen for study.Their monitoring data range from June 18 th to September 9 th 2007 was derived to form the 1 985 sets of sample respectively.BP (back propagation) neural network models were established according to the theory of automaton network of discrete dynamic system,the target output of which was sap flow velocity and the inputs of which consisted of five influencing factors,ie,air temperature,relative humidity,light intensity,stem diameter growth and soil water potential.To improve the generalization quality of networks,Bayesian regularization and early stopping modes were involved in the training process.After training in two modes above,the linear regression between simulated outputs and the corresponding targets of test sample sets showed good fits (R>0.85),which indicated a high forecasting precision of the models established,specifically when 11 neurons in hidden layer.Models demonstrated fine generalization under the two training modes in that the fit of test sample was equivalent to that of training sample,which further indicated their availability in practice.展开更多
文摘The increasingly mature nonlinear technique can facilitate accurate forecasting of transient sap flow process of plant.In this paper,the dominated tree species,Pinus tabulaeformis and Platycladus orientalis in Beijing mountainous area were chosen for study.Their monitoring data range from June 18 th to September 9 th 2007 was derived to form the 1 985 sets of sample respectively.BP (back propagation) neural network models were established according to the theory of automaton network of discrete dynamic system,the target output of which was sap flow velocity and the inputs of which consisted of five influencing factors,ie,air temperature,relative humidity,light intensity,stem diameter growth and soil water potential.To improve the generalization quality of networks,Bayesian regularization and early stopping modes were involved in the training process.After training in two modes above,the linear regression between simulated outputs and the corresponding targets of test sample sets showed good fits (R>0.85),which indicated a high forecasting precision of the models established,specifically when 11 neurons in hidden layer.Models demonstrated fine generalization under the two training modes in that the fit of test sample was equivalent to that of training sample,which further indicated their availability in practice.
