基于BP神经网络的沙柳地上生物量预测模型

目的 以库布其沙漠沙柳为研究对象，建立基于BP神经网络的沙柳生物量模型，探究不同建模因子下的沙柳生物量估算模型变化，以期探究沙柳生物量估算模型的最优形式。 方法 选取6种沙柳生长因子，并根据与生物量相关性大小加入输入变量，从而组成6组不同输入变量，输入变量包含因子数量逐步增加（1 ～ 6种）。对比BP神经网络沙柳生物量模型不同输入变量所拟合模型的性能，确定最佳输入变量，并在最优输入变量的基础上，确定BP神经网络隐层数量，经过反复训练，建立基于BP神经网络的沙柳生物量估算模型。 结果 基于BP神经网络的沙柳生物量模型最优结构，即输入层节点数（Nin）∶隐层节点数（Nh）∶输出层节点数(Nout)为：4∶9∶1。其中训练数据R2=0.97，RMSE=0.67，MAE=0.50；测试数据R2=0.96，RMSE=1.10，MAE=0.77。 结论 基于BP神经网络的沙柳生物量，随着输入变量中输入因子的数量不断增加，发现其R2、RMSE、MAE所表现出的模型性能逐渐变好，但是输入变量每增加1种后，当输入因子数量为5时，模型精度相比输入因子数量为4时提升幅度较小，考虑模型使用时的精度和野外工作的便利性，输入层最优输入因子数为4种，当隐层数为9时模型性能表现为最优。

Aboveground Biomass Model of Salix psammophila Based on BP Neural Network

Objective To accurately predict the aboveground biomass of Salix psammophila in Kubuqi Desert, the aboveground biomass models including different predictors were developed using BP neural network. Methods Six independent variables were selected and added to the biomass model according to their correlation coefficients. The number of input variables gradually increased from 1 to 6. The best BP biomass model with the optimal input variables was selected based on the model performance. Results The best structure of S. psammophila BP biomass model was that the number of input layer nodes (Nin), hidden layer nodes (Nh) and output layer nodes (Nout) were 4∶9∶1, respectively. R2, RMSE, and MAE output from training data was 0.97, 0.68, 0.50, respectively; and 0.96, 1.16, 0.78, respectively from test data. Conclusion The performance of S. psammophila aboveground biomass model based on BP neural network became better with the increasing number of input variables. However, when the number of input variables was 5, the improvement of model performance was slightly better than the model with 4 input variables. Considering the model accuracy and model application, the optimal number of input variables in the input layer is 4, and the model performance is the best when the number of hidden layers is 9.