基于BP神经网络的长白落叶松人工林林分平均高预测

Prediction mean height for Larix olgensis plantation based on Bayesian-regularization BP neural network

【目的】研究BP神经网络模型在树高预测中的应用,分析比较不同森林调查因子及不同神经网络训练算法对平均树高预测的影响,为树高预测提供新的方法。【方法】以吉林省长白落叶松人工林为对象,基于168块固定样地的314个观测数据,运用BP神经网络建模技术建立了林分平均树高生长模型。输入因子首先加入年龄,然后依次加入立地因子及林木竞争因子,分析立地因子及林木竞争因子对树高的影响。基于Matlab R2016b中的Sigmoid函数和线性函数为神经元的传递函数,分别采用贝叶斯正则化算法和Levenberg-Marquatdt算法(简称L-M算法)对网络进行训练,对比分析了贝叶斯正则化算法和L-M算法作为训练函数的差异。【结果】与L-M训练算法相比,贝叶斯正则化训练算法具有更好的泛化能力。模型中依次加入年龄、立地因子、林木竞争因子后,树高的拟合精度呈现出相同的上升趋势。采用贝叶斯正则化训练算法,当年龄作为输入因子时,决定系数R^2为0.521 0,均方根误差(RMSE)为2.091 7,平均绝对误差(MAE)为1.627 6。加入立地因子后,决定系数R^2提高至0.573 6,提高了10.10%,均方根误差(RMSE)为1.973 6,降低了5.65%,平均绝对误差(MAE)为1.5797,降低了2.94%;在此基础上,加入林木竞争因子后,决定系数R^2为0.845 5,增长了47.40%,均方根误差(RMSE)为1.187 9,下降了39.81%,平均绝对误差(MAE)为0.968 5,下降了38.69%。【结论】利用贝叶斯正则化BP神经网络可以准确地预测长白落叶松人工林的平均高。立地因子及林木竞争因子能够较好地提升林木生长预测的精度,且林木竞争因子对树高的影响明显大于立地因子。

【Objective】The neural network model was used for stand height prediction. Different types of forest survey factors and neural network training algorithm were examined. The results were expected to provide new method for stand height prediction.【Method】Stand mean height prediction model was developed based on the 314 observations from 168 permanent sample plots of Larix olgensis plantation in Jilin Province with BP neural network modeling technology. Firstly we input the age,then input the site factor and competitive factor to analyze their contributions to stand mean height. We applied Bayesian-regularization algorithm and Levenberg-Marquatdt algorithm for network training based on tansig function and purelin function in Matlab R2016b as neural transferring function. We compared and analyzed the differences of the two algorithms as training functions. 【Result】The analysis result showed that no matter which training algorithm was adopted,the estimation precision of stand mean height would be improved after the age,site factor and forest competitive factor were added in succession and Bayesian-regularization algorithm had better generalization ability than L-M training algorithm.The coefficient of determination R-2 was 0.521 0,RMSE was 2.091 7,MAE was 1.627 6 when Bayesian-reg-ularization algorithm was adopted and age as input factor;when site factor was added subsequently,R-2,RMSE andMAE was 0.573 6（increased by 10.10%）,1.973 6（decreased by 5.65%）,and 1.579 7（decreased by 2.94%）,respectively;when forest competitive factor was added,R-2,RMSE and MAE was 0.845 5（increased by 47.40%）,1. 187 9（ decreased by 39. 81% ）,and 0. 968 5（ decreased by 38. 69% ）,respectively.【Conclusion】Bayesian regularization neural network modeling could be accurate for stand mean height prediction of Larix olgensis plantation. The precision was improved after site and forest competitive factor were taken into consideration and forest competitive factor gave more contribution than site factor.