杉木人工林胸径生长神经网络建模研究

Modelling tree diameter growth for Chinese Fir plantations with Neural Networks

【目的】探索神经网络技术对杉木人工林胸径生长的模拟和预测能力,以寻求最优模型。【方法】以江西大岗山杉木人工林为研究对象,依据林木生长理论,用林龄(A)、立地指数(SI)和初植密度(N)3个因子构建平均胸径生长BP模型;通过定量和定性分析相结合的方法对模型选优,并将最佳模型与拓展的Richards模型比较;最后将优化模型应用于未参与建模的样地。【结果】最佳BP模型为Levenberg-Marquardt算法3∶5∶1结构模型(LM351),R2=0.984,MSE=0.196;拓展的Richards模型R2=0.964,MSE=0.433。LM351模型经校正后,适合预测福建邵武杉木人工林胸径生长规律(R2=0.995)。【结论】LM351神经网络模型在精度上优于传统Richards模型,适于林龄6～28年、立地指数12～17m、初植密度1 667～9 967株/hm2的杉木林分平均胸径的模拟和预测。

[Objective] The paper developed a growth model of mean diameter（DBH, 1. 3 m above ground） for Chinese Fir plantation with Back-Propagation network（BP）. [Method] Data from even-aged Chinese Fir（Cunninghamia lanceolata （Lamb.） Hook） plantation including conventional measurements of 15 permanent sample plots located at Dagangshan Mountain in Jiangxi province from 1981 to 2008 were collected. BP models employed were one hidden layer with tan-sigmoid function and linearity function as transfer function of hidden and output layer respectively. According to the theory of stand growth, three factors,age （A）,site index （SI） and initial planting density （N）,were used as predictors for DBH of Chi- nese Fir. Then BP models were trained by 14 algorithms combining with different structures. Subsequent- ly, the best was selected by comparison with performances of qualitative and quantitative index. On the oth- er hand,generalized Richards＇s model was introduced as a reference model. Ultimately,the optimal model was examined by independent data from Fujian province. [Result] The optimal BP with one hidden layer containing 3 hidden units was trained by Levenberg-Marquardt algorithm （R2= 0. 984,MSE= 0. 196）, which was better than Richards model （R2 =0. 964,MSE=0. 433）. Practical Result demonstrated calibra- ted model is also reasonable for Chinese fir plantation in different sites. [Conclusion] BP model is suitable for simulation and prognoses of planted forests of Chinese Fir in which stand variables are within A=6-28years , SI=12-17 m,N-1 667-9 967 stems/hm2 ,and ANN generalize well to new data.