论文部分内容阅读
【目的】研究BP神经网络模型在树高预测中的应用,比较分析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(tansig)和线性函数(purelin)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。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,决定系数R2为0.5210,均方根误差为RMSE为2.0917m,平均绝对误差MAE为1.6276m。加入立地因子后,决定系数R2提高至0.5736,提高了10.10%,均方根误差RMSE为1.9736m,降低了5.65%,平均绝对误差MAE为1.5797m,降低了2.94%;在此基础上,加入林木竞争因子后,决定系数R2方为0.8455,增长了47.40%,均方根误差RMSE为1.1879m,下降了39.81%,平均绝对误差MAE为0.9685m,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的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 purpose of this paper is to study the application of BP neural network model in tree height prediction, compare and analyze the influence of different forest investigation factors and different neural network training algorithms on average tree height prediction, and provide a new method for tree height prediction. 【Method】 Based on 314 observation data of 168 permanent sites of Larix olgensis plantations in Jilin Province, an average height growth model was established using BP neural network modeling technique. The input factors first join the age, and then join the site factors and forest competition factors, analysis of site factors and forest competition factors on the tree height. Based on the Sigmoid function (tansig) and purelin in MATLAB R2016b, the network is trained by the Bayesian regularization algorithm and the Levenberg-Marquatdt algorithm (L-M algorithm for short) respectively. The differences between Bayesian regularization algorithm and L-M algorithm as training function are analyzed. 【Result】 Compared with L-M training algorithm, Bayesian regularization training algorithm has better generalization ability. After adding the age, site factor and tree competition factor, the fitting precision of tree height showed the same upward trend. Using Bayesian regularization training algorithm, when the age is taken as the input factor, the coefficient of determination R2 is 0.5210, the root mean square error is RMSE is 2.0917m and the average absolute error MAE is 1.6276m. After adding the site factor, the coefficient of determination R2 was increased to 0.5736, an increase of 10.10%. The root mean square error RMSE was 1.9736m, a decrease of 5.65%. The average absolute error MAE was 1.5797m, a decrease of 2.94%. On this basis, After the forest competition factor, the coefficient of determination R2 was 0.8455, an increase of 47.40%. The root mean square error RMSE was 1.1879m, a decrease of 39.81%. The average absolute error MAE was 0.9685m, a decrease of 38.69%. 【Conclusion】 Bayesian regularization BP neural network can accurately predict the average height of Larix olgensis plantations. Site factors and tree competition factors can better improve the accuracy of forest growth prediction, and the impact of tree competition factors on the tree height is significantly greater than the site factor.