Abstract: Stand volume is an important indicator to measure stand productivity of subcompartment. Using subcompartment data of Larix principis-rupprechtii Mayr plantation, this paper established two stand volume prediction models based on BP neural network and multiple regression, with stand age, site index and stand density as input variables and stand volume as the output variable. At the same time, the prediction results of the two models were compared. The results showed that: ① the optimal parameters combination of the BP neural network was the three layer network structure included the input layer of three neurons. the hidden layer of ten neurons and one neuron, the output layer of one neuron. The batch gradient descent method with momentum method was used in R language or Levevberg-Marquardt method in MATLAB software; ② in the multivariate regression model, the combination of “Logistic + power function”, V=SI0.977 2N0.510 3 0.500 1/\[1+44.226 1exp(-0.146 6t)\] in the modified function based on the growth theory equation was the best, and the coefficient of model fitting determination was R2=0.721 8; ③ in the prediction accuracy the BP model had optimal performance, followed by the multiple regression model, and finally volume table. Based on the above research, in order to improve the practicability of BP model, through JAVA and R language programming, BP neural network prediction model of subcompartment volume was constructed and stored in the stand volume yield prediction model of knowledge base, which could achieve the development of the classic mathematical model from the form of intelligent software, make the forestry staff use the intelligent system to easily fit and call the better fitting effect of the model to improve the BP model in the actual production of the operability, and provide decision support for forest management operations.

Key words: BP neural network, correction function, subcompartment volume prediction model, model library, Intelligent software of JAVA and R language programming

摘要： 林分蓄积是衡量小班林分生产力的重要指标。选择华北落叶松人工林小班数据，对以年龄、公顷株数和立地指数为自变量，小班公顷蓄积为因变量的BP (back propagation)神经网络模型和多元回归模型进行研究。研究结果表明：①BP神经网络参数最优组合：三层网络结构包括输入层3个神经元，隐含层10个神经元和1个神经元，输出层1个神经元，R语言算法选用含有动量的自适应梯度下降法，MATLAB软件算法选择Levevberg-Marquardt法；②多元回归模型中,生长理论方程为基础修正函数“Logistic+幂函数”组合的修正模型V=SI0.9772N0.51030.500 1/\[1+44.226 1exp(-0.146 6t)\]表现最优,其决定系数R2为0.721 8；③BP模型预测精度最高，其次是多元回归模型和材积表法。基于以上研究，为了提高BP模型的实用性，通过JAVA和R语言编程方式,将构建BP神经网络小班蓄积预估模型存储到收获预估模型的模型库中，在人工林收获预估中实现BP模型的调用，实现从经典的数学模型形式向智能化软件方向发展，提高BP模型在实际生产中的可操作性，为森林经营作业提供决策支持。

关键词: BP神经网络, 修正函数, 小班蓄积预估模型, 模型库, JAVA和R语言编程智能化软件