人工神经网络模型及其在农业和生态学研究中的应用

对于一些复杂的农业生态系统,人们对其生态过程了解较少,且这些系统的不确定性和模糊性较大,用传统的方法难以模拟这些系统的行为,神经网络模型因为能较精确地模拟这些系统的行为,而引起生态学者们的广泛兴趣。该文着重介绍了误差逆传神经网络模型的结构、算法及其在农业和生态学中的应用研究。误差逆传神经网络模型一般采用三层神经网络模型结构,三层的神经网络模型能模拟任意复杂程度的连续函数,而且因为它的结构小而不容易产生与训练数据的过度吻合。误差逆传神经网络模型算法的主要特征是:利用当前的输入误差对权值进行调整。在生态学和农业研究中,误差逆传神经网络模型通常作为非线性函数模拟器用于预测作物产量、生物生产量、生物与环境之间的关系等。已有的研究表明:误差逆传神经网络模型的模拟精度要远远高于多元线性方程,类似于非线性方程,而在样本量足够的情况下,有一定的外推能力。但是误差逆传神经网络模型需要大量的样本量来保证所求取参数的可靠性,但这在实际研究中很难做到,因而限制了误差逆传神经网络模型的应用。近年来人们提出了强制训练停止、复合模型等多种技术来提高误差逆传神经网络模型的外推能力,也提出了Garson算法、敏感性分析以及随机化检验等技术对误差逆传神经网络模型的机理进行解释。误差逆传神经网络模型的真正优势在于模拟人们了解较少或不确定性和模糊性较大系统的行为,这些是传统模型所无法实现的,因而是对传统机理模型的重要补充。

ARTIFICIAL NEURAL NETWORK AND ITS APPLICATION IN AGRICULTURAL AND ECOLOGICAL RESEARCH

Artificial neural network appeals to many ecologists because of the complexity of agroecological systems and its precise predicting ability to simulate vaguely understood and highly uncertain ecosystems. We focus on the introduction of the structure, algorithm and the applications for back-propagation artificial neural network (BPN) in this paper. People usually adopt BPN of a three-layer structure that can approximate functions of any complexity and is not easy to overfit because of its simple structure. The main principle of back-propagation artificial neural network is to adjust weights according to the errors of input neurons. In agroecological research, neural networks are always employed to predict crop yield, biomass yield, relationships between organisms and environmental factors and so on. Previous studies indicate that neural networks greatly outperform linear models, while the accuracy of the results produced by neural networks are very similar to that of algorithmic models. Moreover, neural networks can extrapolate to some degree when enough training data are provided. However, neural networks require a large number of samples to guarantee the robustness of its parameterization, which seems unrealistic for complicated networks to collect such large data sets from crop growth experiments involving countless small plot trials over multiple site-years. The requirement of large amount of training data has hindered the application of neural networks. Some techniques, such as early stopping, jittering and metamodels, have been advanced to induce generalization of neural networks, while techniques such as Garson's algorithm, sensitivity analysis and randomization test are advanced to explain the mechanisms of neural networks. The advantage of neural network lies on its ability to precisely simulate the vaguely understood and uncertain ecosystems, which cannot be realized by traditional approaches. As a nonlinear approximator, artificial neural network is an important tool complementary to comprehensive models.