应用集成BP神经网络进行田间土壤空间变异研究

以英国北爱尔兰Hayes的一块牧草地为研究区，将所有样点分为独立的训练和检验数据集，并在训练样点集的基础上设计了其他4种样点布局方案，以研究神经网络集成技术应用于田间土壤性质空间变异性的可能性。与广泛应用的克里格法的试验结果相比，集成BP神经网络的插值结果精度与之基本相当，尤其是在样点分布较稀疏和样点数较少的情况下，集成BP网络表现出明显的优势；由于神经网络集成方法对样本数据的分布没有任何要求，因此具有较广泛的应用前景和潜力，并在不符合克里格法对样本数据分布要求的情况下是一种可行的替代方法。

Spatial variety of soil properties by BP neural network ensemble

A 7.7 hectare silage field at Hayes, Northern Ireland, UK, was selected for the study, the samples were divided into training and validation dataum sets, several sampling distributions were designed based on the whole training sample distribution. Compared with the Root Mean Square Error (RMSE) achieved by Kriging, the accuracy achieved by BP neural network ensemble was very near or even better. While the interval between samples enlarged, the accuracy by BP neural network ensemble exceeded the accuracy achieved by Kriging. And one of the advantages was no statistical inference to samples for neural network ensemble. The potential ability of neural network ensemble was also discussed.