基于改进深度置信网络的大棚冬枣病虫害预测模型

发布时间：2018-05-28 10:35

本文选题：病害 + 预测　；参考：《农业工程学报》2017年19期

【摘要】：导致冬枣病虫害发生的原因很多而且很复杂,利用传统的数学方法和神经网络(neural network,NN)很难建立正确的病虫害预测模型。由于典型的深度置信网络(deep belief network,DBN)的各层之间缺乏有监督训练,使得网络误差逐层向上传递,降低了预测模型的预测率。针对这些问题,引入冬枣病虫害的先验信息,提出一种基于环境信息和改进DBN的冬枣病虫害预测模型。在该模型中,通过无监督训练和有监督微调从冬枣生长的环境信息序列中获取可表征冬枣病虫害发生的深层特征的隐层参数,并形成新的特征集,然后在预测模型的顶层通过一个后向传播神经网络(back propagation neural network,BPNN)进行病虫害预测。从2014—2017年的4 a时间内,利用农业物联网传感器采集30个大棚冬枣常见的2种虫害和3种病害发生的环境信息序列6 000多条,由此验证所提出的预测模型,平均预测正确率高达84.05%。与基于强模糊支持向量机、改进型NN和BPNN的3种病虫害预测模型进行了试验比较,预测正确率提高了20多个百分点。试验结果表明,该模型极大提高了大棚冬枣病虫害的预测正确率。该研究可为大棚冬枣病虫害预测提供技术参考。
[Abstract]:There are many and complicated reasons for the occurrence of diseases and insect pests in winter jujube. It is difficult to establish a correct prediction model of diseases and pests by using traditional mathematical methods and neural network (NN). Due to the lack of supervised training among the layers of the typical deep belief network, the network errors are transmitted upward, and the prediction rate of the prediction model is reduced. In order to solve these problems, a prediction model of winter jujube disease and insect pests based on environmental information and improved DBN was proposed by introducing the prior information of winter jujube pests and diseases. In this model, the hidden layer parameters can be obtained from the environmental information sequence of Dongzao jujube growth by unsupervised training and supervised fine-tuning, and a new feature set is formed. The disease and insect pests are predicted at the top of the prediction model through a back propagation neural network (BPNN). In the period of 2014-2017, more than 6 000 environmental information sequences of two common pests and three diseases of winter jujube in greenhouse were collected by using agricultural Internet of things sensors, and the proposed prediction model was verified. The average prediction accuracy was 84.05%. Compared with three forecasting models based on strong fuzzy support vector machine, improved NN and BPNN, the accuracy of prediction is increased by more than 20 percentage points. The experimental results show that the prediction accuracy of winter jujube diseases and insect pests in greenhouse is greatly improved by this model. This study can provide technical reference for prediction of diseases and pests of winter jujube in greenhouse.
【作者单位】：西京学院信息工程学院;
【基金】：国家自然科学基金项目(61473237) 陕西省自然科学基础研究计划(2016GY-141)
【分类号】：S436.65;TP18

【相似文献】