ISSN 1000-1239 CN 11-1777/TP

• 论文 •

一种新的神经网络集成方法及其在精准施肥中的应用

1. 1(吉林大学计算机科学与技术学院 长春 130012) 2(符号计算与知识工程教育部重点实验室(吉林大学) 长春 130012) 3(吉林农业大学信息技术学院 长春 130118) (liudy@jlu.edu.cn)
• 出版日期: 2010-09-15

A Novel Neural Network Ensemble Method and Its Application in Precision Fertilization

Yu Helong1,2,3, Liu Jie1,2, Jin Di1,2, Yang Yupeng1,2, and Liu Dayou1,2

1. 1(College of Computer Science and Technology, Jilin University, Changchun 130012) 2(Key Laboratory of Symbolic Computing and Knowledge Engineering (Jilin University), Ministry of Education, Changchun 130012) 3(College of Information Technology, Jilin Agricultural University, Changchun 130118)
• Online: 2010-09-15

Abstract: In order to solve the problem of precise fertilization rate determination, a novel neural network ensemble method is introduced in this paper. In this method, the method of sampling with replacement is used to produce neural network individual set. A novel formula measuring the network similarity is given and Freys clustering algorithm AP is used to select the networks with high precision and greater diversity. Then by the Lagrange multiplier ensemble(LME) and forecasting effective measure mensemble(FEME) method, these selected networks are combined. The experiment on the standard dataset shows that, LME algorithm has higher accuracy and stronger generalization than the single neural network. Furthermore, as a linear weighted ensemble method, LME is better than FEME. Generally, the root mean squared error (RMSE) decreases when the subnets number and cluster number increases, but when the number reaches a level, the decreasing trend of RMSE becomes slow. Then according to the fertilizer effect data on maize field in black soil, LM ensemble is used to build the precision fertilization model, where soil nutrient and target yield are taken as inputs and fertilization rate is taken as output. The practice shows that LM ensemble based fertilization model is better than traditional fertilization models and the existing neural network based fertilization models.