A Selective Approach To Neural Network Ensemble Based On Network Clustering Technology And The Application Of The Agriculture

This paper is supported by National High Technology Research and Development Program of China (Digital Agricultural Knowledge Grid Technology Research and Application Topic). In this paper, a new method of neural network ensemble called NCSNNE is proposed, it can effectively solve the problem of precise fertilization rate determination. We study on how to use the network clustering algorithm to select the networks with high precision and great diversity. Then the selected networks are ensembled separately with the linear weighted ensemble and nonlinear ensemble methods. With the fuzzy system analysis method we implement the linear weighted ensemble method, in this method, the time-variant weight is considered to be a fuzzy coefficient. Then we introduced some concepts. Then we transform the target function into a linear programming model for combination forecasting model. With the neural network method we implement the nonlinear ensemble method.The basis of national economy is agriculture. To ensure the food safety and the sustainable development of the agriculture is the most important task in our country. Precision agriculture is a sophisticated technology based on modern information technology and intelligent equipment technology. It can be used to implement quantitative decision making,variable rate technology and located fertilization technology. In the case of lack of resources, people implement precise fertilization technology through modern industrial technology, so it is an intensively and efficient technology in agriculture. Precision agriculture is an innovative of modern agricultural technology, it is the inevitable outcome of the development of modern agriculture. Precision agriculture follows the latest high-tech developments, has advanced technology for production, and its integration of technology is better. Precision agriculture can improve the resource utilization, so by using the technology there have been enormous increases in agricultural productivity. Because of precision agriculture can effectively solve many problems in agriculture, the goal of our project is to build a knowledge consultation platform, by using this platform, users can get the imformation of precision agriculture,can analyze and process the data of precision agriculture and also can use appropriate software of the system. Finally, users can get lots of accurate and useful knowledge of precision agriculture by using our platform.Precision agriculture has many related technologies. We found precision fertilization is one of the most important technologies because it has the most extensive range of applications and it is the most mature technology in precision agriculture. By using the technology of precision fertilization, users can effectively solve this problem of precise fertilization rate determination, so it can increase the soil productivity. Users can get a big deal of benefit and at the same time they can save fertilizer. Obviously, precision fertilization is very efficient. There are some methods to solve the problem of precision fertilization, but all of them have some faults because they can not accurately fit the nonlinear relationship between soil nutrients,fertilization and other factors. So these methods are difficult to promote. Because of the method of neural network ensemble can improve on prediction accuracy, we use this method to solve the problem of precision fertilization.We have studied that if we want to improve the result of the ensemble, we have to make the errors of the neural networks to ensemble on different parts of the input space as well. To solve the difficult issue of crop precise fertilization rate, a novel neural network ensemble method called NCSNNE is presented. In this method, complex network clustering method is firstly applied to neural network ensemble. This method is a new solution of selective neural network ensemble. After individual neural networks are trained, a novel formula measuring the network similarity is given and it is used to calculate the similarity matrix. Convert the similarity matrix to a K-Nearest-Neighbor network, and then Yang's network clustering algorithm FEC is used to select the networks with high precision and great diversity. Then we have got some clusters, so we can choose the network who has the highest precision in each cluster. These networks are called "elite network", they can be used for ensemble.This paper implemented the linear weighted ensemble method and nonlinear ensemble method. We implemented both of the two methods because they are all very popular in neural network ensemble. With the fuzzy system analysis method we implement the linear weighted ensemble method and with the generalized regression neural network we implement the nonlinear ensemble method. The selected networks are ensembled separately with the linear weighted ensemble,nonlinear ensemble methods and amalgamating the results produced with the two ensemble methods. Finally, we found the last method's result is the best. So the prediction result is achieved by amalgamating the results produced with the two ensemble methods. This method was developed to solve the difficult issue of crop precise fertilization rate. The test result carried out at the corn plot in Yushu City revealed that it is better than the traditional fertilization model, linear weighted ensemble model and nonlinear ensemble model. In addition, its generalization ability is also very strong.Finally, we design and implement the Precision Fertilizer Decision Service Component(PFDSC). This service component is a very important element in our project. And it is the application of the study result. In the course of implementing the service component, we use the method NCSNNE. So the service component can provide users with the capability of precision fertilizer. The service component has three function modules:thaining module,choosing module and forecasting module. They correspond to the three key steps of the NCSNNE method. Thaining module provides a way to train the networks; choosing module provides a way to select the networks with high precision and great diversity and forecasting module provides a way to calculate the result. This project's architecture is based on Service-Oriented Architecture(SOA). We choose WCF technology to design it, so we have to use service component technology. By using this architecture, the system can be extended easily.