The Forecasting Of Cotton Bollworm's Occurrence Degree Based On Improved Neural Network Model

Cotton bollworm, Helicoverpa armigera(Hübner),widely distributed in the world as a key target pest, seriously damage many kinds of agricultural crops,making heavy loss in agricultural economy. The forecasting of cotton bollworm can be considered an input/output system and simulated by a neural net. We can improve the conventional neural network model for high efficient operation and accurate prediction. Cotton bollworm occurrence data in 1959-2007 were collected from Yuncheng County, Shandong Province,an important cotton growing area,which are characterized by high dimension with 49 samples and 40 candidate prediction factors (including meteorological factors,base numbers of cotton bollworm, etc).By using dimension-reduction algorithms in neural nets,We built three models,FABP, FA-RBF and PCA-Elman,to predict the pest. The FABP is BP algorithm model, while the FA-RBF is RBF algorithm model, which is based on FA. The PCA-Elman is the Elman algorithm based on PCA. Our aim is to get lower dimension, less input, easy design, simple structure, good convergence-speed, short run-time, and high efficiency of neural nets.Similarly, we built CA-BP model based on BP algorithm and CA. With cluster analysis, CA-BP model can cluster training samples for treatment,divide samples into 3-6 categories, even clarify samples, train more precise net model. The simulation samples were taken DA, respectively, and predicted by the corresponding networks, in order to improve prediction accuracy.Likewise, we explored BP algorithm based on GA, to optimize neural net with algorithm, to overcome neural net drawbacks. We used GA to optimize weigh values and number of neurons in hidden layer of BP net, in order to prevent BP from trapping into local minimum, low convergence speed and uncertainty number of neurons in hidden layer.Finally, we got models based on FA-CA-BP algorithm and FA-CA-GABP algorithm, combining the advantage of neural net with reducing dimension of character, clustering analysis and genetic algorithm. We aimed at improving the neural net performance to get good prediction on pest forecast.Through reducing-dimension analysis, the result showed as follows: though it lose certain amount of information, it decrease data redundancy by the prediction factors. Under the circumstance of the same intensity, it saves run-time without lowing prediction accuracy, increases convergence speed; the new model can also improve prediction accuracy to some certain extent, although clustering analysis adds complicated operations; the new model can raise success rate for train, improve efficiency of neural net, although GA optimization wastes a lot of time as well. Combining the above mentioned advantages, we can build more practical models and provide new models for studying pest ecology, by means of powerful computing software and programming tools. In the end, we can get good results easily in spite of complicated operations.At last, by the tools of via C++ and MATLAB, we developed operation platform with good human-computer interaction, simple operation, easy extension and application.