| Precise predicting and forecasting methods are necessary in fruit pest and disease control. At present, practical experience control is mainly used in the fruit production, resulting in delays in the treatment procedures and a decrease of fruit production and quality. In order to improve the accuracy of prediction, Grapholitha molesta (Busck) was selected as a study object, the key affecting factors were screened out by correlation analysis and expert knowledge, according to the meteological factors. Prediction models of the adult peak period and the occurrence degree of G molesta were established by Support Vector Machine (SVM), based on the statistical learning theory, in order to provide an efficient prediction method. Finally, a prediction system of G molesta was designed and developed based on SVM and Geographic Information System (GIS) through the integration of poly-information technology, providing a prediction reference method and technology platform for other fruit pests. The main results obtained were as follows:1. The prediction accuracy was ensured by proposing the methods and processes in each prediction link of G. molesta. Considering the shortage of previous methods of analysis, in which the meteorological data was separated by 10-day periods or months, puffing treatment technology was applied to deal with meteorological factors, and as a result, the factor number increased evidently. Consequently, the difficulty of screening the affecting factors due to the short interval of the pest's occurrence was overcome successfully. The continuity and accumulates of the meteorological factors were also manifested fully, which will be helpful to screen out the factors more satisfied with the biological rules of G. molesta. To avoid the significant level as the unique screening standard, the biological rule, correlation coefficient and the period of the meteorological factors were used as the screening standards to select the relative factors. Therefore, the screening efficiency of the affecting factors was increased.2. The key affecting factors of the adult peak period and occurrence degree of G. molesta were selected and confirmed by combining the mathematical statistics method and expert knowledge. The continuity of most selected factors demonstrated the continuity of occurrence and development of G. molesta. For the adult peak period, the affecting factors of different generations were varied. The affecting factors were not limited to the range of temperature and humidity. Temperature, a co-factor, was inversely related to the adult peak period, and humidity, a proportional factor, turned out to be the key affecting factor of the adult peak period from 2nd to 4th generation. At the same time, low temperature and rainfall had an effect on adult peak period of some generations. For the occurrence degree, the affecting factors, mostly related with humidity or rainfall, were proportional; the more rainfall or the higher of humility, the higher occurrence of G. molesta. Nevertheless, in fall and winter, the influence of humility and rainfall were inverse with the occurrence degree. In addition, temperature and sunshine affected the occurrence degree of some generations. The relationship of occurrence and development of G. molesta with the mathematical factors in corresponding periods were described quantitatively through the screened affecting factors.3. Based on the regression and classification of LibSVM, the models of adult peak period and occurrence degree of G molesta from the overwintering generation to 4th generation were established. Through the parameters selected, the models were optimized, and therefore, the accuracy of the prediction was enhanced. The results were compared with the models established by BP Neural Networks. The accuracy of the prediction model of the adult peak period for each generation based on SVM was 93.6% on average, about 10.9% higher than that by BP Neural Networks (82.7% on average). Similarly, the accuracy of the prediction model of the occurrence degree for each generation, established by SVM (82% on average) was significantly higher than that by BP Neural Networks (63.2% on average). The mean square error of the model of SVM was less than that of BP Neural Networks. In conclusion, the models based on SVM could predict more precisely and stably than that of BP Neural Networks.4. The prediction system of G. molesta based on SVM and GIS was established by the development platform of MapObjects and C#.NET. The functions of data management, inquiry, statistics, prediction and forecasting, control decision, thematic map development, information publishing etc. were included. The accuracy of prediction was increased by using the SVM, and the system overcame the shortage of BP Neural Networks. This system could be extended to predict other pests and provide a technology platform of prediction and information management of fruit pests.
|
PDF Full Text Request