基于BP人工神经网络方法的广西稻飞虱发生等级预测  

Predicting occurrence degree of rice planthoppers in Guangxi Province based on BP artificial neural network method

作　　者：何燕 何慧[2] 孟翠丽[1] 谢茂昌[3] 龙梦玲[3] 李玉红[1] 

机构地区：[1]广西壮族自治区气象减灾研究所/国家卫星气象中心遥感应用示范基地,南宁530022 [2]广西壮族自治区气候中心,南宁530022 [3]广西壮族自治区植保总站,南宁530022

出　　处：《生态学杂志》2014年第1期159-168,共10页Chinese Journal of Ecology

基　　金：广西自然科学基金项目(2011GXNSFA018098和桂科自0832204);广西科技攻关项目(桂科攻10100004-6);国家公益性(气象)科研专项经费项目(GYHY201006026)资助

摘　　要：利用广西45个农业病虫测报站1988—2012年稻飞虱发生等级及1987—2012年气象要素、大气环流特征量等资料,采用模糊聚类分析、BP人工神经网络等方法,将广西早稻稻飞虱发生等级分为桂东、桂西南和桂西北3个区域,分别对各区域稻飞虱发生等级进行预测。结果表明:各区域稻飞虱发生等级与气象要素及大气环流密切相关,冬春季气温高、雨日多、湿度大、光照少等因素均利于稻飞虱发生,副热带高压、印缅槽和西南气流等均对稻飞虱发生等级有影响;各区域稻飞虱发生等级序列从冬春季气象要素、大气环流特征量中选择初选预测因子,对初选预测因子作EOF展开构造综合预测因子,分区建立预测模型并进行交叉检验表明,3个区域的人工神经网络模型平均拟合绝对误差比逐步回归模型分别小0.07、0.1和0.02,2011、2012年独立样本预测试验表明,人工神经网络模型和逐步回归模型的实际预测绝对误差为0.42和0.5,可见稻飞虱发生等级的BP人工神经网络预测模型比传统逐步回归模型有更好的拟合和预测效果,为稻飞虱与气象要素之间的非线性关系研究开拓新的思路。Based on data of occurrence degree of rice planthoppers from 45 agricultural pest monitoring stations in Guangxi Province during 1988 to 2012 as well as data of meteorological factors and atmospheric circulation characteristics during 1987 to 2012, three zones with different occurrence degrees of early rice planthoppers were divided： east Guangxi, southwest Guangxi, and northwest Guangxi. Occurrence degree of early rice planthoppers was predicted in each zone by fuzzy cluster analysis, and BP neural network. The results showed that the occurrence degree of rice planthoppers was closely correlated with meteorological factors and atmospheric general circulation in Guangxi. High temperature, frequent rainy days, high humidity and insufficient sunshine in winter and spring seasons were beneficial to the occurrence of rice planthoppers, and subtropical high, IndiaBurma trough and southwest airflow also affected the occurrence degree of rice planthoppers. Original predictive factors for the occurrence degree of early rice planthoppers in each zone were selected from the meteorological factors in winter and spring seasons and the atmospheric circulation characteristics to build comprehensive predictors using EOF decomposition method, and then prediction models for the occurrence degree of rice planthoppers were established in each zone. The crosstest showed that the average absolute fitting error was lower in BP neural network model than in step regression by 0.07 in east Guangxi, 0.1 in southwest Guangxi, and 0.02 in northwest Guangxi. The prediction using independentsamples in 2011 to 2012 showed that the mean predicted absolute errors were 0.42 for BP neural network model and 0.5 for step regression, indicating that the nonlinear correlation between rice planthoppers and meteorological factors is better predicted by BP neural network model.

关 键 词：模糊聚类 气象要素 综合预测因子 人工神经网络 稻飞虱发生程度 拟合误差 

分 类 号：S435.112.3[农业科学—农业昆虫与害虫防治]