基于遗传神经网络的全国小麦条锈病长期气象预测

为了提高反向传播（BP）神经网络模型预测小麦条锈病发病率的准确性和效率，以上年1月-当年3月组合的120个大气环流特征量为基础，定量分析大气环流特征量与全国小麦条锈病发病率之间的相关性并从中筛选出主要的影响因子；对影响因子进行主成分分析（PCA），提取累计贡献达到85．46％的前10个主成分作为预测因子；利用逐步回归、BP神经网络及遗传算法（GA）优化的BP神经网络三种模型进行预测，三种模型的预测精度均在80％以上，其中GA—BP神经网络模型的精度最高，达92．6％，而其训练步长仅为标准BP神经网络的1／4左右。通过PCA简化网络结构，同时运用GA优化网络初始权值和阈值，GA—BP神经网络模型可以较好的预测小麦条锈病的发病率。

Long-term Meteorological Prediction of Countrywide Wheat Stripe Rust by Genetic Neural Network

This study aims at improving the accuracy and efficiency of the back propagation （BP） neural network in wheat-stripe-rust prediction. First, the correlation between the atmospheric circulation and the occurrence of wheat stripe rust in China were examined quantitatively. The significant covariates, i.e. factors for the occurrence of wheat stripe rust, were then identified from scores of atmospheric circulation variables at multiple time scales from last January to March. Through the principal component analysis （PCA） , the first several components that together explained over 80% of the identified factors were employed as predictors. The prediction experiment was carried out by three models, i.e. the stepwise regression model, the back propagation （BP） neural network model, and the genetic algorithm （GA） optimized BP neural network （GA-BP） model. The experimental accuracy was above 80% with each of the models; while the GA-BP model, with only a quarter of the training epochs of the standard BP neural network, scores the highest accuracy of 92.6%. It was indicated that the prediction of the occurrence of wheat stripe rust could be much promoted using the BP neural network with its network structure simplified through PCA and with its initial weights and threshold optimized by GA.