中文摘要：

为解决利用回归模型进行作物种植密度和施肥量优化时存在的拟合精度差和准确性低等问题,该文提出一种基于BP神经网络的优化方法。以玉米作物为研究对象,选取种植密度、施氮量、施磷量、施钾量为试验因素,玉米产量为影响指标,设计4因素5水平正交旋转试验方案进行田间试验,获取不同种植密度和施肥量水平下的玉米产量。利用BP神经网络模型对试验数据进行函数拟合,拟合后运用该文提出的优化方法获得试验条件下红星农场德美亚1号玉米最佳种植密度9.32×10~4株/hm2、施N量139.5 kg/hm2、施P2O5量85.4 kg/hm2、施K2O量70.8 kg/hm2,该参数组合下玉米的最优产量为16 308.53 kg/hm2,高于二次回归模型优化得到的最高产量16 009.00 kg/hm2。以BP神经网络优化结果在试验区进行验证试验,获得试验方案下玉米产量为15 948.3 kg/hm2,试验与优化结果相对误差仅为-2.21%,表明该优化方法拟合函数精度高,优化结果准确,为解决农业生产领域中类似优化问题提供了一种可靠方法。

英文摘要：

Planting density and fertilizer application rate are the important factors affecting crop yield,and the unreasonable utilization has caused a series of serious consequences such as high cost,agriculture resources waste,agricultural non-point source pollution,and ecological environment deterioration and so on.In this study,a BP neural network-based optimization method of planting density and fertilizer application rate was proposed and tested for its feasibility by field experiments.The field experiment was carried out in Hongxing Farm of Heilongjiang,China（126°47′E,48°01′N）in 2014 and 2015.The experiment of 4 factors and 5 levels was designed by using the quadratic orthogonal rotation method.Four factors included planting density,N,P and K application rate.Five levels were considered as the equally spaced values taken from the planting 32.2-151.8 kg/hm2 and the K2O application rate of 25-115 kg/hm2.Among the 5 levels,the 0 level referred to the local experience value.A total of 36 plots were prepared and each plot had the width of 4.4 m and the length of 5 m.Maize（variety of Deyamei No.1）was planted on ridges in the width of 100 cm and in the height of 15 cm.Irrigation was not conducted during the experiment.The rainfall during the growing season of maize was 471.4 mm in 2014 and 460.7 mm in 2015.At harvest,the maize yield was determined.The field data was fitted using BP neural network model and regression method,respectively for optimization of planting density and fertilizer application rates.The BP neural work optimization method included model establishment and global optimization.The data was processed in Matlab.The results showed that the BP neural network model had higher determination coefficient of 0.98（P〈0.01）than the regression model（R2=0.87,P〈0.05）.Meanwhile,the former had smaller root-mean-square error of 189.89 kg/hm2 than the latter（464.25 kg/hm2）.It indicates that the BP neural network model was better in fitting the relationship between maize yield and fertilizer ap