基于改进的BP神经网络集成的作物精准施肥模型

作物最优施肥量与土壤养分含量、产量之间存在复杂的非线性关系。为更加准确地模拟这种关系，提出一种改进的的BP神经网络集成方法。该方法采用K-均值聚类优选神经网络个体，采用拉格朗日乘子方法计算待集成的神经网络个体的权值。然后，基于农田肥料效应试验数据，以土壤养分含量和施肥量作为神经网络的输入，以产量作为神经网络的输出，建立了作物精准施肥模型。该模型通过求解一个非线性规划问题，能同时获得最大产量和最优施肥量。试验结果表明，在施肥模型的拟合精度方面，改进的神经网络集成方法（其均方根误差为64.54）明显优于单个神经网络方法（其均方根误差为169.74）。而且，作为一种定量模型，基于改进的神经网络集成的施肥模型优于传统施肥模型，能有效地指导精准施肥。

Crop precision fertilization model based on improved BP neural network ensemble

There exists obvious nonlinear relation between the optimal fertilization rate and soil and yield. In order to simulate this relation more accurately, a novel neural network ensemble method was proposed, where the K-means clustering was used to select better network individuals and Lagrange multiplier was used to compute the weight of network individuals. Based on the fertilizer effect data in the experimental field, taking soil nutrient and fertilization rate as inputs and taking yield as output, a crop precision fertilization model was constructed. By solving a nonlinear programming problem, both the maximum yield and the optimal fertilization rate were achieved. The results showed that the simulation error of the fertilization model based on neural network ensemble (root mean square error was 64.54) was much less than that of the fertilization model based on individual neural network (root mean square error was 169.74). Also, as a quantitative model, it is better than the traditional fertilization models and can be used to guide precision fertilization effectively.