基于改进BP神经网络的大豆病害检测

Soybean Diseases Detection Based on Improved BP Neural Network

大豆病害诊断是有效防治的先决条件。为此,针对传统BP神经网络在处理高维大豆病害数据时存在的时间复杂度高、诊断准确率低以及误差收敛缓慢且容易出现震荡现象的问题,提出了一种改进方法。该方法首先对高维大豆病害数据进行特征选择,去除"贡献"较小的特征,实现数据降维;然后,对传统BP算法进行改进,根据输出误差动态调整学习速率,并使用改进后的算法建立大豆病害检测模型。经实验测试,该方法在大豆病害诊断测试中准确率达96%以上,且各项统计指标、误差收敛速度及平稳性均优于传统BP神经网络,证明了其可靠性和高效性。

Soybean disease diagnosis is a prerequisite for effective prevention and treatment. Since traditional neural net- work has many weaknesses, such as high time complexity, low diagnostic accuracy, slow convergence and apparent con- cussion when detect the soybean disease with high dimension data, an improved method is proposed. Firstly, it reduces data dimension by selecting the features which are more important for disease detection. Secondly, it improves the tradi- tional BP algorithm by modifying the learning rate dynamically according to output error. Then it uses the improved algo- rithm to build soybean diseases detection model. The experimental results show that the detection accuracy is higher than 96%, what＇s more, the improved methods are superior to the traditional BP neural network in many aspects, such as, statistical indicators, convergence speed and stability. Thus, the reliability and efficiency of the proposed method are proved.