基于栈式自编码神经网络的包衣单籽粒玉米品种识别

常规近红外定性识别研究中,玉米籽粒为表皮裸露状态,未经种衣剂覆盖处理,但是在实际农业生产中,为抵御病虫害侵袭,提高玉米种子发芽率,达到保产增产的功效,玉米种子常需经种衣剂包裹处理。玉米种衣剂的类型多样,对近红外光谱具有一定的吸收,因此种衣剂对近红外定性识别具有干扰作用。本文针对种衣剂对玉米品种识别准确性影响的问题,提出了一种基于栈式自编码神经网络（SAE）的近红外光谱定性建模方法。首先采用无种衣剂玉米籽粒光谱作为训练集,通过栈式自编码无监督学习算法与softmax分类器构建栈式自编码网络定性分析模型,再利用所建模型对有种衣剂玉米籽粒进行品种识别。实验结果表明,基于SAE的建模方法能够将种衣剂对玉米籽粒识别率的影响降低至3%以内。

Varietal Identification for Single Maize Seed Based on Stacked Auto Encoder Neural Network

In the conventional near infrared qualitative identification, the maize seed kernel epidermis was not treated with seed coating agent. However, in the actual agricultural production, in order to resist the invasion of diseases and insect pests, improve the germination rate, and achieve the effect of maintaining and increasing yield, maize seeds often need to be coated with seed coating agents. In reality, on the market, it is usually necessary to model maize seed kernels without seed coating to identify the ones with seed coating, so as to achieve the purpose of cracking down fake and shoddy products. The maize seeds coating usually consist of a mixture of insecticides, fungicides, fertilizer, plant growth regulators and other ingredients. Their types are diverse and the components are different. These components contain hydrogen group organic compounds, which have certain absorption to near infrared spectrum. Therefore, the seed coating agent had an interference effect on near infrared spectroscopy qualitative identification, which reduced the performance of some conventional shallow learning model. According to the effects of seed coating on maize variety authenticity identification accuracy, a method of near infrared spectroscopy qualitative modeling based on stacked autoencoder (SAE) neural networks has been proposed. Firstly, taking maize seed spectrum without seed coating agent as the training set, a qualitative analysis model was constructed through SAE unsupervised learning algorithm and Softmax classifier. Then, based on this model, the authenticity of maize seeds with seed coating agents was identified. The experimental results showed that, by using the method based on SAE, the effect of seed coating on maize varietal authenticity recognition rate was controlled within 3%.