基于近红外光谱的粳稻种子快速鉴别方法研究

黑龙江省是我国最大的粳稻产区和商品粮生产基地。水稻种植过程中，选择合适的水稻品种是实现高产的关键环节。在农业生产中，水稻品种的选择受多方面因素影响，一般说来，同一积温带所种植的不同水稻品种在外观上差别不大，甚至没有差别，很难通过肉眼观察进行准确区分。为了快速鉴别肉眼不便区分的不同类别粳稻种子，提出了一种基于近红外光谱技术的粳稻品种快速无损鉴别方法。以黑龙江垦区大量种植的3种不同品种的粳稻种子（垦粳5号、垦粳6号和绥粳4号）作为研究对象，每个品种选取40个样本，其中30个样本做为建模集，10个样本作为预测集，扫描获取全部120个样本的近红外光谱数据。对原始光谱数据（11 520～4 000 cm-1）两端进行裁剪，选取吸光度较强的8 250～5 779cm-1范围内的光谱数据进行研究。首先建立参照模型，即直接对光谱数据建立BP模型1, 同时光谱数据经过一阶导数和Savitzky-Golay平滑预处理后建立BP模型2。模型1的分类正确率为93.3%，预测集均方根误差RMSEP=0.232 8，迭代时间t=3 882.9 s。模型2的分类正确率为100%，RMSEP=0.070 6，迭代时间t=954.5 s。比较两种模型的评价参数RMSEP发现FD+SG预处理可以提高模型的预测能力，但是由于两种模型未进行降维处理，数据量过大，模型的输入节点过多，迭代时间太长，不利于实际应用。因此利用小波变换多分辨率的特点对数据进行降维处理，采用预测集残差平方和Press值作为评价指标，在多个小波类别和参数中选取分解尺度为5的sym2（symlet2）小波对光谱数据进行压缩和降维处理，将光谱数据由601维降到21维。以小波变换结果作为神经网络输入，建立模型3，并与模型1比较，模型3的分类正确率为93.3%, RMSEP=0.225 0, 迭代时间t缩短至198.9 s，比较结果显示小波降维可以减少神经网络的输入，简化神经网络的结构，从而提高迭代速度，但对提高模型的预测能力效果不明显。上述三种模型比较结果表明，FD+SG预处理可以提高模型的预测能力，小波降维可以提高模型的迭代速度，综合上述三种模型的比较结果分析，最终建立“FD+SG+小波降维”的21输入、15个隐层、3个输出的神经网络鉴别模型4，其分类正确率达100%，RMSEP=0.029 3, 迭代时间为98.8 s，表明模型4能够完全实现对三种不同水稻品种的快速、准确、无损鉴别。因此，所提出的基于近红外光谱的小波降维和反向传播人工神经网络鉴别模型的方法完全可以用于粳稻种子的快速无损鉴别，同时也为其他农作物种子的快速鉴别提供了参考。

Rapid Discrimination of Japonica Rice Seeds Based on Near Infrared Spectroscopy

Heilongjiang Province is the largest japonica rice producing area and commodity grain base in China. In the process of rice planting, selecting suitable rice varieties is the key to achieving high yield. In agricultural production, the selection of rice varieties is influenced by factors in many aspects. Generally speaking, different rice varieties planted in the same temperate zone have little difference in appearance, or even no difference. It is difficult to make an accurate distinction by visual observation. In order to accurately distinguish different varieties of japonica rice seeds that are difficult to distinguish by naked eyes, a rapid non-destructive discrimination method for japonica rice based on near-infrared spectroscopy (NIRS) was proposed. 3 varieties of japonica rice seeds (seeds 5th, seeds 6th and Sui japonica 4th) planted in Heilongjiang reclamation area were selected as the research object. For each variety, 40 samples were selected, 30 of which were used as modeling set and 10 as prediction set. The NIRS data of all 120 samples were obtained by scanning. The noise at both ends of the original spectral data (11 520～4 000 cm-1) were clipped, the spectral data in the range of 8 250～5 779 cm-1 with strong absorbance were selected as the research band. Firstly, a reference model was established, that is, BP model 1 was established directly from raw spectral data, and BP model 2 was established from the spectral data preprocessed by first derivative (FD) and Savitzky-Golay (SG). The classification accuracy of model 1 was 93.3% with RMSEP=0.232 8, and the iteration time was t=3 882.9 s. The classification accuracy of model 2 was 100% with RMSEP=0.070 6, and the iteration time was t=954.5 s. Comparing the evaluation parameter RMSEP of the two models, it was found that FD+SG preprocessing can improve the prediction ability of the model. However, because the two models do not reduce the dimension, the amount of data is too large, the input nodes of the model are too many and the iteration time is too long, which is not conducive to the practical application. Therefore, the wavelet transform with multi-resolution characteristic was used to reduce the dimension of the data. The residual sum of squares of the prediction set (Press value) were used as the evaluation index. Sym2(symlet2) wavelet with decomposition scale 5 was selected to compress and reduce the dimension of the spectral data from 601 dimension to 21 dimension. The results of wavelet transform were used as the input of BP model 3, which was compared with model 1. The classification accuracy of the model 3 was 93.3% with RMSEP=0.225 0, and the iteration time was shortened to 198.9 s. The comparison results showed that dimensionality reduction based on wavelet transformation can reduce the input of the neural network, thus simplifying the structure of the neural network and improving the iterative speed, but the effect of improving the prediction ability of the model is not obvious. The comparison results of the three models showed that FD+SG preprocessing can improve the prediction ability of the model, and the wavelet transform can improve the iteration speed of the model. Based on above analysis results, a neural network discrimination model 4 with 21 inputs, 15 hidden layers and 3 outputs of FD+SG+wavelet transform was established. Moreover, its recognition rate of classification was 100% with RMSEP=0.029 3 and the iteration time was t=98.8 s, which could identify three different japonica rice varieties quickly, accurately and non-destructively. Therefore, the method of wavelet reduction and back propagation artificial neural network (BP) discrimination model based on near infrared spectroscopy can be used for rapid and nondestructive discrimination of japonica rice seeds, providing a reference method for other crop seeds recognition.