近红外光谱的水稻抗性淀粉含量测定研究

用化学法测定水稻抗性淀粉含量耗时长、成本高，为此，探索了基于近红外光谱技术(NIRS)的水稻抗性淀粉含量测定新途径。首先，采集了62份抗性淀粉含量差异较大的水稻的光谱数据，将光谱数据和已测定的化学值数据导入化学计量学软件，采用偏最小二乘法(PLS)建立了抗性淀粉含量的近红外定标模型，对不同预处理得到的预测模型进行了内部验证和外部验证。结果如下：内部交叉验证方面，未处理、MSC+1thD预处理、1thD +SNV预处理的决定系数(R2)分别为0.920 2，0.967 0，0.976 7，预测均方根误差(RMSEP)分别为1.533 7，1.011 2，0.837 1。外部验证方面，未处理、MSC+1thD预处理和1thD +SNV预处理的决定系数(R2)分别为0.805，0.976，0.992，绝对误差平均值分别为1.456，0.818，0.515，预测值和化学值之间没有显著差异(Turkey法多重比较)，说明以近红外光谱分析法代替化学测定法是有可能的。在不同预处理方法之中，1thD+SNV的预处理方法无论内部验证还是外部验证都具有较高的决定系数和较低的误差值，定标模型精度更高，误差更小。

Research on Resistant Starch Content of Rice Grain Based on NIR Spectroscopy Model

A new method based on near-infrared reflectance spectroscopy (NIRS) analysis was explored to determine the content of rice-resistant starch instead of common chemical method which took long time was high-cost. First of all, we collected 62 spectral data which have big differences in terms of resistant starch content of rice, and then the spectral data and detected chemical values are imported chemometrics software. After that a near-infrared spectroscopy calibration model for rice-resistant starch content was constructed with partial least squares (PLS) method. Results are as follows: In respect of internal cross validation, the coefficient of determination (R2) of untreated, pretreatment with MSC+1thD,pretreatment with 1thD+SNV were 0.920 2,0.967 0 and 0.976 7 respectively. Root mean square error of prediction(RMSEP)were 1.533 7,1.011 2 and 0.837 1 respectively. In respect of external validation, the coefficient of determination (R2) of untreated, pretreatment with MSC+1thD, pretreatment with 1thD+SNV were 0.805, 0.976 and 0.992 respectively. The average absolute error was 1.456, 0.818, 0.515 respectively. There was no significant difference between chemical and predicted values (Turkey multiple comparison), so we think near infrared spectrum analysis is more feasible than chemical measurement. Among the different pretreatment, the first derivation and standard normal variate (1thD+SNV) have higher coefficient of determination (R2) and lower error value whether in internal validation and external validation. In other words, the calibration model has higher precision and less error by pretreatment with 1thD+SNV.