Introduction
Corn is an important food crop. The content of water, protein, fat, sugar and other major chemical components in corn will directly affect the economic benefits of corn. Determination of chemical content has become an important link in the evaluation of raw material quality. As the most basic material in production, the quality of corn seed directly affects the yield and quality of corn. The detection of maize quality indicators (moisture, protein, starch, etc.) is commonly used physical and chemical methods, the detection of safety indicators (toxins, etc.) using physical or chemical methods such as liquid phase, available cold immersion method, etc. to analyze the quality of the germplasm, but all of these methods will cause damage to the sample itself, there are longer processing time as well as need to be operated by professional personnel, the cost of the instrument and other shortcomings. Therefore, it is particularly important to explore a non-destructive and rapid detection technology for maize. Near-infrared (NIR) spectroscopy has the advantages of no complex and time-consuming pre-treatment of samples, no loss, simultaneous analysis of multiple components, and non-polluting detection, and has received widespread attention in recent years. Near-infrared spectroscopy is a detection technology that uses the absorption, scattering, reflection and transmission characteristics of the material to be measured to analyze the sample, through the absorption spectrum of the sample and the results of physical and chemical analysis of the sample can be qualitative or quantitative analysis. The detection steps of NIR spectroscopy are to use chemometrics to preprocess the NIR spectral data and establish a model, to test the prediction set of the sample through the model, to verify whether the model is accurate, and to evaluate and optimize the model. Commonly used processing methods in NIR spectroscopy technology, due to the noise interference caused by the strong background information in the NIR spectrum and the presence of redundant variables, resulting in more difficult to extract the information related to the detection target from the NIR spectrum of the sample, so the spectral data need to be preprocessed. Commonly used spectral preprocessing methods include denoising autoencoder (DAE), orthogonal signal correction (OSC), standard normal transform (SNV), multiple scattering correction (MSC), and so on.
Evaluation indexes of near-infrared spectroscopy technology models
Evaluation index of quantitative model
The essence of evaluating the prediction accuracy of NIR spectroscopy quantitative model is the proximity of the prediction results of the model to the results of the samples. Evaluation of prediction models generally uses parameters such as correction coefficient of determination (R2c), verification coefficient of determination (R2v), correction coefficient of correlation (Rc), verification coefficient of correlation (Rv), corrected root-mean-square error (RMSEC), verified root-mean-square error (RMSEV), and relative analysis error (RPD). （The coefficient of determination (COD) and correlation coefficient (RC) are the criteria for the correlation between the predicted values and the sample set of true values detected by the chemical method, and the larger the R2c, R2v, Rc, and Rv are, the better the modeling effect is considered to be; the RMSEC and RMSEV are the measures for the size of the differences between the predicted values of the calibration set and the validation set, and the true values detected by the chemical method, and the smaller the RMSEC and RMSEV are, the better the modeling effect is considered to be. RMSEC and RMSEV are measures of the size of the difference between the predicted values in the correction set and the validation set and the true values detected using the chemical method. The smaller the RMSEC and RMSEV, the better the performance of the constructed model is considered to be; RPD is a measure of the reliability of the model.
Qualitative model evaluation index
NIR spectroscopy is mostly used for sample classification in qualitative analysis, and the commonly used determination indexes are correct rate, sensitivity, specificity, etc.