A model for soluble protein content detection of walnuts based on near infrared spectroscopy

Abstract: 【Objective】The primary goal of this research was to compare the modeling methods of Support Vector Regression (SVR) and Back-Propagation network and seak for the best pre-processing combination method with the modeling method. The protein content prediction model of walnut kernel was established using near-infrared spectroscopy technology. The protein content of walnut kernels is one of the important indicators affecting the quality of walnuts. At present, the detection method for protein content mainly depends on the national standard method, the process is cumbersome, and multiple indicators can not be determined at the same time.【Methods】180 walnut samples from 9 different or-chards were collected as research materials, the row spacing of the walnut trees in each orchard is 4 m × 6 m, and the tree age is 10 years. Firstly, the diffuse reflectance spectra of the samples were collected at room temperature (around 25 ℃) using an Antaris Ⅱ Fourier transform NIR spectrometer made in the United States, the spectra were obtained in the wave number range of 4000-10 000 cm-1 (780-2500 nm) with a resolution of 8 cm-1 and a gain of 2x. With the built-in background of the instrument as the reference, each sample was scanned 3 times repeatedly as the original spectrum of the sample. The average spectrum was obtained after 32 final scans. Secondly, the protein content of 180 walnuts was determined by the Kaumas Brilliant Blue method. After the six outliers were removed by the Marginal distance, the SPXY algorithm was used to divide 174 samples into 132 Correction sets and 42 Validation sets in a 3∶1 ratio. The Competitive Adaptive Reweighted Sampling (CARS) method was used to extract the feature wavelengths. The spectral information was processed by six different pretreatment combination methods: Standard Normal Variables transformation (SNV), First-Derivative (FD), Multivariate, Scattering, Correction (MSC) +First-Derivative (FD), Second-Derivative (SD), Savitzky Golay convolution smoothing (SG)+Second-Derivative (SD), Standard Normal Variables transformation (SNV)+ Second-Derivative (SD). The Root Mean Square Error (RMSE), coefficient of determination (R2 ), Residual Prediction Deviation (RPD) were used to determine the optimal model and to compare the walnut protein prediction models established by different preprocessing combination methods with BP neural network method and SVR.【Results】Comparing the SVR with the BP neural networks, the R2 of the SD+BP neural network, MSC+BP neural network, SG+SD+BP neural network and SNV+SD+BP neural network Correction set and Validation set were below 0.8, only the R2 of the MSC+FD+BP neural network Correction and Validation sets reach was above 0.8. Moreover, the maximum RPD was 2.856. Although R2 of the Correction set for the FD+BP neural network was 0.845 7, it was quite discrepancy from the Validation set. The R2 of Correction set and Validation set for SD+SVR, MSC+SVR, MSC+ FD+SVR were all lower, Only SG+SD+SVR and SNV+SD+SVR Corrected and Validated sets had R2 above 0.8, although the R2 of the Correction set of the FD+BP neural network was 0.820 0, but the R2 of Validation set was only 0.770. Compared with SG+SD+SVR and SNV+SVR, MSC+FD+BP neural network had smaller differences in R2 of Correction set and Validation set, RMSE of Correction set and Validation set, the highest RPD. That is, the R2 of the MSC+ FD+BP neural network Correction set was 0.871, RMSEC was 0.089 5, and the RPD was 2.875; the R2 of the validation set was 0.825, RMSEP was 0.105 9, and the RPD value was 2.233. Therefore, MSC+FD+BP neural network built prediction model performs better than SG+SD+SVR and SNV+SVR.【Conclusion】The results showed that the BP neural network algorithm had better model quality than the SVR algorithm in walnut kernel soluble protein content prediction modeling in characteristic bands. The MSC + FD + CARS + BP neural network modeling method would be more suitable for the prediction of the soluble protein content of walnut kernel, which would provide a reference for the analysis of walnut kernel quality using near-infrared spectroscopy.