Research On Nondestructive Detection Of Mature Tomato By Near Infrared Spectroscopy Technology

The tomato is the most important economic crops in solanaceae, and one of the most popular fruits and vegetables cultivated worldwide. Tomato fruit is rich nutrition, can be eaten raw, cooked, processed into tomato sauce, juice or whole fruit canning. In Europe and Amerca countries, China and Japan have a large area of greenhouse, plastic greenhouse and other prtected area facilities to cultivate tomatoes. Chinese cities generally have planting, cultivtion area continues to expand. Ripe tomatoes contain large amounts of lycopene, which is the main source of natural lycopene. Lycopene is not only important natural food colorants in today’s industry, the more important is that it is a strong antioxidant. The content of soluble solids, vitamin C and free amino acid is an important indicator to measure the quality of tomato, and is an important index to measure processing features. In order to guide people to eat tomato scientifically, and provide the basic data for tomato processing enterprises, it has great signiicance to detect the content of nutrient components in tomato nondestructively.In this study, nine varieties of tomatoes as the research object,108 samples collected, using near infrared spectroscopy and chemometrics methods to detect nondestructively the contents of lycopene, total free amino acid, vitamin C and soluble solids in tomato. Analyzed the different role of the different conventional spectra pretreatment methods in near infrared spectral modeling by partial least squares(PLS) regression, determined the optimum pretreat-ment method for different components according to the model’s evaluation parameters. Four kinds of methods were adopted to choose different composition characteristic spectrum inter-val, determined the best characteristic wavelength of each component in the different meth-ods. Using wavelet and wavelet packet to study de-noising of tomato spectrum, and deter-mined the best parameters of different wavelet and wavelet packet.The main contents and conclusions of this study are as follows:(1) Elimination of abnormal samples. Using spectral out point diagnosis function, to di-agnose the abnormal samples, the results show that the spectral acquisition without exception. Leverage value and residual test combined with the one by one recycling method, analyzed and eliminate the abnormal concentration samples. At last, three lycopene concentration ab-normal samples, one vitamin C concentration abnormal sample, two free amino acid abnormal samples, were eliminated from samples.(2) The best conventional pretreatment methods of near infrared spectrum. In 10 kinds of pretreatment methods, the optimal processing method of lycopene’s near infrared spectrum was average centralized and Norris first derivative method (Mean Centering+NorrisFD), the correlation coefficient R of the PLS model was 0.7894, the standard deviation correction (RMSEC) was 17.12 ug/g, the standard errors of prediction (RMSEP) was 18.56 ug/g; In the pretreatment method of vitamin C spectra, the optimal processing method was average cen-tralized and Norris first derivative method (Mean Centering+NorrisFD), the correlation co-efficient R of the PLS model was 0.8768, the RMSEC was 0.5804 mg/100g, the RMSEP was 0.5948mg/100g; In the total free amino acid in the near infrared spectra pretreatment method, the optimal treatment method was the variance ratio and Norris first derivative (Variance Scaling+NorrisFD), the correlation coefficient R of the PLS model was 0.8635, the RMSEC was 16.04 ug/100g, the RMSEP was 16.50 ug/100g; In soluble solids in the near infrared spectra pretreatment method, the optimal processing method was average centralized and Sa-vitzky-Golay first derivative (Mean Centering+FD+S-G), the correlation coefficient R of the PLS model was 0.8913, the RMSEC was 0.454%, the RMSEP was 0.480%.(3) Spectrum characteristic wavelength selection methods research. Using backward in-terval PLS (Bi-PLS), Synergy interval PLS (Si-PLS) and uninformative variable elimination PLS (UVE-PLS) and genetic algorithm PLS (GA-PLS), and combining with Matlab toolbox to analyze the characteristic wavelength. According to the evaluation parameters, the lyco-pene’s model which based on GA-PLS was the optimal, the number of variables was 142, the value of R was 0.9072, and its model’s RMSECV was 8.76, the RMSEP was 8.93. The Vita-min C’s model based on UVE-PLS was the optimal, the chose variables was 493, the value of R was 0.9043, the RMSECV was 0.4856, and the RMSEP was 0.4872; Based on GA-PLS, the best model of the total free amino acid used variable 137, its R was 0.9163, the RMSECV was 11.79, the RMSEP was 11.95. Based on UVE-PLS, the optimal model of soluble solids, using variable 336, its R was 0.9122, the RMSECV was 0.4257, and the RMSEP was 0.4282.(4) Wavelet and wavelet packet de-noising research. In the study of dbN wavelet, the db9 was the best, and the best decomposition layer number was 2. By hard threshold Stein unbiased likelihood estimation or heuristic de-noising method, and according to the different layers of the noise estimation to adjust the threshold could get the best de-noising effect, the SNR and RMSE values were 66.7466 and 0.0001 respectively. In the research of coifN wave-let, the best wavelet base was coif5, the best decomposition layer was one. By selecting dif-ferent threshold method and different threshold adjustment way, got the different values of SNR and RMSE by adjusting threshold valve in coif5 wavelet. The results showed that the estimation of the noise floor was conducted by using the coefficient of the first layer to adjust the threshold and other different layers, both de-noising effect consistent. General soft threshold method which wasn’t adjustment was the worst effect, and the hard threshold which could be adjusted, using stein unbiased likelihood estimation and heuristic threshold method, could get the best SNR and RMSE value,64.4761 and 0.0002 respectively. In the study of symN wavelet, the sym5 wavelet was the best, and the best decomposition layer was 1. Es-timation of noise layer by using the coefficient of the first layer to adjust the threshold and estimation to adjust the threshold by the noise of different layers, the de-noising effect was the same. Universal soft threshold method which was not adjusted effect was the worst, the threshold could be adjusted by the hard way, stein unbiased risk estimate and heu-ristic threshold method, could get the best SNR and RMSE values, which were 67.1901 and 0.0001 respectively.Separately on the db9, coif5 and sym5 wavelet decomposed by wavelet packet, in the selection of "Shannon" type of entropy, the signals were decomposed and reconstructed. By the SNR, the optimal wavelet packet determined was sym5.