Non-destructive Detection Of Quality In Lingwu Jujube Based On Hyperspectral Imaging Technology

The application of hyperspectral imaging technology in fruit quality nondestructive detection becomes an important trend. In this paper, hyperspectral imaging technique was investigated for non-destructive determination of sugar and moisture content, different maturity stages and origins in ’Lingwu’ long jujube, combined with chemometric methods. The detection algorithms offered the comprehensive quality evaluation of fresh long jujube to provide the technical support for high-end fruit market position of ’Lingwu’ long jujube. The main research results are as follows:(1) Compared to PCR models, PLSR models have an excellent ability to predict the sugar and moisture content in ’Lingwu’ long jujube. The correlation coefficient of calibration and validation models about sugar content prediction models based on 400～1000 nm spectral with multiplicative scatter correction and 900～1700 nm spectral with Savitzky-Golay smoothing are 0.938,0.823 and 0.916,0.864, respectively. The correlation coefficient of calibration and validation models about moisture content prediction models based on 400～1000 nm spectral with Savitzky-Golay smoothing and 900～1700 nm spectral with multiplicative scatter correction are 0.913,0.874 and 0.913,0.904, respectively.(2) The linear discriminant analysis and supporting vector machine models based on 400～1000 nm spectrum pretreated with Savitzky-Golay have a better performance. Optimal wavelengths were selected by PCA. Fourteen characteristic wavelengths (415.75nm、478.17nm、521.38nm、535.79nm、 641.42nm、670.23nm、675.03nm、699.04nm、703.84nm、742.25nm、747.06mm、963.12nm、967.93nm、 948.72nm) were selected to build the linear discriminant analysis and supporting vector machine models. The discrimination accuracy of identification models are 90% and 92.14%, respectively.(3) The linear discriminant analysis based on original spectrum and supporting vector machine models based on the spectrum pretreated with standard normalized variable have a better performance. Optimal wavelengths were selected by PCA. Identification models were developed based on linear discriminant analysis and supporting vector machine using these optimal wavelengths. The accuracy of identification models are 92.5%.