Detection Of Soluble Solid Content For Apples Based On NIR Spectroscopy And Model Transfer Method

China is the country with the largest apple planting acreage and production in the world.Due to the poor quality of apples,the apple export rate is low.Soluble solids content(SSC)is an important indicator to evaluate the internal quality of apples.Traditional SSC detection methods are destructive and time-consuming.Therefore,it is very important to explore a rapid detection method of the internal quality of apples to promote the development of China's apple industry.A large number of studies at home and abroad have shown that near-infrared spectroscopy can achieve non-destructive detection of the internal quality of fruits.However,due to the characteristics of its point-like spectral measurement,its application in factory online inspection is greatly limited.Hyperspectral imaging technology,which has the advantage of combining spectra and image data,is currently a popular technology for fruit quality detection.However,because of the instrument is expensive,most studies only use spectra in visible near-infrared region(VIS-NIR,400-1000 nm),and the near-infrared region(NIR,1000-1700 / 2500 nm)was reported by few studies.In addition,whether the spotted near-infrared spectrometer or array hyperspectral imager,the spectra measured by the two types of sensors are different.Constructing a fruit quality detection model which suitable for multiple types of sensors is a new research direction currently.Focusing on these issues,in this paper apple was chosen as the research object,and hyperspectral imager and Fourier transform infrared spectrometer were used to explore the rapid detection method of apple internal quality and establish a universal model.The main research contents are as follows:(1)Based on near-infrared hyperspectral(NIR-HSI)sensor,through spectral analysis and preprocessing,combined with characteristic wavelength selection algorithm,the effective wavelength used to measure apple SSC was extracted.Competitive adaptive weighted sampling(CARS),continuous projection algorithm(SPA),random leapfrog(RF),and combined CARS-SPA and CARS-RF algorithms were used to extract the effective wavelengths of apple hyperspectral images.Based on the selected effective wavelength,partial least squares(PLS)and least squares support vector machine regression(LS-SVR)were applied to establish models to predict SSC in apple and compare the performance of different models.The research results showed that among all models,the characteristic wavelength model selected by CARS-SPA is the best,and both PLS and LS-SVR could be used to develop calibration models to predict the SSC of apple.(2)Using Fourier transform spectrometer and the same batch of apples,non-destructive prediction of the SSC was implemented.Through spectral analysis and preprocessing,combined with CARS,SPA,RF and CARS-SPA combination algorithms,PLS detection models based on characteristic wavelengths and full-spectrum were established,and compared with the results of model built on hyperspectral imager.The results showed that the SPA has the best effect on FT-NIR spectral characteristic wavelength extraction,which was similar to the hyperspectral prediction results but was different from the characteristic wavelengths selected by the hyperspectral imager.(3)Based on the two different NIR spectrometers,studied the universality of the spectral model among different instruments.Because the characteristic wavelengths of the two sensors were inconsistent,and the preprocessing method could not eliminate the difference between them,the feasibility of applying the NIR-HSI in the FT-NIR model was explored based on the original spectrum.The Fourier transform spectra has 3112 bands and high resolution 0.2 nm,when the hyperspectral has 256 bands with low resolution 8 nm.On the one hand,this paper discussed the effect of model transfer at low resolution by reducing the FT-NIR spectral resolution;on the other hand,the effect of model transfer was focused at high resolution by increasing the HSI spectral resolution.Using direct standardization(DS)and piecewise direct standardization(PDS)to reduce the spectral difference between master(Fourier transform spectrometer)and slave(hyperspectral imager)spectrometers,the results showed that at low resolution,the conversion spectrum established by DS obtains a prediction effect similar to the original model.In summary,near-infrared spectroscopy can achieve non-destructive detection of SSC in apple,and the model transfer algorithm improves the universality of the near-infrared spectroscopy model.These results provide important technical support for achieving online detection of internal quality and improving the apple's market competition.