Multispectral imaging for predicting sugar content of 'Fuji' apples
This research investigated a usage of multispectral imaging to predict sugar content of ‘Fuji’ apples. A visible/near-infrared spectroscopy (350–1200 nm) was used to select optimal wavelengths for the multispectral imaging system. The spectral data were analyzed using the backward interval partial least square to generate a subset composed of several most sensitive wavebands. Four optimal wavelengths (461 nm, 469 nm, 947 nm and 1049 nm) were determined from this subset using stepwise multiple linear regression. A multispectral imaging system was developed based on these effective wavelengths. The scattering areas of the multispectral images were extracted by using the image histogram and the camera response function. The scattering profiles were calculated from the scattering areas by radial averaging. The modified Lorentzian distribution function was used to fit the scattering profiles. The parameters of the Lorentzian functions were used as the data base of multiple linear regression to create the prediction model. The multiple linear regression model predicted sugar content with r = 0.8861 and RMSE (root-mean-square-error of calibration) = 0.8738° Brix.