Non-destructive visible and short-wave near-infrared spectroscopic data estimation of various physicochemical properties of Fuji apple (Malus pumila) fruits at different maturation stages

measurement of physicochemical properties of fruits during maturation stages can help having proper fruit management. Spectroscopy data analyzing and processing is among the commonly used methods that enable non-destructive accurate property estimation. Non-destructive linear (partial least squares regression, PSLR) and non-linear (artificial neural network, ANN) regression estimation of different physicochemical properties including firmness, acidity (pH) and starch content of 160 Fuji (Malus pumila) apple fruit samples at various maturity stages using visible and short wave near infrared (VSWIR) spectroscopic data in wavelength range 400–1000 nm is investigated with the following steps: (1) harvesting 160 Fuji apple samples at four different maturation levels; (2) extracting spectral data in wavelength range of 400–1000 nm; extracting physicochemical properties of tissue firmness, acidity (pH) and starch content; (3) pre-processing the reflectance spectra from each sample; (4) selecting effective wavelength values for each chemical property; and (5) non-destructive estimation of tissue firmness, acidity (pH) and starch content using spectral data range 400–1000 nm and spectral data based on effective wavelengths, by means of an ensemble average artificial neural network method. Results show that the neural ensemble reached similar results when using VSWIR spectral data content (wavelength range) and fixed effective selected NIR wavelengths. Correlation coefficients estimating tissue firmness, acidity (pH), and starch content were 0.800, 0.919, and 0.940 for VSWIR spectral data (linear PLS regression), 0.826, 0.947, and 0.969 for VSWIR spectral data (non-linear ANN), 0.827, 0.946, and 0.969 for fixed NIR effective wavelengths (non-linear ANN). Mean ± std. Regression coefficients for tissue firmness, acidity (pH), and starch content were 0.717 ± 0.113, 0.786 ± 0.131, and 0.941 ± 0.013 for Vis/NIR spectral data (linear PLS regression), 0.849 ± 0.017, 0.930 ± 0.017, and 0.967 ± 0.007 for Vis/NIR spectral data (non-linear ANN), 0.852 ± 0.016, 0.929 ± 0.015, and 0.966 ± 0.006 for fixed effective NIR wavelengths (non-linear ANN).