Non-invasive assessment of fruit quality by near-infrared spectroscopy for fruit grading in an in-line setting

Instrument criteria, in terms of wavelength range, wavelength resolution, signal to noise ratio, sensitivity and illumination/detector were defined for the use of near infrared (NIR) spectroscopy in the assessment of soluble solids content (SSC) in intact fruit in an in-line system. Techniques for predictive model generation and transfer of predictive models across a number of systems were assessed in tenns of root mean squared error of prediction (RMSEP). A comparative study of components making up an NIR spectroscopic system established that for the application of assessment of SSC in intact fruit, quartz halogen light sources could provide adequate low cost radiant energy and prisms could provide cheap, more efficient dispersion and higher throughput than flat diffraction gratings. Three wavelength dispersion elements (single equilateral prism, two prisms in series and a ruled diffraction grating) were separately assessed. Calibration perfonnance for sucrose in a water-cellulose matrix was significantly degraded by a signal to noise ratio (SNR) <5000: 1, and when wavelengthresolution was decreased beyond a FWHM of 16 nm (at 912 nm). Therefore either photo- diode arrays or binned charge-coupled devices could be used as photodetecting elements if SNR is maintained above this level. A body-transmittance optical path was preferred over reflectance optics to eliminate the 'noise' from specularly reflected light. However, physicaly contacting the fruit with an optical barrier to separate illuminated and detected regions constrained process rates. Therefore, an illumination/detector configuration was designed to allow rapid, non-contact spectral measurements to be made. This configuration supported comparable calibration statistics for assessment of SSC of intact melons as a 'contact' configuration (e.g. root mean squared error of cross validation (RMSECV) of 0.740 and 0.650 Brix non-contact and contact, respectively). Predictive models developed using partial least squares (PLS) regression were significantly more accurate than those developed using multiple linear regression, principal component regression or parallel regression. Wavelength selection techniques were examined. Predictive PLS models based on knowledge of spectrally important wavelengths (for SSC, 630 to 1040 nm) were superior to models based on other wavelength selection techniques. Assessment of data pre-treatment techniques showed that, in most cases, mean centred and autoscaled absorbance data provided the best results. Nine methods for transfer of calibration between instruments were compared against the performance of a simple model updating (MU) technique. While MU gave consistently better predictions on slave instruments, this approach requires maintenance of calibrations on every instrument. Of the established standardisation methods, direct standardisation of the wavelet coefficients was the most efficient. These design criteria were used in the construction of a prototype fruit sorting system, with performance assessed over a period of two years. The hardware components of this system proved adequately robust to endure the rigours of a pack-house environment and the accuracy of the sorting achieved an RMSEP of 11 0.7° Brix (standard deviation and range of sse in sample set, 1.5° and 8.5°, respectively). 111