Comparison of optimal wavelengths selection methods for visible/near-infrared prediction of apple firmness and soluble solids content

Abstract

Abstract. Visible and near-infrared (Vis-NIR) spectroscopy is a promising technique for noninvasive measurement of quality attributes of agricultural products. The technique relies on selection or extraction of optimal spectral features or wavelengths for the development of calibration models. Five wavelengths selection algorithms, namely, uninformative variable elimination (UVE), partial least squares projection analysis (PLSPA), standard genetic algorithm (SGA), successive projections algorithm (SPA), and affinity propagation (AP), were investigated for extracting optimal wavelengths from the spectra of 460 - 1,100 nm to evaluate their ability for prediction of firmness and soluble solids content (SSC) in apples using partial least squares (PLS) method. More than 6,500 apples of ‘Delicious’, ‘Golden Delicious’ and ‘Jonagold’ varieties harvested in 2009 and 2010 were used for analysis. Overall, the prediction results from each wavelength selection algorithm were not as good as those obtained by full-spectrum PLS models. A simple fusion method, which averaged over the prediction results from the five wavelengths selection algorithms, improved prediction results for firmness and SSC by 0.4%-4.8% and 0.4-5.6%, respectively, compared with the full-spectrum PLS models for the three varieties of apples. This fusion method provides a simple and robust means for improving firmness and SSC prediction results.