SPECTRAL BAND SELECTION FOR OPTICAL SORTING OF PISTACHIO NUT DEFECTS

Abstract

A technique using near-infrared spectroscopy (NIR) was developed for selecting the optimal spectral bands for use in dual-wavelength sorting machines commonly found in food processing plants. A variation of a nearest-neighbor classification scheme selected the two optimal spectral bands given NIR spectra from both sides of an object. The optimal bands were determined for two cases: when both sides contain the defect of interest (AND logic), or when the defect appears on a single side (OR logic). A commercially available sorting machine was used to compare the sorting accuracy using the spectral bands determined with this technique to the accuracy using bands recommended by the manufacturer. The product stream tested was the removal of “small inshell” (small nuts with the shell intact) and shell halves from the stream of nuts with no shells (“kernels”). Results for the selected spectral bands averaged 1.20% false negative (fn) for small inshell and 1.80% fn for half shells with 0.15% false positive (fp) vs. 1.70%, 2.40%, and 0.70%, respectively, using the spectral bands recommended by the manufacturer. Optimal spectral bands were also determined and reported for a variety of other defects and unwanted materials commonly sorted in the pistachio processing plant, including adhering hull, stained, sticks, mold, insect damage and/or webbing, and black spots. Given the success of this technique in pistachio sorting experiments, it is believed that it could be applied to any commodity sorted using commercially available, dual-wavelength, NIR sorting devices.