NIR SPECTROSCOPY IS SUITABLE TO DETECT INSECT INFESTED CHESTNUTS

In this study, the feasibility of using NIR spectroscopy to detect hidden insect damage is demonstrated. Using a genetic algorithm for feature selection (from 2 to 6 wavelengths) in combination with a linear discriminant analysis routine, classification error rates as low as 16.81% false negative, 0.00% false positive, and 8.41% total error were achieved, with an AUC value of 0.952 and an Wilk’s λ of 0.403 (P<0.001). A Savitzky-Golay first derivative spectral pretreatment with 13 smoothing points was used. The optimal features corresponded to Abs[1582 nm], Abs[1900 nm], and Abs[1964 nm]. These results represent an average of 55.3% improvement over a traditional floatation sorting system.