Detection of pits and pit fragments in fresh cherries using near infrared spectroscopy

Abstract

Near infrared spectroscopy in the wavelength region from 800 to 2600 nm was evaluated as the basis for a rapid, nondestructive method for the detection of pits and pit fragments in fresh cherries. Partial least squares discriminant analysis following various spectral pretreatments was applied to spectra of cherries with either no pit, a whole pit, a half pit, or a quarter pit to test various classification schemes. An iterative algorithm tested all combinations of pretreatments and parameters as input to the partial least squares discriminant analysis. In addition, a step forward feature selection algorithm was used to identify the most significant wavebands in order to isolate small sets (<10) of spectral bands that represent the entire spectra. The highest accuracy was achieved for a binary model in which the samples were combined into only two classes (no pit versus whole pit + half pit + quarter pit) using all features (reflection at each wavelength) with no false positive error, 4% false negative error, and 98% overall accuracy. Overall accuracy of the same model was reduced only slightly to 96% when employing only the four most significant features. Accuracy declined when models attempted to separate the classes of fragments, with the lowest being 92, 83, 86, and 99% accuracy, respectively, in discriminating no pit, quarter pit, half pit, and whole pit classes separately. The high accuracy achieved under the binary model using only four features indicates that reflection of light at specific near infrared wavelengths is a suitable basis for high-speed, nondestructive detection of pits and pit fragments in cherries.