Hyperspectral imaging and spectrometry-derived spectral features for bitter pit detection in storage apples

Abstract

Bitter pit is one of the most important disorders in apples. Some of the fresh market apple varieties are susceptible to bitter pit disorder. In this study, visible-near-infrared spectrometry-based reflectance spectral data (350-2500 nm) were acquired from 2014, 2015 and 2016 harvest produce after 63 days of storage at 5 ◦C. Selected spectral features from 2014 season were used to classify the healthy and bitter pit samples from three years. In addition, these spectral features were also validated using hyperspectral imagery data collected on 2016 harvest produce after storage in a commercial storage facility for 5 months. The hyperspectral images were captured from either sides of apples in the range of 550-1700 nm. These images were analyzed to extract additional set of spectral features that were effective in bitter pit detection. Based on these features, an automated spatial data analysis algorithm was developed to detect bitter pit points. The pit area was extracted, and logistic regression was used to define the categorizing threshold. This method was able to classify the healthy and bitter pit apples with an accuracy of 85%. Finally, hyperspectral imagery derived spectral features were re-evaluated on the visible-near-infrared reflectance data acquired with spectrometer. The pertinent partial least square regression classification accuracies were in the range of 90-100%. Overall, the study identified salient spectral features based on both hyperspectral spectrometry and imaging techniques that can be used to develop a sensing solution to sort the fruit on the packaging lines.