Robustness of near infrared spectroscopy based spectral features for non-destructive bitter pit detection in honeycrisp apples

• Non-destructive apple bitter pit detection through optical sensing was evaluated.
• Robust feature selection and validation algorithms were implemented.
• Spectral bands that may aid in apple bitter pit detection were identified.
• Classifier results validated usefulness of selected bands.

Bitter pit is a serious disorder in apples. The current technique involves manual inspection of fruits prior to packaging for fresh market. Therefore, the main objective of this study was to evaluate the near infrared (NIR) spectroscopy for bitter pit detection in apples. The spectral reflectance data were collected from healthy and bitter pitted honeycrisp apples from two different locations. Apples were stored in cold storage and spectra were acquired at 0, 35 and 63 days after harvest (DAH). On each of the DAH, each of the 40 apples (20 healthy and 20 bitter pitted) were analyzed to acquire three spectra per location with three marked locations per fruit. Suitable spectral features were selected using stepwise multilinear regression and rank feature technique. The spectral bands of 971.2, 978.0, 986.1, 987.3, 995.4, 1131.5, 1135.3, 1139.1 and 1142.8 nm were identified as the bands thought to be associated with bitter pit in honeycrisp apples. Feature datasets were evaluated using quadratic discriminant analysis and support vector machine classifiers to evaluate robustness of these features in bitter pit detection. Overall, classifiers performance comparison revealed that bitter pitted honeycrisp apples can be distinguished with average accuracy in the range of 78–87%. Based on spectral features of this study, spectra related to cell membrane water-soaked regions that contribute to spectral variation might have been identified. Our on-going studies are further validating those bands on Honeycrisp and other apple cultivars and using different spectral band selection methods towards developing a portable sensing module for apple bitter pit detection.