Supervised Multivariate Analysis of Hyper-spectral NIR Images to Evaluate the Starch Index of Apples

Abstract

Fruit maturity indexes are crucial in harvest time determination and commercial context. The harvest time of apples, matching the desired commercial characteristics, is assessed through starch–iodine test in practice. Fruit halves are dipped into iodine solution and patterns are visually evaluated by experts comparing them to reference charts. Aim of the work was to study the relationships of near infrared (NIR) spectral images (1,000–1,700 nm), starch/starch-free patterns visually assessed and RGB color images. Spectral images of 88 Golden Delicious Klon B apples were sampled at seven different maturity stages. Partial least-squares discriminant analysis (PLSDA) technique was used on hyper-spectral NIR images to classify single pixels using its NIR reflectance spectrum. The response variable (i.e. the classification for each pixel) was identified through the matching of single pixel obtained with the color images, segmented in two classes (starch and starch-free), and the NIR hyper-spectral matrix. Mean hyper-spectral classification obtained through PLSDA modeling on individual apple correctly classified 80.81% of the total pixels, while the unique model, i.e. a single model including all the fruits, resulted in 66.33%. In the latter case, the relationship with the RGB classification showed high values (Pearson correlation coefficient r = 0.95). The present work shows the feasibility of NIR imaging spectroscopy as a tool for apple fruit maturity determination, avoiding expert’s subjective interpretation by traditional starch index assignments.