Classification of blueberry fruit and leaves based on spectral signatures

Blueberry spectral analysis can provide necessary wavelengths, for use in multispectral imaging that could be applied in blueberry yield estimation system. Samples of fruit and leaves were obtained from a commercial blueberry field in Waldo, Florida and an experimental field in Citra, Florida, USA in 2011. Samples were also collected in 2010 in Waldo. Seven representative southern highbush varieties were chosen for the experiment. Spectral reflectance was measured in the 200–2500 nm with an increment of 1 nm. Samples were divided into leaf, mature fruit, near-mature fruit, near-young fruit and young fruit. Normalised indices were used as the candidate variables for classification. Each index was composed of the two wavelengths that had the greatest difference in reflectance between two classes. Classification tree, principal component analysis (PCA) and multinomial logistic regression (MNR) were conducted to develop classification models. An MNR model with six wavelengths (233, 551, 554, 691, 699 and 1373 nm) performed the best for the 2011 dataset, with a prediction accuracy of 100% for leaf and mature fruit, 97.8% for young fruit, 97.9% for near-young fruit and 94.6% for near-mature fruit. Four wavelengths (553, 688, 698 and 1373 nm) were used in the classification models of two yearsʹ data with four classes (mature fruit, intermediate fruit, young fruit and leaf), and accuracies of 100%, 100%, 99%, and 98.5% were obtained for the classification of leaf, mature fruit, intermediate fruit and young fruit, respectively. An easy-to-use and low cost blueberry fruit detector could thus be developed using multispectral imaging.