Spectrum classification of citrus tissues infected by fungi and multispectral image identification of early rotten oranges

Abstract

Citrus fruit is susceptible to postharvest rot by fungal infection. The detection of early rot is difficult due to similar skin characteristics to sound area, which limits the ability of the grading system to evaluate the comprehensive quality of citrus. In this study, the visible and near infrared hyperspectral imaging system with the wavelength range of 325–1000 nm was used to collect hyperspectral images of oranges. Hyperspectral data of three types of tissues including sound tissue from 80 samples, rotten tissue infected by Penicillium digitatum from 100 samples and rotten tissue infected by Penicillium italicum from 100 samples were extracted. The bootstrapping soft shrinkage (BOSS) and BOSS-SPA (BOSS-Successive Projections Algorithm) combination algorithm were separately used to optimize spectrum variables. The partial least squares discriminant analysis (PLS-DA) model for classifying three types of tissues and PLS-DA model for classifying two types of tissues (sound tissue and rotten tissue) were constructed based on full-spectrum and the selected informative variables. Model comparisonshowed that the BOSS-PLS-DA model can effectively identify three types of tissues with the classification accuracy of 97.1%, while the BOSS-SPA-PLS-DA model was more effective for the binary classification of sound and rotten citrus tissues with the accuracy of 100%. Furthermore, the wavelength images corresponding to the nine informative variables extracted by BOSS-SPA were performed the principal component analysis (PCA), and four feature wavelength images (508, 568, 578 and 614 nm) were obtained by analyzing the weighting coefficients of each single-wavelength images constituting the optimal principal component (PC) image. Finally, a fast multispectral image processing algorithm combined with the global threshold theory was proposed for the rotten orange detection based on the extracted four wavelength images. A total of 280 samples including 80 sound and 200 rotten samples were used to evaluate the classification ability, which showed the proposed multispectral image detection algorithm can successfully differentiate between sound and rotten oranges with an overall classification accuracy of 98.6%.