Abstract:Citrus black spot (CBS), which is one of the most common fungal diseases of citrus, causes lesions on the rind and early fruit drop before its mature stage. This disease can significantly reduce crop yield, making blemished fruit unsuitable for market. The objective of this research was to study the reflectance spectral characteristics of healthy and infected citrus fruits to identify diseased fruit from healthy ones. A portable USB2000+spectrometer was used to acquire spectral reflectance of citrus fruit in the laboratory with wavelength ranged from 340nm to 1030nm. However, the spectra contained thousands of wavelengths, and many of them would be considered as redundant, which may even decrease the classification accuracy. To reduce the data dimensionality and select the useful bands for further application, principal components analysis (PCA) and four band ranking methods, i.e., T-test, Kullback—Leibler distance, Chernoff bound and receiver operating characteristic (ROC) were applied. One important wavelength (525nm) was selected and used to classify healthy and CBS infected fruits. Sequential minimal optimization (SMO), radical basis function network (RBF), and C4.5 classification methods were used to evaluate the performance of the selected band, and SMO achieved the highest accuracy of 99.37%. In order to compare the performance of classification accuracies according to optimal wavelengths selected by using different methods, two other methods, i.e., sequential floating forward selection (SFFS) and mutual information (MI), were applied. Wavelengths of 527nm and 917nm were selected based on SFFS, while the MI method selected 513~531nm as the optimal wavelength range, and the highest recognition accuracy was 99.06%, which was lower than that of using 525nm. Then SFFS was applied to find the optimal wavelengths for further distinguishing three CBS symptoms. C4.5 method was used to evaluate the performance of distinguishing CBS infected and healthy fruits based on selected wavelengths, and the highest overall classification accuracy was 73.77%.
