ASABE Technical Library - Abstract

Abstract. Citrus black spot (CBS), one of the most common fungal diseases of citrus, affects lesions on the rind and early fruit drop before its mature stage. This disease can significantly reduce crop yield, making the highly blemished fruit unsuitable for market, and influence the fresh fruit market eventually. Identification of this devastating disease before harvesting or shipment to other district can avoid healthy fruit infected by diseased fruit or spread this disease to new areas. And CBS identification can also assure fruit quality and safety and enhance the competitiveness and profitability of the citrus industry. To this end, the objective of this research was to develop an efficient algorithm using selected wavelengths based on hyperspectral image data, and to detect citrus black spot ahead of harvesting. A hyperspectral imaging system was used to acquire hyperspectral images of citrus samples in the wavelengths range between 396 and 1010 nm. However, hyperspectral image contains hundreds of wavelengths, and many of them would be considered as redundant information, which may even decrease the classification accuracy. Band selection was used to reduce the dimensionality of hyperspectral images and select the useful bands for further application. In this study, a principal components analysis (PCA) and four band ranking methods, T-test, Kullback-Leibler distance, Chernoff bound and Receiver Operating Characteristic (ROC) were applied and 533 nm was selected to classify healthy and CBS infected. Kernel SVM, KNN and C4.5 classification methods were used to evaluate the performance of the selected band, and Kernel SVM achieved the highest accuracy of above 98%. Then Sequential Forward Selection (SFS) and Greedy Stepwise methods were carried out to select bands to distinguishing five CBS symptoms, and they selected 14 and four bands, respectively. After selecting optimal bands using each method, C4.5 was used to test the performance of distinguishing CBS infected and healthy based on selected bands. The overall classification accuracy was about 61%.