Skip to main content
ARS Home » Research » Publications at this Location » Publication #288339

Title: Band selection using forward feature selection algorithm for citrus Huanglongbing disease detection

item KATTI, ANURAG - University Of Florida
item LEE, WON SUK - University Of Florida
item EHSANI, REZA - University Of Florida
item Yang, Chenghai
item Mangan, Robert

Submitted to: Biosystems Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/21/2015
Publication Date: 11/5/2015
Citation: Katti, A.R., Lee, W., Ehsani, R., Yang, C., Mangan, R.L. 2015. Band selection using forward feature selection algorithm for citrus Huanglongbing disease detection. Biosystems Engineering. 40(4):417-427.

Interpretive Summary: Citrus greening, also known as Huanglongbing (HLB), is a destructive citrus disease that has caused substantial economic losses to the citrus industry. Timely and accurate detection of HLB-infected citrus trees is important for the management and control of the disease. This study compared image classification results using a few selected image bands versus using all the available bands. It also compared the performance of simple and commonly used classification techniques with more complex methods with selected bands. Results showed that a few well-chosen bands yielded better results than when all bands were used, but the overall performance of all the algorithms was quite similar. Therefore, band selection techniques need to be combined with appropriate classification algorithms for effective identification of HLB infection.

Technical Abstract: This study attempted to classify spectrally similar data – obtained from aerial images of healthy citrus plants and the citrus greening disease (Huanglongbing) infected plants - using small differences without un-mixing the endmember components and therefore without the need for endmember library. However, the large number of hyperspectral bands has high redundancy which has to be reduced through band selection. The objective, therefore, was to compare the classification performance when using a few bands chosen using the Forward Feature Selection Algorithm against classification using all the available bands; and comparing the performance of simple, traditionally used classification techniques (naive Bayes classifier and K-nearest neighbor) with more complex methods (support vector machines, generalized local discriminant bases and multi-modal Bayesian classifier) with band selection. It was observed that a few well-chosen bands yielded much better results than the case when all bands were chosen. But the overall performance of all the algorithms was quite similar with small differences in accuracies of individual classes. Hence, the choice of classification algorithms must be done on a case-by-case basis. But it was observed that simple techniques can work as well as the more computationally intensive ones.