Airborne hyperspectral imaging based citrus greening disease detection using different dimension reduction methods

Abstract

<abstract> <bold>Abstract.</bold> Hyperspectral (HS) imaging has become an efficient tool for agriculture applications. While HS image can provide large amounts of information, it has high redundancy because the bands of HS image are highly correlated. To solve citrus greening disease (Huanglongbing, or HLB) detection problem using airborne HS image, an efficient dimension reduction method needs to be applied, to improve the accuracy of HLB detection and simplify the image acquisition step. There are many dimension reduction methods have been proposed. The objective of this study was to compare the classification performance when using different bands selected using different dimension reduction methods, utilizing the HS airborne images of citrus groves in Florida in 2011. In this study four dimension reduction methods including principal component analysis (PCA), maximum noise fraction (MNF) transformation, forward feature selection algorithm (FFSA) and Kullback-Leibler divergence (KLD) based method, were applied on the obtained HS image. While PCA and MNF are feature extraction methods, which means a new and reduced dataset representing the transformed initial information will be obtained by using them, FFSA and KLD based methods are feature selection, or band selection methods, which means a subset of bands from the original information will be selected. To analyze the band reduction results, both pixel based classification and tree based classification were applied on the 2011 HS data, and the results obtained from these four methods were compared. After applying K-nearest neighbor (KNN) classification on the pixel bands chosen by these four methods, they all showed a 100% accuracy in a calibration dataset for both HLB detection and healthy sample detection. KLD showed the highest HLB detection accuracy of 63.3%, while PCA and KLD based method showed the highest healthy detection accuracy of 74% using a validation dataset. After applying Mahalanobis distance (MahaDist) classification on the transformed image data, MNF and KLD gave the same tree based HLB detection accuracy, which was 93.3%.