Monitoring and classification of citrus Huanglongbing based on UAV hyperspectral remote sensing

Abstract: Citrus Huanglongbing (HLB) is an extremely destructive disease without cured medicine. Finding citrus trees suffering from HLB as soon as possible and eradicating the infected tree timely is an effective measure to prevent citrus from HLB. For large-scale citrus orchards, monitoring HLB is a heavy workload that requires a lot of time and effort.- Using remote sensing by unmanned aerial vehicle (UAV) to monitor the citrus orchards is a feasible measure which could reduce a lot of work and cost. In this study, a hyperspectral image of citrus in orchard was obtained by a UAV equipped with hyperspectral camera, flying at a height of 60 m. 26 healthy trees and 26 trees infected HLB were selected from the hyperspectral image, which was radiational corrected and geometric corrected. 10 regions of interest (ROIs) were created (5×5 pixel size) on each selected citrus canopy and the mean reflectance spectra in every ROI was calculated. The abnormal spectra were removed by observing the mean reflectance spectra, and the remaining spectra were smoothed and denoised by Savitzky-Golay. The ground reflectance spectra captured by ASD FieldSpec HandHeld 2 Spectroradiometer was used as a reference to verify the effect of the spectra by hyperspectral camera in UAV and it was found that the reflectance spectra of hyperspectral camera had a same trend with the spectra from gound. The first-orderderivative reflectance spectra (FDR) and the inverse logarithmtic reflectance spectra (ILR) were obtained by spectral transformation. The dataset was divided into a training set and a test set by a ratio of 3:1, and the training set was used to train the discriminant model. In the training phase of the model, K-folds cross validation was used internally. Finally, the test data was used to predict in the discriminant model. The k-Nearest Neighbor (kNN) and support vector machine (SVM) model were adopted as classifiers respectively, and 3 kinds of spectra transformed from the full-band spectra and first 3 principal components after PCA were compared as input variables to establish the discriminant model. Different input variables in different classifiers, different kernels in SVM model and different distance calculating way in the kNN were compared. The parameters of different models were gradually tried, and the parameters with the highest training accuracy were selected for modeling. Some conclusions were gotten in the paper. First, the reflectance spectra acquired by remote sensing by UAV could be used to establish the discriminant model for trees injected HLB after a series of processing. Such as the SVM classification model with the quadratic kernel had a classification accuracy of 94.7% for full-band FDR and the predictive error rate for the test data was 3.36%. Second, spectral variable obtained by spectral transformation can improve the classification accuracy of the model. For example, the classification accuracy with FDR as the input variables was the highest in each model. Third, principal component analysis (PCA) dimensionality reduction on spectral variables can significantly improve the recognition speed. We could find that the error rate had decreased for model with ILR after PCA and increased for models with other spectra after PCA. Last but not least, in the SVM classification model with the quadratic kernel, the discriminative accuracy for healthy plants was 100%, and the plants with misjudged just were plants injected HLB, the impact of this result was likely to come from part with HLB in the canopy of fruit trees. In summary, hyperspectral remote sensing by UAV was used to monitor the cultivation of orchards in large areas. It was an effective management method to monitor citrus HLB by establishing a discriminant model.