Rapid Detection Citrus HLB By Developing A Handheld Device Based On Spectral Imaging Technology

Citrus Huanglongbing(HLB)is one of the most devastating diseases in citrus production.HLB infected trees not only cause low yield and short life span,but also act as a source of infection in orchards.Currently,timely removal of HLB infected trees is considered as one of the most effective strategies for citrus orchard management.Therefore,it is essential to remove the infected trees as early as possible.Spectral imaging techniques are promising for rapid and non-destructive detection of plant diseases.The aim of this study is to develop a handheld device based on the spectral images acquisition,and build a rapid detection model(Mobile Net V3)for citrus HLB detection.The contents and conclusions of this study are listed as follows:At first,a handheld device was developed to collect dynamic chlorophyll fluorescence,multicolor fluorescence and multispectral reflectance images on citrus leaf synchronously.Compared to the laboratory instrument,(V10E,Isuzu Optics Corp,Taiwan,China)it shows that the data collected by the two instruments have consistent trends.Secondly,we explored the effect of citrus HLB on the chlorophyll fluorescence characteristics and changes of visible and near infrared spectral reflectance.Due to the infection of pathogenic bacteria,the large amount of chlorophyll on the leaves was destroyed.Compared with healthy leaves,F440 and F520 intensities increased with a significant difference of1.05,1.14,1.41 and 1.83 times for HLB asymptomatic and symptomatic in Navel orange leaves,respectively.Similar results can be also found in Ponkan.Compared with HLB-symptomatic leaves,the reflectance of HLB-asymptomatic ones in the visible(460,520 and 680 nm)region was slightly higher than that of healthy ones but lower than nutrient deficient ones in both Navel orange and Ponkan.In the NIR,the reflectance of Mgdeficient Navel orange leaves was higher than healthy,while N-deficient Ponkan leaves was lower than healthy leaves.Additionally,the overall detecting accuracies of Mobile Net V3 models trained by dynamic chlorophyll fluorescence,multicolor fluorescence and multispectral reflectance images were 73.78%,85.06% and 91.46%,respectively.Finally,dynamic chlorophyll fluorescence,multispectral fluorescence and multispectral reflectance images are integrated to obtain physiological and structural information related to citrus HLB from different perspectives,and to establish a rapid detection model for multi-source image fusion.The results show that the lightweight convolutional neural network model built with multispectral fluorescence and multispectral reflection images has the best discrimination,with the overall identification accuracy of 92.07% and the false negative rate of HLB reduced to 12.19%.Although the overall accuracy was slightly higher than that of multispectral fluorescence images or multispectral reflection images alone,the false negative rate decreased by 11.59%.However,the identification effect decreased substantially when detecting other orchards used established models.Therefore,using the transfer learning algorithm to solve this problem.The transfer learning method of fine-tuning model obtained a superior transferring ability than that of reuse-model with the overall accuracy of 96.50% after 46 Ponkan samples being added and the false negative rate decreased to 4.70%.The findings in this research demonstrated that the feasibility of combining multicolor fluorescence imaging with multispectral reflectance imaging for rapid citrus Huanglongbing detection based on lightweight convolutional neural network using a handheld device.