Plant Leaves Diseases Dection Using Spectral Imaging Technology

When a plant is under the influence of organisms or abiotic factors, it brings the pathological changes in morphology, physiological and biochemical. Plant disease, which impede the normal growth and development of a plant, occurred on leaves, fruits and roots. It is so important to detect the leaves disease effectively that can prevent disease developing fast. In this study, the objects are two typical plants’ leaves (tomato and strawberry leaves). The plant diseases are yellow leaf curl disease (TYLCD), colletotrichum, late blight, target and bactreria spot. These plant leaves disease are widly spread in south area because of the hot and humid climate in the South. This study was conducted in the past four years (2012～2015).Itincluded discrimination of plant leaves diseases, the classification of different degrees of leaves disease and early detection of asymptomatic infected leaves using spectral imaging technology. The results brought benefits to the further study of in-field crop management, which can improve the quilty of fruits and vegetables. Besides, the conculsions of this study can be used as a reference to other plant leaves diseases.The main contents and conclusions are as followed:(1) It recommend a method of calualating spectral features using the combination of spectral pre-processing, spectral analyzing and peak/bottoms in spectra. By comparing the discrimination results of each spectral feature, it improved that the method was useful to discriminat plant leaves diseases like TYLCD. The plants were infected with tomato yellow leaf curl virus using tootpicks in a controlled climate chamber. Two leaf areas (leaf vein and leaf edge) were choosed as region of interests (ROIs). The reflectance spectra of ROIs (400-1000 nm) were normalized before analysing. The first derievtive spectra and the curves of absolute different values were analyzed. The sensitive wavelengths were selected from the peak/bottom of spectra and curves. There were 2 sensitive wavelenths,4 band ratios and 4 spectral vegetation indices caluculated as spectra indexes. Each of the spectra indexes were evaluated by binary logestic regression (BLR). The results improved this method was useful to discriminate healthy and infected leaves whose overall accuracies reached 100%.(2) It recommened a spectral analyzing way that using spectral pre-processing, spectral feature calculation and classification algorithm. This method realized the classification of different infected degrees of leaves and ealy detection of plant disease like late blight, target and bacteria spot. Three diseased tomato leaves (late blight, target and bacteria spot) were picked in the field. There were four classes of these leaves:healthy, asymptomatic, ealy stage and late stage. The reflectace spectra (400～2500 nm) of leaves were collected in the lab.56 spectra indexes including sensitive wavelengths, band ratios and vegetation indices were calculated from rthe normalized reflectance spectra. By principal component analysis (PCA), the spectra indexes were selected and grouped in 12 ways. Finally, the classification was done by K nearest neighbors (KNN) in 12 ways. The results improved the ability of PCA-KNN algorithm to classify different stages of infected leaves and early detection of asymptomatic leaves with overall accuracies of 100%.(3) It recommend a method of using texture features extracting from images at sensitive wavelengths and band ratio images to discriminate the plant disease like TYLCD. The images at sensitive wavelengths (720 nm and 852 nm) and band ratio images (560/575 nm and 720/840 nm) were studied according to spectra indices. The binary mask images were created from images at 852 nm. The leaf pixies of both images at 720 nm and band ratios (560/575 nm and 720/840 nm) were segmented by adding original images and binary mak images.24 texture features were extracted by gray-level co-occurrence matrix (GLCM) and evaluated by Lenven’ test. Finally,22 texture features with significant difference were used for discrimination. The accuracies of Younden’s index, K nearest neighbors (KNN) and stepwise dicriminant analysis (SDA) were compared. According to the results, KNN algorithm had a high accuracy and was less affected by the character of features; the accuracy of Younden’s index was completely depended on the selection of features and it required high standard of features; SDA algorithm has the highest accuracy, while it was greatly influenced by the features. This mehod improved the possibility of using texture features to discriminat plant disease infected by virus and it had high accuracies of healthy, infected and overall leaves (100%).(4) It recommend a method extracting leaf vein feature using hit-or-miss transiformation and grey level co-occurrence matrix (HMT-GLCM). It improved the feasibility of early detection of plant disease using leaf vein features. Blue LEDs (centeral wavelength:430 nm) were used to excitat cholroplly fluorescence in 690 nm. Collecting chlorophyll fluorescence images was conducted in the lab. Leaf vein images were subtracted by automatic interative threshold method and hit-or-miss transformation. Leaf vein features were extracted in the way texture features as a reference. Three classification algrotihms including K nearest neighbor (KNN), binary logestic regression (BLR) and liner discriminant analysis (LDA) were tested before and after features extraction, respectively. The results showed the accuccries of KNN and LDA of infected leaves were both rised to 100% after feature extraction, while that of BLR decreased to 93.2%. Considering the randomness of samples in test group, this method offered an effective way to realize early detection of plant disease based on leaf vien features which can use less features to reach the same or even higher accuccies.(5) It improved the feasibility of early detection of plant disease in the field. An in-field device was developed to collect reflectance spectra of strawberry leaves (400～2500 nm). After normalization of raw spectra,33 spectra indexes were calculated as inputs to the classifiers K neast neighbor (KNN), stepwise discriminant analysis (SDA) and Fisher discriminant analysis (FDA) to classify healthy, asymptomatic and symptomatic leaves.The experiment was also conducted in the lab to test the results of in-field dataset. The results showed that the accuries of three leaves classes of in-field dataset ranging from 75.8%-77.3%, which were 10%-20% lower than that of indoor dataset. For the indoor dataset, FDA had the accuries raning from 81.6%-89.7%. The accuracy of healthy leaves of SDA was 93.1%, while the accuracy of asymptomatic leaves was 84.2%. For the in-field dataset, the accuracies of FDA were 10%-15% lower and the same to SDA. When using unbalanced dataset, there was a misclassification in KNN which increased the accuracy of large quantity samples (asymptomatic leaves) and decresed the accuracy of small quantity samples (healthy and symptomatic leaves). After adjusting the samples to a banlanced dataset, the error was eliminated. By comparing three classifiers, KNN was the best to detect plant disease of both the indoor and in-filed detection.