| Citrus is one of the world’s largest fruits industry, which has a pivotal role in rural economic development of main producing areas and the world trade. However, there are many kinds of citrus diseases and physiological diseases which prevented the development of citrus production and domestic and international markets. In recent years, nondestructive rapid detection of plant diseases has become a hotspot of plant disease detection research, and as one of the new non-destructive detection techniques, spectroscopic techniques have been applied to the diagnosis of plant diseases. Spectral detection techniques of crop diseases were well developed at home and abroad in recent years, but spectral detection studies related citrus diseases were relatively less. While some symptoms of citrus diseases were similar and easily confused, knowledge of hyperspectral imaging, spectral analysis, pattern recognition, chemometrics, citrus diseases and many other areas were integrated in this study. At the same time, detection of several important citrus diseases and nutrient deficiencies (physiological diseases) using hyperspectral imaging nondestructive techniques were studied, in order to provide theoretical basis for real time detection of citrus diseases and the development of rapid diagnostic systems. The main contents and results of the study are described as follows:1. Using hyperspectral imaging technology (400-1000nm), combined with chemometric methods, recognition of disease infected and healthy leaves, different disease infected leaves of several important citrus diseases were studied.2. Choosing Newhall as the research object, normal and diseased leaves including green and yellow citrus leaves were collected for from the huanglong disease onset orchards and normal orchards. Hyperspectral images were obtained from the front and back of the citrus leaves, at the same time molecular detection was performed to verification. Three models including LDA, BPNN and LSSVM citrus greening detection models were developed using the hyperspectral imaging reflectance data. The results showed that among the three models, the LSSVM models obtained the best results: Based on the front blade spectra and 9 point smoothing spectra preprocessing method, the recognition accuracy of modeling sets and prediction set were 100% and 92.5%; Based on the back blade spectra and the second derivative spectra preprocessing method, the recognition accuracy of the modeling and prediction set were 100% and 92.5%. Based on the front blade spectra, using the nine selected optimal wavelengths (400.19,403.17,406.15,407.64,412.12,721.14,730.74,740.34 和 823.98 nm) as inputs, the input variables of the developed LSSVM model were greatly reduced and the results were good, the recognition accuracy of modeling sets and prediction set were 98.57% and 92.5%. The results showed that the front blade spectra and the back blade spectra both can be used for citrus greening detection, and it was possible to identify citrus greening PCR positive leaves with normal leaf color. It showed that early detection of citrus greening infected leaves was possible using the hyperspectral imaging technology.3. Hyperspectral imaging of citrus brown spot, citrus black spot and citrus canker were collected, and the spectral reflectance characteristics of lesions tissues and tissues nearby the lesions in different places were analyzed and compared. The optimal wavelengths (404.66, 421.1, 428.6, 434.62, 436.12, 446.68, 618.04, 700.40, 719.55, 727.54, 864.38, 938.93 and 998.96 nm) were extracted to distinguish the three diseases lesions. Using the optimal wavelengths and Fisher multi-class linear analysis, the recognition rate of the three diseases lesions reached 100%, which showed the advantages of hyperspectral imaging technology.4. Hyperspectral reflectance images of healthy and diseased leaves infected with different strains of CTV including TRL514, CT30, CT32 and CT11A were collected. Comparative analysis was performed among supervised classification models, including MD, LDA and BPNN. Classifier models including MD, LDA and BPNN can discriminate the healthy and CTV-infected leaves with the highest classification accuracies of 100% in the spectral range of 400-1000nm and 760-1000 nm. Nine optimal wavelengths (405.4, 424.1, 920.28, 947.04, 957.59, 972.19, 978.68, 980.3 and 998.15 nm) selected by stepwise regression resulted in 97.33% total classification accuracy for differentiation of healthy and CTV-infected leaves and showed great potential in CTV diagnosis. However, the best overall classification accuracy of different CTV strains infected leaves resulted in 70%.5. Using hyperspectral imaging technology, the spectral response characteristics of ’Hamlin’ Sweet Orange leaves lacking Mn and/or Zn at different extents were detailedly characterized and then three recognition models were also built using LDA, BPNN and LSSVM, subsequently the recognition accuracy of which was compared. Our results showed that the model built by LSSVM with the whole spectral range wavelength could reach 91.88% recognition accuracy in modeling set and 90.00% in prediction set; in addition, the model built by LSSVM with forty wavelengths selected by successive projections algorithm (SPA) resulted in 82.50% recognition accuracy. The results in this study indicated that it was possible to identify Mn and/or Zn deficiency using hyperspectral imaging technology, meanwhile which laid a foundation for application of hyperspectral imaging technology to the identification of other nutritio’ n deficiency in citrus. |
PDF Full Text Request