<%BANNER%>

Early Detection of Citrus Greening (hlb) Using Ground Based Hyper-Spectral Imaging and Spectroscopy

Permanent Link: http://ufdc.ufl.edu/UFE0041616/00001

Material Information

Title: Early Detection of Citrus Greening (hlb) Using Ground Based Hyper-Spectral Imaging and Spectroscopy
Physical Description: 1 online resource (153 p.)
Language: english
Creator: Mishra, Ashish
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: hlb, hyperspectral, logistic, multi, pls, spectroscopy, support
Agricultural and Biological Engineering -- Dissertations, Academic -- UF
Genre: Agricultural and Biological Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Citrus greening, also known as Huanglongbing or HLB, is a major threat to the U.S. citrus industry. Currently, scouting and visual inspection are used for screening infected trees. However, this is a time-consuming and expensive method for HLB disease detection. Moreover, as it is subjective, the current method may involve high detection error rates. The objective of this research was to evaluate the optical sensors for the detection of HLB, other diseases and nutrient deficiencies in citrus. This dissertation describes the status of citrus in Florida, current HLB status in FL, various advanced techniques for plant disease detection. It further reviews various diseases and nutrient deficiencies in citrus that may be confused with HLB. Initially, the spectral characteristics of healthy and HLB infected tree canopies were investigated. A FieldSpec spectroradiometer (350-2500 nm) was used to detect HLB-infected trees. Discriminability, spectral derivative analysis and spectral ratio analysis were used to distinguish HLB. It was found that the spectral bands of green to red and near infrared have the ability to discriminate HLB-infected trees from healthy trees. These wavelength regions include green peak wavelengths at around 530-564 nm, 710-715 nm (red edge), and near infrared wavelengths of 1041 nm and 2014 nm. In the next step partial least squares (PLS) and discriminant statistical analyses were used to identify and discriminate spectral characteristics of HLB infections in citrus trees. Results suggest that both techniques have the potential to discriminate HLB for different varieties of citrus. Overall, the full range of data gave more accurate results compared to a narrower range of reflectance data with both statistical techniques. However, the narrower, visible, range (400 nm to 900 nm) data produced better results with PLS modeling. In contrast, discriminant analysis produced better overall results with the full reflectance range. Machine learning techniques like k-nearest neighbors (KNN), logistic regression, and support vector machines (SVM) were applied for classifying the HLB data. Analysis showed that with one spectral measurement, none of the classification methods was successful in discriminating healthy from infected trees, because of the large variability in the spectral measurements. When five spectra from the same tree were used for classification, SVM and weighted KNN methods classified spectra with 3.0 and 6.5 percent error, respectively. The results from this study indicated that the canopy visible and near infrared (VIS-NIR) spectral reflectance can be used for detecting HLB infected citrus trees. However, high classification accuracy ( > 90%) requires multiple measurements from a single tree. Since ASD and SVC spectroradiomters are very expensive and difficult for growers to use in field data collection, a rugged, low-cost, multi-band active optic sensor was used to identify the HLB infected trees from the healthy trees. The sensor consisted of four bands: two visible bands at 570 nm and 670 nm, and two NIR bands at 870 nm and 970 nm. Extensive field measurements were conducted using this sensor. Analysis of the data showed that due to the large variability in the data, it was not possible to discriminate healthy and infected trees based only on a single measurement from a tree. Using multiple measurements from a tree, however, it was possible to achieve high classification accuracy. With five measurements from a tree, classification methods such as k-nearest neighbors, support vector machines, and decision trees achieved classification errors of less than 5 percent. The results demonstrated the potential of a multi-band active optic sensor for detecting HLB-infected citrus trees under field conditions. This research further investigated the application of hyperspectral camera for HLB detection. Hyperspectral images of HLB infected trees and healthy trees were collected with a Specim hyperspectral camera (Autovision Inc., Los Angeles, CA, USA) having a spectral range from 306.5 nm to 1067.1 nm with 2.7 nm spectral resolution. These images were processed in ENVI 4.5 (ITT Visual Information Solutions, Boulder, Colorado). Various vegetation indices were estimated. ANOVA was used to compare the mean vegetation indices of healthy and HLB trees. Results showed that hyperspectral imaging have a potential to discriminate HLB from healthy samples.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Ashish Mishra.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Ehsani, M Reza.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2012-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041616:00001

Permanent Link: http://ufdc.ufl.edu/UFE0041616/00001

Material Information

Title: Early Detection of Citrus Greening (hlb) Using Ground Based Hyper-Spectral Imaging and Spectroscopy
Physical Description: 1 online resource (153 p.)
Language: english
Creator: Mishra, Ashish
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: hlb, hyperspectral, logistic, multi, pls, spectroscopy, support
Agricultural and Biological Engineering -- Dissertations, Academic -- UF
Genre: Agricultural and Biological Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Citrus greening, also known as Huanglongbing or HLB, is a major threat to the U.S. citrus industry. Currently, scouting and visual inspection are used for screening infected trees. However, this is a time-consuming and expensive method for HLB disease detection. Moreover, as it is subjective, the current method may involve high detection error rates. The objective of this research was to evaluate the optical sensors for the detection of HLB, other diseases and nutrient deficiencies in citrus. This dissertation describes the status of citrus in Florida, current HLB status in FL, various advanced techniques for plant disease detection. It further reviews various diseases and nutrient deficiencies in citrus that may be confused with HLB. Initially, the spectral characteristics of healthy and HLB infected tree canopies were investigated. A FieldSpec spectroradiometer (350-2500 nm) was used to detect HLB-infected trees. Discriminability, spectral derivative analysis and spectral ratio analysis were used to distinguish HLB. It was found that the spectral bands of green to red and near infrared have the ability to discriminate HLB-infected trees from healthy trees. These wavelength regions include green peak wavelengths at around 530-564 nm, 710-715 nm (red edge), and near infrared wavelengths of 1041 nm and 2014 nm. In the next step partial least squares (PLS) and discriminant statistical analyses were used to identify and discriminate spectral characteristics of HLB infections in citrus trees. Results suggest that both techniques have the potential to discriminate HLB for different varieties of citrus. Overall, the full range of data gave more accurate results compared to a narrower range of reflectance data with both statistical techniques. However, the narrower, visible, range (400 nm to 900 nm) data produced better results with PLS modeling. In contrast, discriminant analysis produced better overall results with the full reflectance range. Machine learning techniques like k-nearest neighbors (KNN), logistic regression, and support vector machines (SVM) were applied for classifying the HLB data. Analysis showed that with one spectral measurement, none of the classification methods was successful in discriminating healthy from infected trees, because of the large variability in the spectral measurements. When five spectra from the same tree were used for classification, SVM and weighted KNN methods classified spectra with 3.0 and 6.5 percent error, respectively. The results from this study indicated that the canopy visible and near infrared (VIS-NIR) spectral reflectance can be used for detecting HLB infected citrus trees. However, high classification accuracy ( > 90%) requires multiple measurements from a single tree. Since ASD and SVC spectroradiomters are very expensive and difficult for growers to use in field data collection, a rugged, low-cost, multi-band active optic sensor was used to identify the HLB infected trees from the healthy trees. The sensor consisted of four bands: two visible bands at 570 nm and 670 nm, and two NIR bands at 870 nm and 970 nm. Extensive field measurements were conducted using this sensor. Analysis of the data showed that due to the large variability in the data, it was not possible to discriminate healthy and infected trees based only on a single measurement from a tree. Using multiple measurements from a tree, however, it was possible to achieve high classification accuracy. With five measurements from a tree, classification methods such as k-nearest neighbors, support vector machines, and decision trees achieved classification errors of less than 5 percent. The results demonstrated the potential of a multi-band active optic sensor for detecting HLB-infected citrus trees under field conditions. This research further investigated the application of hyperspectral camera for HLB detection. Hyperspectral images of HLB infected trees and healthy trees were collected with a Specim hyperspectral camera (Autovision Inc., Los Angeles, CA, USA) having a spectral range from 306.5 nm to 1067.1 nm with 2.7 nm spectral resolution. These images were processed in ENVI 4.5 (ITT Visual Information Solutions, Boulder, Colorado). Various vegetation indices were estimated. ANOVA was used to compare the mean vegetation indices of healthy and HLB trees. Results showed that hyperspectral imaging have a potential to discriminate HLB from healthy samples.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Ashish Mishra.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Ehsani, M Reza.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2012-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041616:00001


This item has the following downloads:


Full Text

PAGE 1

1 EARLY DETECTION OF CITRUS GREENING (HLB) USING GROUND BASED HYPERSPECTRAL IMAGING AND SPECTROSCOPY By ASHISH RATN MISHRA A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010

PAGE 2

2 2010 Ashish Ratn Mishra

PAGE 3

3 To my parents my loving wife and daughter

PAGE 4

4 ACKNOWLEDGMENTS I would like to bestow my sincere gratitude to advisor and dissertation chair Dr. Reza Ehsani, Assistant Professor of Agricultural and Biological Engineering, Citrus Research and Education Center (CREC), University of Florida (UFL) for his consistent guidance, encouragement and support throughout this research work at UF His thorough and thoughtful coaching was unselfishly tireless, and his enthusiasm has left me an everlasting impression. I am greatly indebted to my supervisory committee Dr. Lee, Associate Professor of Agricultural an d Biological Engineering, UFL, Dr. Masoud Salyani Professor of Agricultural and Biological Engineering, CREC, UFL, Dr. John Schueller, Professor of Mechanical and Aerospace Engineering, UFL and Dr. Amr Abd Elrahman, Assistant Professor of Geomatics, UFL f or their guidance and suggestions to complete this work. Their ideas wisdom, and suggestions have helped me sail smoothly through the graduate studies Dr. Gene Albrigo, Emeritus Professor, CREC helped with numerous and detailed criticisms of parts of the dissertation. The field experiment would not have successful without the help of Sherrie Buchanon, Dr. Joe M Maja, Raghav Panchapakesan, Sajith Udumala, Bhargav Prasad, Andre Colaco and John Pilkey. In addition, I would like to thank Dr. Davood Karimi and Dr. Sindhuja Sankaran and for their guidance in my data analysis. Cecile Robertson at CREC deserves special thanks for her assistance in greenhouse study and PCR testing. I also thankful to Jennifer Dawson and Kathy Snyder at CREC library for providing me literature and other lab support whenever I required. I am greatly thankful to the staff of the Agricultural and Biological Engineering department and CREC family for sharing their insights into getting through graduate

PAGE 5

5 research. I would like to thank all of those not explicitly mentioned here who have aided my intellectual and social growth throughout my academic career. Finally, I would like to extend my special thanks to my parents and my wife Ekta for their continued moral support, love and care throug hout this milestone of my life. I bow my head before the Almighty, for all the blessing showered on me during the entire course of this work.

PAGE 6

6 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 9 LIST OF FIGURES ................................ ................................ ................................ ........ 10 LIST OF ABBREVIATIONS ................................ ................................ ........................... 13 ABSTRACT ................................ ................................ ................................ ................... 15 CHAPTER 1 GENERAL INTRODUCTION ................................ ................................ .................. 18 In troduction ................................ ................................ ................................ ............. 18 Dissertation Organization ................................ ................................ ........................ 18 Literature Review ................................ ................................ ................................ .... 19 Huanglongbing or Citrus Greening ................................ ................................ ... 19 Interaction of Light with Vegetation, Soil and Water ................................ ......... 21 Advance Techniques for Detecting Plant Disease ................................ ............ 27 Spectroscopic and Imaging Techniques ................................ .................... 27 Visible Near Infrared Spectroscopy ................................ ........................... 28 Hyperspectral and Multispectral Imaging ................................ ................... 31 Application of Spectroscopy and Imaging in Citrus ................................ .......... 33 2 REVIEWS ON HLB AND OTHER SIMILAR DISEASE AND DEFICIENCIES ......... 35 Chlorosis ................................ ................................ ................................ ................. 35 Perchlorate Chlorosis ................................ ................................ ....................... 35 Biuret Tox icity ................................ ................................ ................................ ... 36 Arsenic Toxicity ................................ ................................ ................................ 37 Fluorine Toxicity ................................ ................................ ............................... 37 Mechanism of Chlorosis ................................ ................................ .......................... 38 Ultrastructural Changes in Chloroplasts during Senescence ............................ 38 Autonomous Degradation of Chloroplasts ................................ ........................ 38 Chlorophyll Degradation ................................ ................................ ................... 39 Changes in Lipids during Chloroplasts Senescence ................................ ........ 39 Changes in Stromal Enzymes during Leaf Senescence ................................ ... 39 Changes in the Components of the Chloroplast Thylakoid Membranes During Foliar Senescence ................................ ................................ ............. 39 Chloroplasts Protein Degradation ................................ ................................ ..... 40 Leaf Conductance and CO 2 Assimilation in Senescense Leaves ..................... 40 Chlorosis Due to Nutrient Deficiency ................................ ................................ ...... 41 Iron ................................ ................................ ................................ ................... 41

PAGE 7

7 Nitrogen ................................ ................................ ................................ ............ 42 Calcium ................................ ................................ ................................ ............ 43 Manganese ................................ ................................ ................................ ....... 44 Magnes ium ................................ ................................ ................................ ....... 45 Molybdenum ................................ ................................ ................................ ..... 46 Potassium, Phosphorus and Sulfur ................................ ................................ .. 47 Zinc ................................ ................................ ................................ .................. 51 Boron ................................ ................................ ................................ ................ 51 Copper ................................ ................................ ................................ .............. 53 Diseases in Citrus ................................ ................................ ................................ ... 54 Alternaria Bro wn Spot ................................ ................................ ...................... 54 Black Spot ................................ ................................ ................................ ........ 55 Canker ................................ ................................ ................................ .............. 56 Mal Secco ................................ ................................ ................................ ......... 58 Melanose ................................ ................................ ................................ .......... 59 Powdery Mildew ................................ ................................ ............................... 60 Scab ................................ ................................ ................................ ................. 60 Huanglongbing (Greening) ................................ ................................ ............... 61 Leprosis ................................ ................................ ................................ ............ 63 Citrus Variegated Chlorosis ................................ ................................ .............. 64 3 SPECTRAL CHARACTERIS TICS OF CITRUS GREEN ING (HUANGLONGBING) ................................ ................................ .............................. 66 Introduction ................................ ................................ ................................ ............. 66 Objective ................................ ................................ ................................ ................. 67 Materials and Me thods ................................ ................................ ............................ 67 Data Analysis ................................ ................................ ................................ .......... 68 Discriminability ................................ ................................ ................................ 68 Spectral Derivative Analysis ................................ ................................ ............. 69 Spectral Ratio Analysis ................................ ................................ ..................... 70 Results and Discussion ................................ ................................ ........................... 71 Discrimina bility ................................ ................................ ................................ 71 Spectral Derivative Analysis ................................ ................................ ............. 72 Spectral Ratio Analysis ................................ ................................ ..................... 73 Conclusion ................................ ................................ ................................ .............. 75 4 SPECTRAL DISCRIMINAT ION OF HEALTHY VS. H LB INFECTED CITRUS TREES IN THE VIS NIR RANGE ................................ ................................ ........... 76 Introduction ................................ ................................ ................................ ............. 76 Material and Methods ................................ ................................ ............................. 79 Results and Discussion ................................ ................................ ........................... 81 Conclusions ................................ ................................ ................................ ............ 90 5 IDENTIFICATION OF CITRUS GREENING (HLB) INFECTED CITRUS TREES USING SPECTROSCOPY AND STATISTICAL CLASSIFICATION ....................... 92

PAGE 8

8 Introduction ................................ ................................ ................................ ............. 92 Materials and Methods ................................ ................................ ............................ 94 Field experiments ................................ ................................ ............................. 94 Data Analysis ................................ ................................ ................................ ... 96 Spectral pretreatment and feature selection ................................ .............. 96 Classification ................................ ................................ .............................. 98 Weighted K Nearest Neighbors (KNN) ................................ ...................... 98 Logistic Regression (LR) ................................ ................................ ............ 99 Support Vector Machines (SVM) ................................ .............................. 100 Reducing the C lassification E rror by U sing M ultiple M easurements ........ 107 Results and Discussion ................................ ................................ ......................... 109 Conclusi on ................................ ................................ ................................ ............ 111 6 AN ACTIVE OPTIC SENSOR FOR DETECTION OF HUANLONGBING (HLB) DISEASE ................................ ................................ ................................ .............. 113 Introduction ................................ ................................ ................................ ........... 113 Material and Metho ds ................................ ................................ ........................... 116 Data C ollection with M ulti B and S ensor ................................ ......................... 116 Data Analysis ................................ ................................ ................................ 120 Decision Trees ................................ ................................ ......................... 121 k Nearest Neighbors (KNN) ................................ ................................ ..... 121 Logistic Regression ................................ ................................ .................. 121 Neural Networks ................................ ................................ ...................... 122 Support Vector Machines (SVM) ................................ .............................. 122 Results and Discussion ................................ ................................ ......................... 123 Conclusions ................................ ................................ ................................ .......... 125 7 APPLICATION OF HYPER SPECTRAL IMAGING FOR THE DETECTION OF HLB IN THE FIELD ................................ ................................ ............................... 126 Introduction ................................ ................................ ................................ ........... 126 Material and Methods ................................ ................................ ........................... 128 Data collection ................................ ................................ ................................ 128 Preprocessing ................................ ................................ ................................ 132 Processing ................................ ................................ ................................ ...... 135 Results and Discussion ................................ ................................ ......................... 135 Conclusions ................................ ................................ ................................ .......... 137 8 SUMMARY AND RECOMM ENDATIONS ................................ ............................. 139 165

PAGE 9

9 LIST OF T ABLES Table page 1 1. Absorption features in visible and near Infrared wavebands that have been related to particular foliar chemical concentration ................................ ............... 25 2 1. Summary of the effects of mineral deficiencies on chloroplasts structure .............. 50 3 1 Discriminability of wavelengths for HLB and healthy trees ................................ ..... 71 3 2. Identified wavelengths for separating HLB trees from healthy trees ...................... 73 4 1. PLS modeling for HLB and healthy trees showing total samples, correct classifications (June 13, 2007). ................................ ................................ .......... 82 4 2. PLS modeling for HLB and healthy trees showing total samples, correct classifications (June 14, 2007) ................................ ................................ ........... 84 4 3. PLS modeling for HLB and healthy trees showing total samples, correct classifications in a greenhouse with artificial light (Aug 2, 2007). ....................... 85 4 4. PLS modeling for HLB and healthy trees showin g total samples, correct classifications in a greenhouse with natural light (Aug 3, 2007) ......................... 87 4 5. Number of misclassified spectra i n discriminant analysis. ................................ ...... 89 5 1. Spectral data from healthy and HLB infected trees used in this study ................... 95 5 2. Average classification error for three classification techniques. ........................... 110 6 1. List of vegetation indices used in analysis. ................................ .......................... 120 6 2. Average classification error for different classification techniques. ...................... 124 7 1. Means of vegetation indices o f HLB infected and healthy trees ........................... 136

PAGE 10

10 LIST OF FIGURES Figure page Figure 1 1. Symptoms of HLB in citrus leaves ................................ ............................ 20 Figure 1 2. HLB infected citrus fruits ................................ ................................ ............ 20 Figure 1 3. Spectral reflectance of soil, green and dry vegetation ............................... 22 Figure 2 1. Biuret chlorosis in citrus leaves ................................ ................................ .. 37 Figure 2 2. Iron deficiency in Orange leaves ................................ ................................ 41 Figure 2 3. Nitrogen defic iency in citrus leaves ................................ ............................ 43 Figure 2 4. Calcium deficiency ................................ ................................ ..................... 44 Figure 2 5. Manganese deficiency ................................ ................................ ............... 45 Figure 2 6. Deficiency symptoms of magnesium in grapefruit ................................ ...... 46 Figure 2 7. Molybdenum deficiency in Orange leaves ................................ .................. 47 Figure 2 8. Potassium deficiency in Orange leaves ................................ ..................... 48 Figure 2 9. Phosphorus deficiency ................................ ................................ ............... 48 Figure 2 10. Sulfur deficiency in Orange leaves ................................ ............................ 49 Figure 2 11. Zinc deficiency in orange leaves ................................ ............................... 51 Figure 2 12. Boron deficiency ................................ ................................ ....................... 52 Figure 2 13. Copper deficiency ................................ ................................ .................... 54 Figure 2 14. Alternaria brown spot in orange fruit ................................ ........................ 55 Figure 2 15. Black spot ................................ ................................ ................................ 56 Figure 2 16. Citrus canker ................................ ................................ ............................. 57 Figure 2 17. Mal secco ................................ ................................ ................................ .. 58 Figure 2 18. Melanose ................................ ................................ ................................ .. 59 Figure 2 19. Powdery mildow ................................ ................................ ........................ 60 Figure 2 20. Scab ................................ ................................ ................................ .......... 61

PAGE 11

11 Figure 2 21. Huanglongbing (HLB) ................................ ................................ ............... 62 Figure 2 22. Leprosis ................................ ................................ ................................ .... 64 Figure 3 1. A sample spectra of a healthy and HLB infected tree canopy ..................... 72 Figure 3 2. Spectral ratio of HLB infected and healthy trees ................................ ......... 74 Figure 4 1. Canopy reflectance of healthy and HL B infected tree with FieldSpec 3 spectroradiometer. ................................ ................................ .............................. 81 Figure 4 2. Percentage of correct classification of HLB and healthy trees in full NIR range (400 2450 nm) and narrow NIR range (400 900 nm) on June 13, 2007 ... 83 Figure 4 3. Percentage of correct classification of HLB and healthy trees in full range (400 2450 nm) and narrow NIR range (400 900 nm) on June 14, 2007. .. 84 Figure 4 4. Percentage of correct classification of HLB and healthy trees in full range (400 2450 nm) and narrow NIR range (400 900 nm) on Aug 2, 2007. ..... 86 Figure 4 5. Percentage of correct classification of HLB and healthy trees in full range (400 2450 nm) and narrow NIR range (40 0 900 nm) on Aug 3, 2007. ..... 88 Figure 4 6. Canoniocal plot shows the points and the multivariate means of HLB and healthy trees in full range for Aug 2, 2007. ................................ .................. 90 Figure 5 1. Variance of data explained by principal component analysis ...................... 98 Figure 5 2. The logistic curve describing logistic regression model ............................ 100 Figure 5 3. A) An example of a linearly separable set of data, and B) the maximum margin classifier for this data set. ................................ ................................ ..... 101 Figure 5 4. Representative spectroradiometer spectra: the spectra from two healthy and two HLB infected trees. ................................ ................................ ............. 109 Figure 5 5. Contour plots of classification error for finding the optimum values for the parameters C and ................................ ................................ ................... 11 0 Figure 6 1. HLB symptomatic ................................ ................................ ...................... 113 Figure 6 2. Healthy leave s ................................ ................................ .......................... 114 Figure 6 3. Multi band active optic sensor ................................ ................................ ... 119 Figure 6 4. Field measurements using the four band sensor ................................ ...... 119 Figure 7 1. Schematic representation of principle behind hyperspectral imaging ....... 128

PAGE 12

12 Figure 7 2. Data collection sites in Florida ................................ ................................ .. 129 Figure 7 3. Hyperspectral camera ................................ ................................ ............... 130 Figure 7 4. Hyperspectral image acquisition in field ................................ .................... 130 Figure 7 5. Flow chart of methodology ................................ ................................ ........ 131 Figure 7 6. Raw image acquired from hyperspectral camera ................................ ...... 133 Figure 7 7. Final mask for removing background as sky, soil, grass etc. .................... 134 Figure 7 8. Processed image used for data analysis ................................ .................. 135 Figure 7 9. Vegetation indices of healthy and HLB trees ................................ ............ 137

PAGE 13

13 LIST OF ABBREVIATION S ADAR Airborne Data acquisition and registration ANOVA Analysis of Variance ANN Artificial Neural Network BP NN Back Propagation Neural Network CART Classification And Regression Tree CO 2 Carbon di Oxide DPLS Discriminant Partial Least Squares ELISA Enzyme Linked Immunosorbent Assay FASS Florida Agricultural Statistics Survey HLB Huanglongbing or greening G Greenness Index GAE G allic A cid E quivalent IFOV Instantaneous Field of View IF Immunofluorescence KNN K Nearest Neighbors LDA Linear Discriminant Analysis LR Logistic Regression LIF L aser I nduced F luorescence LVQ Learning Vector Quantization mPLS Modified Partial Least Squares MCARI Modified Chlorophyl l Absorption in Reflectance Index MNF Minimum Noise Fraction MTVI Modified Triangular Vegetation Index NASS National Agriculture Statistics Survey

PAGE 14

14 NDVI Normalized Difference Vegetation Index NIR Near Infrared nm Nanometer PCA Principal Component Analysi s PCR Polymer Chain Reaction PDA Procrustes Discriminant Analysis PLS Partial Least Squares PNN Probabilistic Neural Network (PNN) QDA Quadratic Discriminant Analysis R Reflectance RDVI Renormalized Difference Vegetation Index SAM Spectral Angle Mapping SFF Spectral Feature Fitting SIMCA Soft Independent Modeling of Class Analogy SIPI Structure Intensive Pigment Index SOM Self Organizing Maps SR Simple Ratio Index SVM Support Vector Machine TVI Triangular Vegetation Index UV Ultra Violet VIS Visible

PAGE 15

15 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy DETECTION OF CITRUS GREENING (HLB) USING GROUND BASED HYPER SPECTRAL IMAGING AND SPECTROSCOPY By Ashish Ratn Mishra December, 2010 Chair: Reza Ehsani Major: Agricultural and Biological Engineering Citrus greening, also known as Huanglongbing or HLB, is a major threat to the U.S. citrus industry. Currently, scouting and visual inspection are used for screening infected trees. However, this is a time consuming and expensive method for HLB disease detection Moreover, as it is subjective, the current method may involve high detection error rates. The objective of this resea rch w as to evaluate the optical sensors for the detection of HLB, other diseases and nutrient deficiencies in citrus. This dissertation describes the status of citrus in Florida, current HLB status in FL, various advanced techniques for plant disease dete ction. It further reviews various diseases and nutrient deficiencies in citrus that may be confused with HLB. Initially, the spectral characteristics of healthy and HLB infected tree canopies were investigated A FieldSpec spectroradiometer (350 2500 nm) was used to detect HLB infected trees. Discriminability, spectral derivative analysis and spectral ratio analysis were used to distinguish HLB. It was found that the spectral bands of green to red and near infrared have the ability to discriminate HLB infe cted trees from healthy

PAGE 16

16 trees. These wavelength regions include green peak wavelengths at around 530 564 nm, 710 715 nm (red edge), and near infrared wavelengths of 1041 nm and 2014 nm. In the next step p artial least square s (PLS) and discriminant statistical analyses were used to identify and discriminate spectral characteristics of HLB infections in citrus trees. Results suggest that both techniques have the potential to discriminate HLB for different varieties of citrus. Overall, the full range o f data gave more accurate results compared to a narrower range of reflectance data with both statistical techniques. However, the narrower, visible, range (400 nm to 900 nm) data produced better results with PLS modeling. In contrast, discriminant analysis produced better overall results with the full reflectance range. Machine learning techniques like k nearest neighbors (KNN), logistic regression, and support vector machines (SVM) were applied for classifying the HLB data. Analysis showed that with one sp ectral measurement, none of the classification methods was successful in discriminating healthy from infected trees, because of the large variability in the spectral measurements. When five spectra from the same tree were used for classification, SVM and w eighted KNN methods classified spectra with 3.0 and 6.5 percent error, respectively. The results from this study indicated that the canopy visible and near infrared (VIS NIR) spectral reflectance can be used for detecting HLB infected citrus trees. However high classification accuracy (> 90%) requires multiple measurements from a single tree. Since ASD and SVC spectroradiomters are very expensive and difficult for growers to use in field data collection a rugged, low cost, multi band active optic sensor w as us ed to identify the HLB infected trees from the healthy trees. The sensor consisted

PAGE 17

17 of four bands: two visible bands at 570 nm and 670 nm, and two NIR bands at 870 nm and 970 nm. Extensive field measurements were conducted using this sensor. Analysis o f the data showed that due to the large variability in the data, it was not possible to discriminate healthy and infected trees based only on a single measurement from a tree. Using multiple measurements from a tree, however, it was possible to achieve hig h classification accuracy. With five measurements from a tree, classification methods such as k nearest neighbors, support vector machines, and decision trees achieved classification errors of less than 5 percent. The results demonstrated the potential of a multi band active optic sensor for detecting HLB infected citrus trees under field conditions. This research further investigated the application of hyperspectral camera for HLB detection. Hyperspectral i mages of HLB infected trees and healthy trees were collected with a Specim hyperspectral camera (Autovision Inc., Los Angeles, CA, USA) having a spectral range from 306.5 nm to 1067.1 nm with 2.7 nm spectral resolution. These images were processed in ENVI 4 .5 ( ITT Visual Information Solutions, Boulder, Colorado). Various vegetation indices were estimated. ANOVA was used to compare the mean vegetation indices of healthy and HLB trees. Results showed t hat hyperspectral imaging have a potential to discriminate HLB from healthy samples.

PAGE 18

18 CHAPTER 1 GENERAL INTRODUCTION Introduction with about 10 percent of the Florida account s for 71 percent of total U.S. Citrus production (NASS, 2009). since 2007 08. Grapefruit production in Florida (21.7 million boxes) has also reduced 18 percent from last season. One of the possible reasons for reduced citrus production could be at tributed to exotic diseases. Huanglongbing (HLB), also known as greening, is a systemic bacterial disease transmitted by a psyllid insect, is considered one of the most devastating citrus diseases in the world. HLB has been translated loosely as yellow sho ot disease because of the characteristic yellow shoots caused by the disease. It is caused by a phloem limited bacterium, Candidatus Liberibactor asiaticus HLB was first detected in Florida in August of 2005 ( Chung and Brlansky, 2005). Since this is a rel atively new disease in USA, very little published information is available on the dynamics, epidemiology, and molecular characteristics of this disease. Dissertation Organization This dissertation consists of a review of the literature and six chapters pr epared for partial fulfillment of the requirement for the degree, Doctor of Philosophy. The author of the dissertation is Ashish R Mishra. Dr. Reza Ehsani served as a major advisor provided all the facilities and constructive suggestions to conduct this re search. The first chapter serves as a detailed literature review on the Huanglongbing, interaction of light with vegetation, soil and water, various advance techniques used in plant disease detection specially hyperspectral spectroscopy and imaging. The se cond chapter

PAGE 19

19 reviews various diseases, nutrient deficiencies that could affect citrus production. Special focus was given on those diseases and deficiencies whose symptoms could be confused with HLB symptoms. The third chapter demonstrates the spectral cha racteristics of citrus leaves and identifies critical wavelength for HLB detection. The fourth chapter describes partial least squares (PLS) technique and discriminant analysis for the HLB identification. The fifth chapter is the manuscript, submitted in t ransactions of ASABE journal, reports the feasibility of hyperspectral spectroscopy and several machine learning techniques in HLB detection. The sixth chapter is the manuscript, submitted in biosystems engineering journal, reports the application of multi spectral sensor to detect HLB in field condition. The seventh chapter discusses the application of hyperspectral imaging and various vegetation indices used in HLB discussion. In the last chapter summary and future directions are included. Literature Revie w Huanglongbing or Citrus Greening Huang means yellow, long means dragon and bing refers to disease. Therefore, Huanglongbing refers names such as greening in many countries. HLB has destr oyed an estimated 60 million trees in Africa and Asia (Bove, 2006) Huanglongbing (HLB) is caused by the gram negative bacterium Candidatus Liberibacter asiaticus (Garnier et al., 2000). Asian citrus psyllid ( Diaphorina citri ) is the vector of citrus gre ening or HLB. The bacteria are restricted to the sieve tubes of infected plants, and are acquired and transmitted by nymphs and adults of Asian citrus psyllid during feeding (Garnier and Bov, 1983). Psyllids prefer feeding and breeding on

PAGE 20

20 younger leaves ( Halbert and Manjunath, 2004) resulting younger trees at a higher risk of infection as they produce newer leaves and flushes throughout the year. Figure 1 1. Symptoms of HLB in citrus leaves Figure 1 2 HLB infected citrus fruits Symptoms of HLB infected citrus include a blotchy mottle or asymmetrical chlorosis (Figure 1 1) and yellowing of leaf veins due to inefficient production of chlorophyll (Brlansky et al., 2007).The angular blotching has been considered specific for the disease and consis ts of blotches of yellow on dark greenish grey leaves. On the same tree, some branches may not be infected by HLB. Fruits of HLB infected trees are affected in various ways. Some of them, when they reach an inch or more in diameter

PAGE 21

21 become lopsided (Fig ure 1 2). Fruit matures only on one side with the immature side remaining green when the fruit ripens, hence the name "greening". Other normal shaped fruits that attain full size may fail to color properly and remain lusterless, greenish yellow, and many of th em fall before harvest. Infected fruits also lose their expected taste. Seeds are also affected from HLB infected trees. Since the outer seed coat does not develop sufficiently to cover the inner brown coat, partially developed seeds have the brown color o f the inner seed coat. Interaction of L ight with V egetation, S oil and W ater Plants absorb the ultraviolet and the visible regions of the spectrum very efficiently. The reflectance and transmittance of plant leaves increases dramatically in near infrared r egion, resulting in the absor b ance falling to a very low value. There are two regions of the spectrum where relatively less absorption occurs. At wavelengths longer than 1200 nm water vapor absorption rises very steeply, whereas in the red and the blue re gions of visible spectrum, pigment absorption is very strong. The presence of pigments other than chlorophyll tends to broaden the domain of absorption throughout the visible region. Scattering is also caused by structures within the leaf. Such structure m ay include mitochondria, ribosomes, nuclei, starch grains, and other plastids. The visible absorbance substantially increases from the lighter to the darker leaves and the NIR absorbance is the highest for the thinner leaf. The most striking feature in the near infrared is the fact that the transmittance of the thinner leaves is greater than the reflectance. A lack of chlorophyll pigmentation can reduce drastically the absorption of the visible light by a leaf. The white leaf exhibits very little absorptio n through the green and red spectral regions and only increases in the blue due to absorption by

PAGE 22

22 protochlorophyll. The carotenoid and chlorophyll pigments are partially lacking in the white leaf. Gates et al. (1965) considered that near infrared reflectanc e is a function of the cell shape and size as well as the amount of intercellular space. Initially, the mesophyll of a very young leaf contains spongy parenchyma with considerable air spaces. It favors the mechanism of internal reflection. When the leaf ma tures, the cell enlarges, crowding together, reducing the intercellular space and reducing the reflectance. It would then appear that during final maturing the cell structure and intercellular space relationship becomes favorable for increased reflectance. Figure 1 3 Spectral reflectance of soil, green and dry vegetation Refl ectance is characterized by a relative maximum in the green band at 550 nm and minimum at 400 and 670 nm caused by absorption of radiation by chlorophyll 0 10 20 30 40 50 60 350 750 1150 1550 1950 2350 Reflectance (%) Wavelength (nm) Citrus leaves Asphalt Grass Wet soil Dry soil

PAGE 23

23 (Figure 1 3). Reflectance within the 690 700 nm range is particularly sensitive to early stress induced decreases in leaf chlorophyll content ( Carter, 1993) and represent the shift of red edge that are closely related to chlorophyll content. Increased leaf reflectance near 530 nm h as indicated pigment transformations and changes in thylakoid processes (Gamon et al., 1990). Reddy et al. (2001) found that chlorophyll concentration of maize, groundnut and soybean crops mainly affected leaf spectral reflectance at 450 520 nm and 620 nm region. However, Zhao et al., (2003) reported that the chlorophyll correlated most strongly with the reflectance at 554 nm and 712 nm. Reflectance in this region is relatively high and mostly constant. This was related to the low absorption by the leaf. Mi nor relative minima were at 950 nm and 1160 nm due to selective absorption by the presence of water. After 1300 nm, the absorbance by water plays a dominant role with maxima at 1450 and 1950 nm. Fouche (1999) suggested that reflectance at the 779 nm wavele ngth may provide the best detection of N deficiency in cotton, tobacco and wheat. Carter and Estep (2002) reported that a simple linear relationship existed between leaf nitrogen (%) and reflectance at 721 nm in corn. Zhao et al. (2003) concluded that on t he b asis of leaf area chlorophyll a chlorophyll b carotenoid, and chlorophyll a+b concentration could be estimated using reflectance ratios in the near infrared region of 712 nm to 1088, 1097, 809 nm, respectively. Table 1 1 revealed that there are 42 minor absorption features t in fresh leaves that would probably have dampened by the five major absorption features. These absorption features are the result of the bending and stretching of the oxygen and hydrogen bond

PAGE 24

24 ( O H ) bond between carbon and nit rogen ( C N ) and single and double bonds between carbon and hydrogen ( C H, C=H ) in different chemicals.

PAGE 25

25 Table 1 1 Absorption features in visible and near Infrared wavebands that have been related to particular foliar chemical concentr ation (Curran, 1989). ( Chemicals in italics have a wavelength of stronger absorption) Wavelength (nm) Electronic transition or bond vibration Chemical(s) Remote Sensing Consideration 430 Electron transition Chlorophyll a Atmospheric scattering 460 Electron transition Chlorophyll b 640 Electron transition Chlorophyll b 660 Electron transition Chlorophyll a 910 C H stretch Protein 930 C H stretch Oil 970 O H stretch Water starch 990 O H stretch starch 1020 N H stretch Protein 1040 C H stretch, C H deformation Oil 1120 O H stretch, C H stretch, C H deformation Lignin 1200 O H bend Water, cellulose, starch, lignin 1400 O H bend Water 1420 C H stretch, C H deformation Lignin 1450 O H stretch, C H stretch, C H deformation Starch, sugar, lignin, water Atmospheric absorption 1490 O H stretch Cellulose, sugar 1510 N H stretch Protein, nitrogen 1530 O H stretch Starch 1540 O H stretch Starch, Cellulose 1580 O H stretch Starch, sugar 1690 C H stretch Lignin starch, protein, nitrogen 1780 C H stretch, O H stretch, H O H deformation Cellulose, sugar starch 1820 O H stretch, C O stretch Cellulose 1900 O H stretch, O H deformation Starch 1960 O H stretch, O H bend Sugar, starch

PAGE 26

26 Table 1 1. Continued Wavelength (nm) Electronic transition or bond vibration Chemical(s) Remote Sensing Consideration 1940 O H stretch, O H deformation Water lignin, protein, nitrogen, starch, cellulose Rapid decrease in signal to noise ratio of sensors 1960 O H stretch, O H bend Sugar, starch 1980 N H asymmetry Protein 2000 O H deformation, C O deformation Starch 2060 N=H bend, N H stretch Protein, nitrogen 2080 O H stretch, O H deformation Sugar, starch 2010 O=H bend, C O stretch, C O C stretch Starch cellulose 2130 N H stretch Protein 2180 N H bend, C H stretch/C O stretch/C=O stretch/C N stretch Protein, nitrogen 2240 C H stretch Protein 2250 O H stretch, C H deformation Starch 2270 C H stretch/O H stretch, CH 2 bend/CH 2 stretch Cellulose, sugar, starch 2280 C H stretch/ CH 2 deformation Cellulose, starch 2300 N H stretch, C=O stretch, C H bend Protein, nitrogen 2310 C H bend Oil 2320 C H stretch/ CH 2 deformation Starch 2340 C H stretch/ O H deformation/ C H deformation / O H stretch Cellulose 2350 CH 2 bend, C H deformation Cellulose, protein, nitrogen

PAGE 27

27 Advance Techniques for Detecting Plant D isease Plants can exhibit a host of symptoms reflecting various disorders that can adversely influence their health, vigor and productivity to varying degrees. Identifying disease symptoms is essential as inappropriate actions may sometimes prove to be costly and detrimental to the yield. Proper disease control actions or remedial measures can be undertaken if the symptoms are identified early. Sankaran et al. (2010) categorized disease detection techniques into direct and indirect method. Direct methods include s erological methods (Enzyme linked immunosorbent assay (ELISA), immunofluorescence (IF) and flow cyrometry) and molecular methods (Polymerase chain reaction (PCR), DNA arrays). Biomarker based disease detection (Gaseous metabolite profiling, plant metabolit e profiling) and plant properties/stress based detection (imaging techniques, spectroscopic methods) can be classified under indirect method. In the current research, techniques related to plant properties/stress detection are relevant. Therefore, these te chniques will be discussed in detail. Spectroscopic and I maging T echniques In the recent years, spectroscopic and imaging techniques are very popular in plant diseases detection ( Graeff et al., 2006; Huang and Apan, 2006; Moshou et al. 2006), and food quality control (Sundaram et. al., 2009; Sighicelli et al. 2009; Shackelford et al. 2004). In the modern era, plant disease detection sensor should be rapid, disease specific and sensitive to the initial stages of disease infection (Lopez et al., 2003). The spectroscopic and imaging techniques are non destructive, fast and inexpensive.

PAGE 28

28 The spectroscopic and imaging sensors could be integrated with an aero plane, micro copter or agricultural vehicle (manual or autonomous) that can m onitor plant health and detect anomaly in early stage to control the spread of plant disease. V isible Near I nfrared Spectroscopy Infrared spectroscopy is a p recise, fast developing and non destructive technology with increasingly wide range of applicatio ns. The scope of infrared spectroscopy in detection of plant diseases is very high. Guo et al. (2009) carried out experiments showing the potential of NIR spectroscopy as a tool to classify plant species. They used princ ipal component analysis (PCA) combined with the establishment of a Mahalanobis distance of plants leaves to create the discrimination model of leaves and obtained 100% classification. Ramon et al. (2002) used NIR refl ectance using a spectrograph and a neural network classifier to discriminate between weeds and the crop for accurate delivery of the herbicide. Polischuk et al. (1997) made an early detection of Tomato mosaic vi rus in Nicotiana debneyi plants using spectral reflectance measurements in the visible and near infrared. Thus, the spectral reflectance between healthy and infected leaves can be used to diagnose plant diseases before visible changes can be observed. Dise ases can influence spectral properties of plants at many wavelengths, making different wavebands apt for disease detection (West et al. 2003) Kobayashi et al. (2000) used mu ltispectral spectro radiometer and airborne multispectral scanner to detect panicle blast in rice. Airborne multispectral scanner consists four bands of 400 460 nm, 490 530 nm or 530 570 nm, 650 700 nm and 950 1100 nm spectral range. Its instantaneous field of view (IFOV) was 2.5 mrad and ground resolution was 0.94 m at an altitude of 300 m. They measured ground reflectance data with multi spectroradiometer (MSR 7000; Opto Research Corp., Tokyo)

PAGE 29

29 with spectral range of 400 to 2000 nm. They concluded reflectan ce ratios (R 470 /R 570 R 520 /R 675 and R 570 /R 675 ) decreased significantly as incidence of panicle blast increased at dough stage. They reported ground base sensor data and airborne multispectral scanner are effective in panicle blast detection. Zhang et al. (2008) predicted total phenolics, flavonoid contents and antioxidant capacity of rice grain using NIR spectroscopy. They u tilized PLS and modified partial least squares (mPLS) and reported standard errors of prediction were 47.1 and 45.9 nm gallic acid equivalent (GAE) for phenolic content and the coefficient of determination ( R 2 ) were 0.849 and 0.864 by PLS and mPLS, respectively. The feasibility of NIR spectroscopy (1100 2500 nm) to identify waxy wheat was done by Delwiche and Grayboscht (2002) in the lab. They applied linear and quadratic discriminant functions of the scores from principal component and demonstrated near perfect separation of fully waxy wheat from non waxy w heat. Sundaram and Kandala (2009) reviewed the application of NIR spectroscopy to peanut grading and quality analysis. They concluded that NIR spectroscopy could be used for measuring protein, moisture, oil content and fatty acid composition in oil seeds. Wang et al. (2002) classified damaged soybean seeds using NIR spectroscopy. They measured reflectance spectra (log(1/R)) from 400 to 1700 nm. Partial least squares (PLS) and arti ficial neural network (ANN) models were developed to classify normal and damaged seeds. They concluded ANN yields higher accuracy than PLS models. Gomez et al. (2006) concluded that by using the VIS NIR measurement technique in the full spectral range (400 2350 nm), it is possible to assess the quality characteristics of mandarin.

PAGE 30

30 VIS NIR reflectance spectroscopy (325 1075 nm) was app lied to the early detection of Botrytis cinerea disease in eggplant leaves before s ymptoms appear (Wu et al. 2008) PCA was used for dimension reduction. Based on the PCs back propagation neural network (BP NN) model was developed. Furthermore, PLS was executed to identify seven potential wavelengths. BP NN model was also developed with these wavelengths. They indicated that it is possible to apply spectral technology to the early detection of Botrytis cinerea on eggplant leaves. Roggo et al. (2003) compared various classification method accuracies using were studied ; disease resistance, geographical origins and crop periods. NIR spectroscopy data were compared by eight classifi cation method ; Linear Discriminant Analysis (LDA), K Nearest Neighbors (KNN) method, Soft Independent Modeling of Class Analogy (SIMCA), Discriminant Partial Least Squares (DPLS), Procrustes Discriminant Analysis (PDA), Classification And Regression Tree ( CART), Probabilistic Neural Network (PNN) and Learning Vector Quantization (LVQ). They reported SIMCA, DPLS and PDA have the highest classification accuracy. LDA and KNN were not significantly different. Dobrowski et al. ( 2005) demonstrated that the simple reflectance indices calculated in the red edge spectral region can track temperature and water induced changes in fluorescence. NIR spectros copy was applied to predict pre visual decline in eastern hemlock trees (Pontius et al. 2005) An ASD FieldSpec spectroradiometer (350 2500 nm) was used to collect spectral data. PLS and reduced stepwise regression techniques with various vegetat ion indices ( Carter miller index, derivative chlorophyll index, NDVI, RVI )

PAGE 31

31 were used in regression analysis. Their results demonstrated that NIR spectroscopy can detect hemlock decline, before visual symptoms are visible to naked eyes. Hyperspectral and Mu ltispectral I maging Lu (2003) studied the detection of bruises in apples using near infrared hyperspectral imaging. He reported that the spectral region between 1000 nm to 1340 nm was most appropriate for bruise detection The author utilized principal component and minimum noise fraction (MNF) transform ation and was able to detect new and old bruises with correct detection rate from 62% to 88% for red delicious and 59 to 94% for golden delicious. Kim et al. (2001) designed and developed a hyperspectral imaging system which is capable to capture reflectance and fluorescence image in the 430 to 930 nm with 1 mm spatial resolution. The adaptability of the hyperspectral imaging system was demonstrated with sample fluorescence and reflectance images of a normal apple and an apple with fungal contamination and bruised spots. Hyperspectral imaging within the wavelength range of 400 1000 nm was used to detect bruises in Jonagold apple (Xing and Baerdemaeker 2005) The authors utilized PCA and report ed the classification accuracy for sound apples between 77.5 to 84.6% for the one day old bruises. Based on the hyperspectral imaging Xing et al. (2005) stated that wavebands centered at 558, 678, 728 and 892 nm have potential to detect bruises in apples. A hyperspectral NIR imaging system (900 1700 nm) was developed to identify bitter pit lesions on apples (Nicolai et al. 2006) Their system was able to identify bitter pit lesions, even when symptoms we re not visible to the naked eye though their system was failed to discriminate bitter pit lesions and corky tissue. Elmasry et al. (2008) also worked in detection of apple bruises on different background colors using hyperspectral imaging and successfully distinguished from the sound apples. They used PLS method

PAGE 32

32 and stepwise discrimination analysi s for dimension reduction and indentify critical wavelengths. They reported three wavelengths in NIR region (750, 820 and 960 nm) are critical for bruise detection in apple. Moshou et al. (2006) tried to detect pla nt stress caused by disease infestation and to discriminate it from nutrient deficiency stress in field conditions using hyperspectral imaging. They compared yellow rust infected winter wheat plants from the nutrient stressed and healthy plants. They utili zed self organizing maps (SOM) and quadratic discriminant analysis (QDA) The authors demonstrated successfully detection of yellow rust from the nutrient stressed and healthy plants. Mahesh et al. ( 2008 ) differentiated d ifferent wheat varieties by NIR hyperspectral imaging (960 to 1700 nm) ) Seventy five relative reflectance intensities were identified from the images and used for differentiation of wheat classes using statistical classifier (LDA & QDA) and ANN cla ssifier. They reported above 90% classification accuracy with statistical classifier and ANN classifier. Early detection of grapevine phylloxera disease was investigated in Australia (Costa et al. 2007) Their finding showed that at the leaf level, hyperspectral spectroscopy (650 1200 nm) can differentiate phylloxera infested vines. However at the canopy level differentiation is challenging with water deficiency and nitrogen deficiency. Larsolle and Muhammed (2007) measured crop status using multivariate analysis of hyperspectral field reflectance (3 60 900 nm). They analyzed their data in two step: a preprocessing step where data was normalized and a classification steps for estim ating the crop variable. They demonstrated that hyperspectral analysis method can be used to extract spectral signatures of disease severity and plant density.

PAGE 33

33 Late blight caused by the fungal pathogen ( Phythphthora infestan s ) in tomato was successfully i dentified in field conditions (Zhang et al. 2005) They used ADAR (a irborne d ata acquisition and registration, using system 5500 airborne sensor from positive system, Inc. Idaho, USA) broadband system to acquire multi spectral image of four broad bands (blue, green, red and NIR). Various vegetati on indices (combination of red and NIR bands) were used to discriminate late blight. Sour skin disease in v idalia sweet onions was detected by NIR Imaging (Wang et al. 2009) They observed the significant change in mean reflectance spectra in the region 1150 to 1280 nm when the onion was stored 3 days after inoculation. Blasco et al. (2007) evaluated NI R, ultraviolet (UV) and fluorescence techniques to identify the most common defects in citrus. With the NIR system anthracnose and sooty mould were detected. Using UV system only stem end injury was detected, while fluorescence images were able to detect d amages caused by green mould, scarring or thrips. The infecti on due to two plant pathogens ( Phytophthora citrophthora and Penicillium italicum ) in orange using laser induced fluorescence (LIF) and hyperspectral imaging (400 800nm) was studied by Sighicelli et al. (2009) .They observed band sensitivity temporally as infection increases. They reported that both techniques are promising. Application of S pectroscopy and I maging in Citrus Earlier Gaffney (1972) worked on the spectral characteristics of citrus. He reported that a wavelength band of 580 nm to 610 nm is suitable for sorting out defects in pineapple oranges. For v alencia oranges, a wavelength band of 570 to 600 nm could possibly be used to detect the defects.

PAGE 34

34 observed between canopy reflectance and the severity due to Rhizoctonia blight or gray leaf spot (Green et al., 1998). Citrus greening (Huanglongbing or HLB) disease was tried to detect by Lee et al. ( 2008) using aerial hyperspectral imaging. An aerial hyperspectral images were collected from HLB infected groves having spectral range 400 to 1000 nm in 128 different spectral bands with 5 nm spectral resolution and 0.7 m spatial resolution. Spectral angle mapping (SAM) and spectral feature fitting (SFF) classification techniques were used in ENVI software. Due to much variability of healthy and HLB infected tree canopies with geo referencing error these classification te chniques did not yield good results.

PAGE 35

35 CHAPTER 2 REVIEWS ON HLB AND O THER SIMILAR DISEASE AND DEFICIENCIES Citrus is one of the most important agricultural products in Florida as it is the largest citrus producing state in United States and second largest in the entire world. Citrus production being a multi million dollar industry accounts significantly for Flo agricultural economy. But recently, it has been threatened by Huanglongbing (HLB), a devastating and rapidly spreading disease of citrus. This chapter reviews various citrus disease s and nutrient deficiencies that may be confused with HLB. This chap ter also discusses various micro activities during chlorosis in senescence, toxicity, deficiencies and diseases Chlorosis In chlorosis, leaves fail to synthesize sufficient chlorophyll resulting in pale, yellow or yellow white appearance of leaves. Chlor osis may occur due to nutrient deficiency, disease infection, poor drainage, damaged or compacted roots, high alkalinity, or excess use of fertilizers. Low level of nutrients in the soil or their unavailability for reasons like injured roots or high pH can cause nutrient deficiencies in plants. Chlorosis hinders carbohydrates synthesis through photosynthesis in plant that may lead to plant death unless it is treated for the cause of its chlorophyll insufficiency. The following leaf abnormalities have been a ssociated with certain chemicals. Perchlorate C hlorosis This chlorosis is induced by impurities in chilean nitrate of soda potash. The yellow tipping of citrus leaves has been observed in Florida for many years and was once thought to be associated with e xcess boron in certain fertilizers. The symptoms first

PAGE 36

36 develop at the leaf tip and may be confused with boron toxicity, but careful examination reveals that the yellow tipping that is, the chlorotic areas, do not blend with adjoining green tissue.The trans ition from green to chlorotic areas is sharp and abrupt, producing a patchy appearance; whereas the boron toxicity pattern shows a gradual change from green to chlorotic areas, producing a blend of colors. The yellow tipped pattern has been experimentally p roven to be due the perchlorate impurities in chilean nitrate of soda potash ( Stewart et al. 1952). Biuret T oxicity The chlorosis has been experimentally shown by Oberbacker (1954) to be due to biuret impurities in commercial grades of urea. The symptoms first develop at the leaf tips and margins, and the early stages may be confused with the early stages of perchlorate chlorosis. The color of biuret chlorosis is yellow compared to an orange color with the perchlorate chlorosis ( Figure 2 1) The advanced cases of biuret chlorosis may show a burning effect which is more severe on immature than mature leaves. The yellow color of the biuret chlorosis is similar to that of boron toxicity, but the biuret colors are somewhat patchy and free from guming on the un der s urface

PAGE 37

37 Figure 2 1. Biuret chlorosis in citrus leaves (Courtesy: Steve Futch, CREC) Arsenic T oxicity The chlorotic leaf patterns frequently observed after arsenic sprays is commonly known as arsenic toxicity. The degree of chlorosis is usually in the order of the amount of arsenic applied. The symptoms show a loss of chlorophyll without any distinct pattern except that of chlorosis. Arsenic deficiency may be confused with manganese deficiency pattern but close examination indicates that the chlorosis due to arsenic extends across the veins whereas the chlorosis from the manganese deficiency is interveined Fluorine T oxicity The early stages of the toxicity are somewhat similar to boron toxicity but the under surface of the leaves show no resinous excretion with the fluorine toxicity whereas it is generally true for boron (Figure 2 12). The fluorine toxicity shows considerable blends of different shades of green which may be confused with

PAGE 38

38 manganese deficiency. No fluoride problems have been reported where liberal amounts of finely ground rock phosphate carrying high levels of fluorine have been applied to the soil. This indicat es that insoluble fluorides added to the soil are not injurious. Mechanism of Chlorosis Achor and Albrigo (2005) reported that the severe biuret induced chlorosis affected leaves have very few chloroplasts and their average size were about one fifth of tha t in the normal healthy plants. They found amount of cytoplasm was reduced even more, with the central vacuole filling 60% of the viewed surface area of the palisade cells. The plastids looked more like chloroplasts with no grana or other internal membrane s and large numbers of plastoglobuli (lipid bodies). F ollowing are the general changes s that occur during leaf senescence. Ultra structural Changes in Chloroplasts during Senescence The earliest and most striking anatomical changes associated with leaf senescence occur in chloroplasts (Woolhouse, 1984). These organelles undergo ordered sequential changes of their photosynthetic capability from maturity through the process of senescenc e. Autonomous Degradation of Chloroplasts Two models have been proposed to explain photosynthetic activity during senescence (Gepstein, 1988). First hypothesis assumes that the chloroplasts number per mesophyll cell decline during senescence. The other h ypothesis is that the autonomous and sequential degradation of the individual chloroplasts constituents leads to the decline in photosynthetic activity. Achor and Albrigo (2005) found both situations in natural senescence and biuret chlorosis. They found f ewer and smaller chloroplasts

PAGE 39

39 in the mesophyll cell and gradual breakdown and release into the vacuole or cytoplasm of the internal constituents of the chloroplasts. Chlorophyll Degradation Disappearance of chlorophyll is one of the most important processe s of senescence, and eventually the rate of chlorophyll degradation is usually considered to be reliable criteria of leaf senescence and a measure of the age related deterioration of the photosynthetic capacity (Thomas and Stoddart, 1980). Changes in Lipid s during Chloroplasts Senescence The striking ultrastructural changes of thylakoids and the concominant rise in the size and number of plastoglobuli during senescencesuggest that fundamental changes occur in the chloroplasts membranes during senescence. Th e lipids in other membranes show both quantitive and qualitive changes with the advance s enescence (Thompson, 1987). Changes in Stromal Enzymes during Leaf Senescence RuBPCase, the enzyme of the photosynthetic carbon reduction cycle constitutes 50% or mor e of total soluble leaf protein. Due to the loss of photosynthetic activity during senescence, activity of RuBPCase also decrease (Woolhouse 1984). Changes in the Components of the Chloroplast Thylakoid Membranes D uring Foliar Senescence The light harvesting and energy transducing functions of the chloroplasts are now belived to be associated with five main protein complexes in the inner membranes of the chloroplasts (Anderson and Anderson, 1982).

PAGE 40

40 Chloroplasts Protein Degradation Net loss of both thylakoid and stomal proteins during senescense is the result of balance between two opposite processes i.e. synthesis and degradation. Leaf Conductance and CO 2 Assimilation in Senescense Leaves Stomata are main entryways for CO 2 from the atmosphere to th e mesophyll cells, where CO 2 assimilation takes place. It was found that leaf diffusion conductance decreases with the progress of senescence. Insufficient CO 2 supply as a result of reduction in leaf conductance may account for the decreased rates of assim ilation, especially when leaves are exposed to high irradiation or stress. In brief during chloroplasts senescence changes in the molecular organization of the thylakoids, differential and sequential changes in the main protein complexes of th y lakoid, ch anges in the activities of key enzymes in the Calvin cycle an d changes in the rates of protein synthesis and/or degradation o f certain chloroplasts proteins takes place (Gepstein, 1988). Senescence process observed in citrus was slight ly different. There i s a loss of plastoglobuli by their liberation in association with membrane vesicles or direct ly through the double membrane in place of the build up and loss of plastoglobuli at the last step when the membrane dissipates Matile (1992) reported that the pl astoglobuli are the final depository of thylakoidal lipids while Wittenbach (1982) stated that the vacuole and cytoplasm may be the final depository within the cell. These bodies were observed in both the cytoplasm, associated with membranes and in the vac uoles. Acher and Albrigo (2005) also found that the chloroplasts lost their store of plastoglobuli and internal membranes in biuret chlorosis and in senescence whereas in Zn deficient leaves, plastglobuli and internal membranes both were retained. Therefor e,

PAGE 41

41 they concluded that biuret chlorosis is more similar to chlorosis due to senescence than chlorosis due to nutrient deficiency in citrus. Chlorosis Due to Nutrient Deficiency Iron The most obvious effect of iron deficiency is that it produces a marked decrease in the amount of green pigments (Abadia, 1986). Total carotenoids are also decreased by iron deficiency, but to a lesser extent than chloroplasts (Terry, 1980). The characteristic yellow color of chlorotic leaves is a consequence of this relative enrichment in carotenoids ( Figure 2 2) Figure 2 2 Iron deficiency in Orange leaves (Courtesy: Steve Futch, CREC) Mildly affected plants become unsightly and grow poorly. Severe ly affected plants fail to flower or fruit and may even die from lack of iron. Iron chlorosis may occur as a result of one or a combination of causes. The condition is often due to high pH, which makes it possible for other elements to interfere with the absorption of iron, rather than

PAGE 42

42 lack of iron in the soil. This occurs in neutral to alkaline soils when the pH is above 6.5. If overwatering or poor drainage are possible causes, they should be corrected. Iron chelates are water soluble forms of iron that remain in the solution once added to the tree. Some formulations of iron c helates can be applied to the foliage; however, this approach is usually not as permanent as soil applications. Follow the manufacturer's recommendations for amount of use. Some fertilizers contain iron chelates, and use of these with plants susceptible to iron deficiency is recommended. Nitrogen A deficiency of nitrogen in citrus is first characterized by a uniform loss of chlorophyll over the entire leaf, with occasional vein chlorosis in early stages ( Figure 2 3) The symptoms range from a pale yellow ish green color in early stages, to old ivory color in the advanced stages. The deficiency extends over the entire plant, with the greatest severity on fruiting branches, the leaves of which may show a slight mottling effect in acute cases. Severely affect ed trees show stunt ing sparse foliage

PAGE 43

43 Figure 2 3 Nitrogen deficiency in citrus leaves (Courtesy: Mongi Zekri, LaBelle) Calcium Calcium deficiency in citrus fruit is shown in Figure 2 4. Calcium deficiency symptoms are characterized by a marked stunted and hard condition of the tree, with small leaves. In severe cases the leaves become chlorotic at the margins and tips, which progress towards the leaf center and base. The calcium deficienc y pattern may be confused with an advanced case of biuret toxicity. The differences consist of smaller leaves with calcium deficiency, and the chlorosis following the leaf margins, whereas the biurate toxicity is somewhat patchy in early stages, beginning in the tip of the leaf and spreading inward. The tips of calcium deficient leaves are often blunt and sometimes under developed

PAGE 44

44 Figure 2 4 Calcium deficiency (Courtesy: Mongi Zekri, LaBelle) Manganese The symptoms of manganese deficiency in citrus ar e usually less distinct than those of magnesium and zinc. The symptoms occur on the both young and mature leaves, without affecting leaf size, whereas zinc deficiency has a marked reduction on size of leaves and magnesium deficiency pattern is characterize d by green veins with light green background and may be confused with iron deficiencies ( Figure 2 5) As the leaves become mature, the leaf develops pattern with bands of green along the main and lateral veins with light green tissue.

PAGE 45

45 Figure 2 5 Manganese deficiency (Courtesy: Mongi Zekri, LaBelle) Magnesium Magnesium deficiency in citrus is characterized by a type of leaf chlorosis as bronzing ( Figure 2 6) This discoloration or loss of chlorophyll occurs only on mature leaves, and is more preva lent on heavily fruiting trees and bra n ches, and is more noticeable in late summer and fall. In a typical case yellow chlorotic areas develop in the initial stage on each side of the mid rib. Later these areas enlarge often at an angle to the midrib and us ually coalesce to form a yellow zone surrounding a wedge shaped green area at the leaf base. As the deficiency advances, the entire leaf becomes yellow or bronze like.

PAGE 46

46 Figure 2 6 Deficiency symptoms of magnesium in grapefruit (Courtesy: Steve Futch, CREC) Molybdenum The symptom of molybdenum (Mo) deficiency appears first as water soaked areas in the spring flush, later developing into interveinal circular chlorotic areas ( Figure 2 7). It is more noticeable during the summer and early fall months. Moly bdenum deficiency is found on acid sands far more than on heavier and better types. Acid fertilizer aggravates the deficiency, whereas neutral fertilizers and lime usually relieve it. The amount of molybdenum necessary for plant growth, including citrus is very small. The actual amount to correct deficiency ranges from 1 to 2 ounces of sodium molybdate per 100 gallons of spray of equivalent amounts from other soluble sources.

PAGE 47

47 Figure 2 7 Molybdenum deficiency in Orange l eaves Potassium, Phosphorus and S ulfur Potassium and phosphorus deficiencies are shown in Figure 2 8 and 2 9. The chloroplasts from potassium deficient plants have a regular ellipsoidal shape and contain osmiophillic globules. They observed starch grains in almost every chloroplasts In phosphorus deficient maize plants, Hall et al. (1972) observed a regular outline and osmiophillic globules in chloroplasts. An extensive system of grana and stroma lamellae were present. The grana lamellae were organized into irregular grana stacks. T he grana discs within a single stack vary considerably in length and many were longer than the disc seen in the chloroplasts of healthy plants. The most important phenomenon they observed was the absence of starch globules in the phosphorus deficient plan ts.

PAGE 48

48 Figure 2 8 Potassium deficiency in Orange l eaves (Courtesy: Mongi Zekri, LaBelle) Figure 2 9 Phosphorus deficiency (Courtesy: Steve Futch, CREC)

PAGE 49

49 The shape of chloroplasts in sulfur deficient plants was irregular and possessed long projection during advanced stage deficiency. Table 2 1 summaries the effect of mineral deficiencies on chloroplasts structure (Vesk et al. 1965). Figure 2 10 Sulfur deficiency in Orange l eaves (Courtesy: Mongi Zekri, LaBelle)

PAGE 50

50 Table 2 1. Summary of the effects of mineral deficiencies on chloroplasts structure Deficiency Grana Intergranum connections Stroma Star bodies Starch Tomato K Reduced no Long frets, parallel or branching Relative increase Increased no Absent Ca Reduced no, Swelling Extensive parallel frets Swelling Increased amount S Swelling Swelling Present N Reduced no Frets and long lamellae Relative increase Present P Swelling Increased no of long parallel frets, Swelling Present occasionally B No grana Reduced no Relative increase Increased no Absent Zn Reduced no Reduced no Relative increase Starch Cu Swelling Absent Mn Reduced no and size, Swelling Reduced no, replaced by vesicles, Swelling Relative increase Present occasionally Fe Greatly reduced no and size Reduced no replaced by vesicles, Swelling Relative increase Increased amount Spinach Mg Reduced no and size Reduced no of frets Present N Reduced no swelling Plastids smaller in size Reduced no of frets Reduced amount Present Starch P Reduced no and size, Swelling Reduced no of frets, Swelling Increased no Starch B, Zn, Cu No results Mo Coalescence of compartments, swelling Swelling Very dense granular Present granular Present occasionally Mn Reduced no, extremely Swollen Reduced no replaced by vesicles, Swelling Relative increase, formation of tails Absent Fe Absent or greatly reduced in no and size Show tubules and vesicles swelling Relative increase Absent Maize K Normal S Increased no Mg Reduced no and size, Swelling Reduced, branching frets, Swelling Tails Absent Fe Greatly reduced in no and size, may be absent Plastids smaller in size, Extensive parallel lamellae

PAGE 51

51 Zinc The initial stages of the deficiency appear as irregular chlorotic areas in the leaf tissue, between the main and lateral veins (Figure 2 11). The tissue immediately adjoining the veins remains green, while the chlorophyll disappears from the rest of leaf. This results in an irregular, mottled or variegated mixture of vivid green and white to yellow colors. In the early stages of the deficiency, the characteristic leaf pattern may occur on apparently normal sized leaves, but as the deficiency becomes more a cute, the new leaves are small, narrow and pointed, with a greater loss of chlorophyll. Figure 2 1 1 Zinc deficiency in orange leaves (Courtesy: Mongi Zekri, LaBelle) Boron Fruit symptoms most indicative of boron deficiency include dark spots in the white albedo of fruit and sometimes in the central core ( Figure 2 1 2) Boron deficient fruit turns hard and dry due to lumps in the rind so this deficiency is also known as "hard

PAGE 52

52 ave brownish discolorations and unusually thick albedo. Older fruit remain undersized and misshapen with an unusually thick albedo. Seeds do not develop and the terminal growing point of the main stem dies. Other symptoms include slight thickening and down ward curling of leaves. Damp spots are found on young leaves that turn translucent as the leaves mature. Defoliation begins at the top of the tree and continues until tree dies. Figure 2 1 2 Boron deficiency (Courtesy: Mongi Zekri, LaBelle) Borax is commonly used to treat boron deficient citrus. It can be applied to soil or to the foliage. Boric acid is preferred for foliar application as it is more soluble than Borax. Foliar spray application is effective in Florida. The spray may be applied either during the dormant period or post bloom. Unlike other micronutrient deficiencies, boron can impact fruit quality and should therefore not be allowed to occur. Slight excess can

PAGE 53

53 cause toxicity, so maintenance or correctional applications should invo lve ground or foliage applications, but not both. Copper Copper deficiency affects the formation of grains, seeds and fruit much more than it affects vegetable growth (Figure 2 13) The main reason for the poor development of seeds and fruits is that a hig h percentage of the pollen from copper deficient plants is not viable. When extractable copper exceeds 100 pounds per acre, trees may begin to decline Unusually large dark green foliage with a "bowing up" of the midrib are among the primary symptoms for c opper deficiency. Fruit symptoms are most evident on oranges. Fruits bear brown spotted areas of hardened gum on rind and fruit splitting is commonly found on the trees with Cu deficiency. The brown stained areas on the fruit may become almost black over t he time and fruit shed s by summer eventually. L eaf and twig symptoms may not be observ ed when Cu deficiency is present along with Zn or Mg deficiency but the typical fruit symptoms will be evident Therefore, fruit symptoms are considered reliable and cons istent indicator of Cu deficiency Foliar sprays or soil applications of Cu fertilizer can prevent or cure Cu deficiency. Spraying a solution containing 2 to 3 lbs per acre of elemental Cu applied during flowering usual ly results in recovery followed by a normal fruit set.

PAGE 54

54 Figure 2 1 3 Copper deficiency (Courtesy: Mongi Zekri, LaBelle) Diseases in Citrus Alternaria Brown Spot Alternaria brown spot causes serious losses of susceptible tangerine and tangerine hybrid s. A similar leaf and fruit spot affects rough lemon and Rangpur lime. This disease affects young leaves, twigs and fruit, and produces brown to black lesions which vary in size from small dots to large expanding lesions ( Figure 2 1 4). Diseased fruit may a bscise, and lesions on remaining fruit may vary from small spots to larger lesions (Whiteside, 1976).

PAGE 55

55 Figure 2 1 4 Alternaria brown spot in orange fruit (Courtesy: Moongi Zakri LaBelle) Moderate to high temperatures and rainfall favors the disease bu t since heavy dews are sufficient for infection, fruit blemishes occur even in semi arid areas where no rainfall occurs after flowering (Timmer et al. 2000). Minimizing the period of leaf wetness of the tree canopy can reduce disease incidence. Nursery tr ees free of the disease should be used for new plantings, and overhead irrigation should be avoided. Excessive nitrogen fertilization and irrigation that promote abundant growth flushes should be avoided. Foliar fungicide applications are needed in most a ffected orchards, with frequency based on disease severity. Black Spot This disease causes fruit loss and a serious external blemish of citrus fruit. Black spot is widespread in the humid to semi arid citrus growing areas in the southern

PAGE 56

56 hemisphere that h ave summer rainfall. Black spot produces lesions on fruit varying from small brown to black spots to large sunken lesions ( Figure 2 1 5). Symptoms may appear in the orchard on fruit, and cause premature fruit drop, or infections may remain quiescent until h arvest. Figure 2 1 5 Black spot (Courtesy: Michael Rogers, CREC) Infections usually occur from early to mid summer and remain latent for some time. Moisture is essential for infection Fungicide applications are the primary means for managing black spot. Often a single, late summer spray of benomyl will provide sufficient disease control except where resistant strains occur. Canker Citrus canker is a serious bacterial disease in humid tropical and subtropical areas. The disease causes external blemishes on the fruit, making them unsuitable for the fresh market, and may cause fruit drop (Figure 2.16). This disease is widespread in

PAGE 57

57 Asia and is spreading in southern South America and in S outh Florida. Canker affects young leaves, stems and fruit of most citrus species, producing water soaked lesions of variable size. Figure 2 16 Citrus canker (Courtesy: Jamie Yates, CREC) This disease is dependent on storms and windblown rain not only for dispersal, but also to force the bacteria into wounds and stomata. Canker is most serious in areas with severe thunderstorms, hurricanes and typhoons. The presence of leaf miners exacerbates canker because tunnels provide entry points for the bacteriu m and expose additional tissue in which it multiplies. Citrus canker is controlled by quarantine and eradication in countries in which it is absent or has limited distribution. Movement of citrus fruit and budwood from infested areas is restricted (Schuber t et al., 2001). Diseased trees are burned in place, and the area is kept free from citrus root sprouts for

PAGE 58

58 6 12 months. Wind breaks are quite effective in reducing spread of the disease and in limiting the amount of infection. Copper fungicides are effect ive in preventing fruit infection if applied frequently (Stall et al., 1981). Mal Secco This disease is primarily a problem of lemons, but can also affect tangerines and their hybrids. Oranges and grapefruit are seldom affected. Mal secco can result in lo sses of tree limbs or, in severe cases, of the entire tree. Infected leaves develop a veinal chlorosis ( Figure 2 18) As the infection proceeds downward in the vascular system, leaves wilt and the shoots die back. Eventually, limbs or the tree may die. Whe n the bark of affected branch is removed, the wood shows a characteristic orange or orange red discoloration. Disease trees and branches should be removed and burned to reduce inoculums. Foliar sprays of benomyl or copper fungicides in the spring and autum n reduce new infection. Figure 2 1 7 Mal secco

PAGE 59

59 Melanose Melanose is most severe on lemon and grapefruit. It is an important disease of fruit produced for the fresh market in humid subtropical areas, but is not major concern in Mediterranean climates or in high rainfall tropical areas. Melanose appears as raised, brick red to brown pustules on the leaves, twigs and fruit ( Figure 2 19) Spores carried down the side of the fruit by water may cause lesions to form in a tearstain or droplet pattern. Relatively long period of wetting (12 18h) are required for infection even at high temperatures. Copper fungicides are the most widely used means to control melanose (Timmer and Zit k o, 1996) because they are highly effective and have a high residual. However, they must be applied frequently when fruit growth is rapid. Figure 2 1 8 Melanose (Courtesy: Jamie Yates, CREC)

PAGE 60

60 Powdery M ildew Powdery mildew occurs throughout the humid areas of Asia and USA. It reduces yield by debilitating trees and causing fruit drop. Whitish powdery patches of mildew occur on the upper surface of lea ves, especially at the edges and on young fruits ( Figure 2 20) Immature leaves and entire shoots may shrivel and drop, and infected young fruit falls permanently. Figure 2 19 Powdery mildow (Courtesy: Megh Singh, CREC) Scab Scab disease affect only t he external quality of the fruit of susceptible citrus and are important primarily on fruit that are grown for fresh market ( Figure 2 21) Citrus scab affects many mandarins and their hybrids, lemons and grapefruit, and occurs in all areas where conditions are favorable. The first symptoms are clear to slightly pink, water soaked areas on leaves or fruit. These grow rapidly to raise pustules that become warty and grey with age. Lesions on fruit tend to flatten with age, especially on

PAGE 61

61 grapefruit, and lesions of sweet orange scab tend to be flatter than those of citrus scab.Fruit s are susceptible to scab until they reach approximately 3 cm diameter. Fungicide application during this period is effective in controlling the disease. The most effective products in clude the sterol biosynthesis inhibiting fungicides, benomyl, ferbam and copper materials (Timmer and Zitko, 1997) Figure 2 2 0 Scab (Courtesy: Mongi Zekri, LaBelle) Huanglongbing (Greening) Huanglongbing ( HLB) was reported in mainland China in 1919, and in South Africa i n 1937 as citrus greening disease (da Graca, 1991). mmon name in e nglish speaking countries. HLB has destroyed an estimated 60 million trees in Africa and Asia.

PAGE 62

62 The most characteristic symptom of HLB is green patches on the pale green background that often begins in one part of the canopy (Figure 2 26) .prob lems. Leaf yellowing and leaf drop result in twig die back. Fruit on affected trees are small, lopsided and poorly colored, hence the name greening. Juice is bitter, low in soluble solids and high in acid. Nursery trees are stunted, terminal leaves are yel lowed, new leaves are small, leathery and upright and old leaves are mottled. As these symptoms take 4 6 months to appear; symptomless trees may be distributed from affected nurseries. Figure 2 2 1 Huanglongbing (HLB) (Courtesy: Jamie Y ates, CREC ) Schne ider (1968) proposed that when leaves are invaded by the greening virus, necrosis of localized pockets of phloem in the leaf vascular system is the first degenerative change induced. This occurs in mature leaves or in the leaves nearing maturity. Several r eactions to the necrosis may occur in the leaf. During this phase

PAGE 63

63 to a thin film that encloses them. Granules occur in the leaf that gives a leathery feel. It is assumed th at grana and chlorophyll of the chloroplasts may be destroyed by being stretched, due to enlargement of the granule within and by crowding chloroplasts due to enlargment of granules. It was assumed that weak chlorotic growth is not a direct result of virus activity, but that is results from the disturbed state of the old shoot from which new shoots grow. Mineral, nutrients and various organic compounds in old leaves moves to newly forming shoots where they support growth. Transmission of HLB occurs by gra fting and by African citrus psyllid, Trioza erytreae and the Asian psyllid, Diaphorina citri Each psyllid is able to transmit L. africanus and L. asiaticus Polymerase Chain Reaction (PCR) and DNA RNA hybridization techniques can now be used to detect the two species Leprosis Leprosis causes chlorotic to necrotic areas on the fruit, leaves and twigs of susceptible cultivars. Initial symptoms are chlorotic lesions that often become necrotic and gum impregnated and show concentric patterns ( Figure 2 23) A chlorotic zone around the lesion may still remain. Leaf and fruit drop occurs when infections are abundant.

PAGE 64

6 4 F igure 2 2 2 Leprosis (Courtesy: Ron Brlansky, CREC) Infections are localized and apparently associated with feeding activity of mites t hat carry the causal virus. The virus does not infect citrus systemically and trees do not develop symptoms on new growth after infective mites are removed. C itrus Variegated Chlorosis The trees affected with the citrus variegated chlorosis (CVC) have mottled leaves on one or more branches, and in chronic stage may be stunted and show twig dieback (Figure 2 2 4 ). Fruits are small and hard and change color prematurely. They are frequently sun burned and may also have sunken brown areas on the surface of the rind.

PAGE 65

65 Figure 2 2 3 Citrus variegated Chlorosis (Courtesy: Mongi Zekri, LaBelle) The disease is caused by a strain of the bacterium Xylella fastidiosa that inhabits xylem and impairs its normal function. It may spread by infected bud wood or by leaf hopper vectors. Control measures include avoiding propagation of CVC infected bud wood for new plantings, removing infected limbs from recently affected trees and removal of affected trees in young plantings. Mandarins, grapefruit and lemons appea r to be less sensitive to CVC than sweet orange and are more susceptible for areas that are severely affected by CVC.

PAGE 66

66 CHAPTER 3 SPECTRAL C HARACTERISTICS OF C ITRUS G REENING (HUANGLONGBING) I ntroduction Citrus is one of the most important agricultura l products in Florida as it is the largest citrus producing state in United States and second largest in the entire world. Citrus production being a multi agricultural economy. But recently, it h as been threatened by Huanglongbing (HLB), a devastating and rapidly spreading disease of citrus. Spectral reflectance characteristics of leaves have been shown to be highly correlated with their chemical composition. Carter and Knapp (2001) showed the im portance of chlorophyll concentration on the spectral signature of leaves. The optical response to stress near 700 nm, as well as corresponding changes in reflectance that occur in the green yellow spectrum (400 500 nm), was explained by the general tenden cy of stress to reduce leaf chlorophyll concentration. The reflection of incident radiation from within the leaf interior of stressed trees increases such that stressed trees appear brighter in the visible region of the spectrum than healthy trees (Cibula and Carter, 1992). Riedell and Blackmer (1999) found that leaf reflectance in the 625 635 nm and the 680 695 nm wavebands, together with the Normalized total Pigment Chlorophyll Index (NPCI) were significantly correlated with the total chlorophyll concent rations in both green bug and Russian wheat aphid damaged trees. Boochs et al. (1990) suggested that high resolution reflectance spectra, especially in the red edge area (reflectance between 680 760 nm), would be useful for the identification of small diff erences in the chemical and morphological status of the trees in the field. Optimal

PAGE 67

67 reflectance at wavebands of 825 nm and 980 nm were determined using stepwise linear discriminate analysis to detect bruises in strawberries (Tallada et al., 2006). Borel an d Gerstl (1994) pointed out that canopy architecture strongly influences illuminated areas for different sun angles, and thus reflectance. This could affect the spectral signature of trees in the field. The use of first, second and higher orders derivativ es have become an established technique for reduction of low frequency background noise and for resolution of overlapping spectra (Butler and Hopkins, 1970). In remote sensing, mostly the first derivative has been used to facilitate the location of critica edge. Hence, derivative analysis may have the potential to discriminate HLB infected trees. Objective This research was aimed at dev eloping a spectral method for the detection of HLB. The specific objective was to identify optimal wavebands (400 nm to 2500 nm) for accurate detection of HLB in citrus. Materials and Methods The study was conducted at a commercial grove in Lake Placid, FL (approx. 27.34386 N and 81.38387 W). Twenty infected leaf samples and 20 non infected leaf samples were collected. Canopy reflectance spectral data was collected with an ASD FieldSpec spectroradiometer ( FieldSpec UV/VNIR, Analytical Spectral Devices, Boulder, CO). This spectroradiometer is a compact and field portable with a spectral range of 350 2500 nm and a rapid data collection time of 0.1 second per spectrum. All data files were

PAGE 68

68 collected in an ASD file format that can then be viewed and post pro cessed in 1000 nm and 2 .0 nm for the spectral region of 1000 2500 nm. Bare fiber optic cable was used in data collection. The integration time was optimized using optimiz e options within the software Dark and white calibrations were conducted prior to data collection in the field. The data were collected between 11:00 a.m. to 12:00 noon to limit the variability due to change in sun angle. Every 10 minutes a reference reading was c ollected to reduce the error due to atmosphere. Bare fiber optic was kept about 50 to 80 cm from the tree canopy. During data collection, the area scanned by the ASD spectroradiomter was approximately 385 to 988 cm 2 During r aw spectral data process ing in ViewSpecPro (Analytical Spectral Devices, Boulder, CO) outliers were identified by two criteria. First, s pectra from one variety of tree s should only contain random errors. Secondly, i f a spectrum had a different shape or curve it was removed from the da ta set. Since the sensor was view ing the tree canopy nearby parallel to ground any background effect due to soil was negligible. Data Analysis Discriminability The discriminability of two probability density functions (pdf) with the same standard distribu tion is defined by (Duda et al., 2000) as: = 2 1 ( 1 ) Where, d = discriminability, = standard deviation, and

PAGE 69

69 2 and 1 = mean spectra of HLB affected trees and healthy tr ees, respectively. In this case, the standard deviations could be different between the two pdfs at the same wavelength, thus the standard formula could not be used. Equation 2 displays an example of two pdfs with different standard deviations from the sam ple data used. For this reason following equations was used. d = ( 2 1 ) ( 1 + 2 ) 2 ( 2 ) Discriminability may be one method by which can determine optimal wavelengths to discriminate HLB trees with healthy trees. Averaging the standard deviations of both pdfs allow the discriminability to scale with magnitude changes in the standard deviation. For better discriminability of HLB should be large. Since reflectance properties at one wavelength shares common properties with neighboring wavelengths, a second wavelength was required to be outside a threshol d of 100 nm, allowing two distinct features (Kane and Lee 2006). Data analysis was performed with Statistical Analytical Software (SAS). performs a stepwise discriminant an alysis to select a subset of the quantitative variables for use in discriminating categories. Spectral Derivative Analysis Among the techniques that have been developed in spectroscopy, derivative analysis is particularly promising for use with remote sen sing data. Demetriades Shah et al. (1990) showed that derivative analysis was better than ratio or difference vegetati on indices. Spectral derivative analysis was used to examine the spectral

PAGE 70

70 differences more closely in reflectance at specific wavelengths. A derivative analysis separated the differences more clearly (Rundquist et al., 1996). The simplest numerical method for generating derivatives is to divide the difference between successive spectral values by the wavelength interval separating them (Dem etriades Shah et al., 1990). This provides an approximation of the first derivative at the central point between the values whose difference is used to calculate the slope. The first order derivative provides information on the rate of change in reflectanc e, which is the slope, with respect to wavelength, while a second order derivative gives the change in slope with respect to wavelength. Spectral R atio A nalysis Spectral ratio analysis was used to identify the wavelengths that are sensitive to tree stress caused by HLB infection. Spectral derivative analysis magnifies the differences in spectral reflectance. By calculating a spectral ratio of healthy trees with HLB infected trees, the effect of noise can be reduced (Zhang et al. 2002). If the ratio of these spectra is close to 1, it means there is no significant difference between the healthy and HLB trees at particular wavelengths. The more deviation of the ratio from one, the more likely separation is possible at particular wavelengths. Taking the mean spe ctrum of HLB infected trees as the numerator and mean spectrum health tree as the denominator, the spectral ratio (S ratio ) was calculated. S ratio = Mean spectrum of HLB plant Mean spectrum of healthy plant ( 3 )

PAGE 71

71 Results an d Discussion Discriminability The discriminabil i ty d of all samples is present ed in T able 1. At the wave lengths of 695 to 705 nm, a discriminability of 0.86 was obtained. It seems that the visible region (400 700 nm) has good (0.89 to 0.85) discrimination. The wide range suggests a great amount of inconsistency with samples. This could indicate a need for more samples or a slight change in reflectance during data collectio n. Table 3 1 Discriminability of wavelengths for HLB and healthy trees Wavelengths (nm) Discriminability ( d ) Wavelength (nm) Discriminability ( ) 695 705 0.86 695 705 0.89 585 595 0.83 585 595 0.85 405 415 0.78 405 415 0.78 2345 2350 0.71 2345 2350 0.69 1980 1990 0.68 1980 1990 0.68

PAGE 72

72 F igure 3 1 A sample spectra of a healthy and HLB infected tree canopy Spectral D erivative A nalysis To examine the differences in spectral reflectance, derivative analysis was used. The rate of change in reflectance in the first derivative, within a 2 .0 nm range was distinctly different for healthy and HLB infected trees Likewise, the second derivatives are also different for healthy and HLB infected trees Derivative analysis was performed by using the finite diffe rence method and the S avitzky G olay method. The results of both the first and second derivative analysis reveal spectral ranges where the response of HLB infected trees has the opposite sign compared to healthy trees 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 350 500 650 800 950 1100 1250 1400 Spectral Reflectance Wavelength (nm) Healthy HLB

PAGE 73

73 The derivatives computed by the finit e difference method and the Savitzky Golay were noisy. The first derivative spectra seem to be less a function of noise than second derivative spectra. The f inite difference method seems have more potential to differentiate HLB infected trees over the Savi tzky Golay method SAS output gave a large range of wavelengths where first derivative spectra and second derivative spectra reveals good separation of HLB infected trees with healthy trees (Table 3 2 ). The f inite difference method computed 747 nm, 1041 n m, 1283 nm 1601 nm and 2283 nm in first derivative. Results from the f inite difference second derivative method revealed that wavelength s of 480 nm, 590 nm, 754 nm, 1041 nm, and 2071 nm have the potential to differentiate HLB. The Savitzky Golay method gave similar results as the finite difference method (Table 3 2) Spectral R atio A nalysis An example of spectral ratio was illustrated in Figure 3 2. Large magnitude differences among spectral ratios can be observed for the wavelength range of 400 nm to 23 50 nm. The results of ratio analysis showed the wavelengths that are most sensitive to HLB and can be better utilized for discriminating healthy and HLB infected trees (Figure 3 2). Table 3 2. Identified wavelengths for separating HLB trees from healthy trees Finite difference method Savitzky Golay method First derivative Second derivative First derivative Second derivative 747 nm 590 nm 747 nm 653 nm 1041 nm 754 nm 1671 nm 754 nm 2283 nm 1041 nm 2014 nm 1039 nm 1283 nm 480 nm 1010 nm 487 nm 1601 nm 2071 nm 487 nm 2073 nm

PAGE 74

74 The spectral ratios of 1.3 to 1.5 correspond to 530 to 564 nm. These wavelengths correspond to the green peak. The HLB infected trees are less green than healthy trees. Reflectance of HLB infected trees at 530 564 nm was higher than that of healthy trees. Therefore, the ratio is higher in this range. A second sensitive point was observed at 710 to 715 nm (red edge). In this range, the ratio was 1.3 to 1.6. This r ange is sensitive due to the chlorophyll absorption. Healthy trees have more chlorophyll; hence, they will absorb more light in this range than HLB infected trees resulting in higher reflectance of HLB infected trees than healthy trees. The wavelengths of 1450 nm and 1990 nm correspond to the water absorption band. Figure 3 2. Spectral ratio of HLB infected and healthy trees This work provides a better understanding of the spectral properties of HLB infection in citrus canopy. The identified wavelengths i n the green region and near infrared sensitive to the change of chlorophyll content and water content in the ratio analysis are consistent with previous work reported by Gitelson and Merzylyak (1997).

PAGE 75

75 The HLB infected trees contain lower chlorophyll which leads to a low photosynthesis rate and lower water content. The changes of these pigments and water content are often indicators of tree stress, which can be used to monitor the conditions of crop growth and site characteristics. Conclusion The practical i mplication of this result is that hyperspectral spectroscopy has the potential to identify HLB infected trees from healthy trees. Further study is necessary to confirm the potential to detect HLB infected trees from healthy trees. Discriminability to separate HLB infected trees from healthy trees was 0 .83 to 0.86 in visible region (695nm to 705 nm and 585nm to 595 nm). Results of second derivative analysis with the finite difference method and the Savizky Golay method were the same. Results of first derivative analysis with these two methods were slight ly different. Higher ratios (1.3 to 1.6) were obtained at green bands (530nm to 564 nm) and red edge region (710nm to 715 nm). Spectral ratio analysis supports the results obtained from the discriminabilty analysis.

PAGE 76

76 CHAPTER 4 SPECTRAL DISCRIMINATION OF HE ALTHY VS. HLB INFECTED CITRUS TREE S IN THE VIS NIR RANGE Introduction Huanglongbing (HLB) or greening is one of the most se rious diseases of citrus billion dollar citrus industry. It was first reported in Flo rida in August 2005 in South Miami Dade County. It affects all citrus cultivars and causes rapid deterioration of trees Ron Muraro (2007) reported that the total production costs of fresh fruit increased from $ 1115.19 to $ 1711.20 per acre in southwest Florida. At present, Polymerase Chain Reaction (PCR) is the only determinant method to detect HLB. For the selection of leaves for PCR testing, an iodine based starch test can be used. HLB infected trees show an increased level of starch accumulation (Sch neider, 1968). Leaves with strong blotchy mottle symptoms of HLB infection stain very dark grey to black along cut surfaces when immersed in iodine solution for two minutes, while healthy citrus leaves show no or very little staining after their immersion in iodine for two minutes (Etxeberria et al., 2007). Currently, there is no cure available for HLB. Groves are scouted regularly and affected trees are removed as soon as possible. Generally scouting is recommended at least four times a year but more frequ ent identification and removal of HLB infected trees would be desirable. Annual field inspection costs for identifying HLB infected trees was reported as $90.92 per acre in 2006 2007 (Muraro, 2007). Near infrared (NIR) spectroscopy has been utilized for fr uit quality assessment and disease detection for many years. Visible and NIR spectroscopy have been largely used to detect plant status such as moisture content, nutrient stress and disease detection. Thomas and Oerther (1972) found a non linear relationsh ip between

PAGE 77

77 reflectance at 550 nm and leaf nitrogen content of sweet pepper leaves wit h a correlation coefficient of 0.93. Hyperspectral spectroscopy is a technique that utilizes hundreds of narrow contiguous spectral bands for the assessment of plant heal th. The data obtained from hyperspectral spectroscopy may detect many plant attributes that were not detectable with multispectral spectroscopy. Spectral properties of potato, bean and barley disease were reported by Lorenzen and Jensen (1989) and Malthus and Madeira (1993). There is very little known about the spectral characteristics of HLB in the visible and NIR regions. Gaffney (1972) worked on the reflectance properties of citrus. He concluded that a wavelength band of 580 nm to 610 nm is suitable for sorting out defects in Pineapple oranges. Hyperspectral imaging was used to detect chill induced damage in whole cucumbers under a variety of conditions (Liu et al., 2005). Band ratio algorithms and principal component analysis (PCA) were attempted to dis criminate the area damaged by chilling injury. They found that a dual band ratio algorithm (R 811nm/R 756 nm) and a PCA model from a narrow spectral region of 733 848 nm can detect chill injured skins with a success rate of over 90%. Zacro Tejada et al. (2 005) showed that the best indicators for chlorophyll content estimation in V. vinifera L. leaves were narrow band hyperspectral indices calculated in the 700 750 nm spectral region, with R 2 ranging from 0.8 to 0.9, with poor performance of traditional indi ces such as the Normalized Difference Vegetation Index (NDVI). Partial least squares (PLS) is a multivariate analysis technique that is commonly used in analyzing the spectral data. It is sometimes referred a s

PAGE 78

78 in contrast with an including lack of multicolinearity among the predictor variables, with well understood relationships to the response variable. PLS balances the two objectives of explaining response variation and predictor variation. Since the focus of PLS is prediction and not explanation lack of well understood relationships of the response to the predictor variable is not a problem. The number of extracted factors depends on the data. Latent vectors, i.e. s uccessive linear combinations of the predictors, explain response variation and predictor variation. Sometimes too many extracted factors can cause over fitting. Discriminate analysis is a technique for classifying a set of observations into predefined cl asses. It is a one way classification based on the known values. The technique is based on how close a set of measurement variables are to the multivariate means of the levels predicted. This technique can also be used to discriminate HLB. This technique i ncludes stepwise selection of variables, choice of linear, quadratic or regularized parameter analyses and a discriminant score to show each point close to a particular group. Therefore, the objectives of this study were as follows: Collect reflectance dat a from HLB and healthy leaves. Investigate the potential of the PLS technique and discriminate analysis in identifying the spectra of HLB infected trees from the spectra of healthy trees. Evaluate the possibility of using the narrow NIR spectral range (400 nm to 900 nm) instead of the full range (400 nm to 2500) in discriminating the spectra of HLB infected citrus leaves from healthy leaves.

PAGE 79

79 Material and Methods Th is study was conducted near Lake Alfred, FL. There were two sets of data collected on June 13 and 14 2007. Two trees were r eported positive with HLB. The canopy reflectances of two neighboring healthy trees were also collected for the comparison. Another two sets of data were collected from a greenhouse on Aug 2 and 3, 2007. The spectra were colle cted in the presence of artificial lights on Aug 2, 2007 and in the presence of natural light on Aug 3, 2007. Healthy and HLB infected plants of eureka lemon, mandarin, madam vinous (MV), and sunchusha were used on Aug 2, 2007 for spectral data collection. Spectra of healthy and HLB infected plants of calamondin, Duncan grapefruit, trifoliate orange, Madam Vinous, Cleopatra mandarin, Mexican Lime, sweet lime, Valencia, Sunchusha and sour orange were collected on Aug 3, 2007. Canopy reflectance was collected with a FieldSpec 3 spectroradiometer manufactured from Analytical Spectral Device (Boulder, CO). This ASD spectroradiometer collects reflectance data from 350 to 2500 nm with rapid data collection rate of 10 scans per second. The ASD spectroradiometer tr ansmits the spectral data wirelessly to a laptop computer. Each individual scan was the result of an average of 10 scans automatically made by the equipment. Bare fiber optic cable with 25 degree field of view was used in data collection. The integration t ime was optimized using optimization options within the software. Dark and white calibrations were conducted prior to and during data collection in the field. The data were collected between 11:00 a.m. to 2:00 pm to limit the variability due to change in s un angle. Spectral reflectance was collected from two sides of the tree to compensate for the effect of shade. On June 14, 2007, data was also collected from the top of the canopy.

PAGE 80

80 Reference readings were collected every 10 minutes to reduce the error due to the atmospheric changes. The bare fiber optic was kept about 50 to 80 cm from the tree canopy. During data collection, the area scanned by the ASD spectroradiomter was approximately 385 to 988 cm 2 Raw spectral data processing was performed in ViewSpecPro (Analytical Spectral Devices, Boulder, CO). The percent reflectance value was obtained in 350 2500 nm range. JMP 7 (Cary, NC) was used for partial least squares modeling. The initial canopy reflectance was obtained from 350 nm to 2500 nm with intervals of 1 nm. Reflectance data was reduced by averaging 50 reflectance values. For example, the reflectance value at 400 nm was calculated by averaging the reflectance values at 375 nm to 424 nm. This reduced the total number of data point from 2152 t o 40. PLS modeling was applied to the full range from 400 nm to 2450 nm and narrow range from 400 nm to 900 nm. There were 113 spectra (49 HLB, 62 healthy) evaluated on June 13, 2007. Out of 49 HLB spectra, 27 spectra were used for calibration and 22 were used for validation. For healthy trees, 32 spectra were used for calibration and 30 spectra were used for validation. A total of 128 spectra (70 HLB, 58 healthy) were collected on June 14, 2007. For HLB trees, 35 spectra were used for calibration and 35 sp ectra were used for validation. In the case of healthy trees, 30 spectra were used for calibration and 28 spectra were used for validation. In JMP 7, specified measurement variables were specified (reflectance at various wavelength) as Y effects and classi fication variables (healthy or HLB) as a single X

PAGE 81

81 effect. The multivariate fitting gives estimation of the means and the covariance matrix for the data, assuming that the covariance is the same for each group. Results and D iscussion Figure 4 1 shows an e xample of spectral differences between healthy and HLB infected tree. The healthy and HLB infected tree canopies were measured from northeast (NE) and southeast (SE) orientations. Each measurement is an average of ten measurements. In the near infrared reg ion, low reflectance was observed. Noise was observed in bands 1350 to 1500, 1750 to 1950 nm and bands after 2350 due to the presence of atmospheric moisture. Figure 4 1. Canopy reflectance of healthy and HLB infected tree with FieldSpec 3 spectroradio meter. Healthy and HLB infected canopies were measured from northeast and southeast direction 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 400 650 900 1150 1400 1650 1900 2150 2400 Reflectance Wavelength (nm) Healthy NE Healthy SE HLB NE HLB SE

PAGE 82

82 The results of PLS calibration and validation for HLB are summarized in Table 4 1 for June 13, 2007. Correct (full range) and correct (narrow range) are the numbe r of spectra classified correctly for healthy and HLB in full NIR range (400 nm to 2450 nm) and narrow NIR range (400 nm to 900 nm), respectively. Table 4 1. PLS modeling for HLB and healthy trees showing total samples, correct classifications (June 13, 2007). Healthy 1 Healthy 2 HLB 1 HLB 2 NE SE NE SE NE SE NE SE Calibration 8 8 8 8 6 5 8 8 Validation 7 7 8 8 4 4 7 7 Total 15 15 16 16 10 9 15 15 Correct (full range) 6 6 8 7 4 4 5 8 Correct (narrow) 6 7 8 8 4 4 5 7 The percentages of correctly classified spectra are shown in Fig 4 2. The PLS model classified 87.1% healthy and 95.5 % of HLB spectra correctly in full range. The PLS model classified 93.5 % healthy and 90.9% HLB spectra correctly in the narrow NIR range. These results support that the narrow NIR range (400 to 900 nm) has almost equal potential to discriminate HLB as the full range (400 to 2450 nm).

PAGE 83

83 Figure 4 2. Percentage of correct classification of HLB and healthy trees in full NIR range (400 2450 nm) and narrow NIR range (400 900 nm) on June 13, 2007 PLS modeling results from June 14, 2007 are shown in Table 4 2. Canopy reflectances were collected from the n ortheast, southeast, and top of the canopy. Figure 4 3 shows the percentage of spectra classified correctly with PLS modeling on June 14, 2007. In the full range, PLS classified 66.7% of healthy and 74.3% of HLB trees, correctly. PLS has classified, 78.6% healthy and 54.3 % HLB spectra correctly in the narrow NIR range. 0 10 20 30 40 50 60 70 80 90 Classification accuracy (%) Full range Narrow range Healthy HLB

PAGE 84

84 Table 4 2 PLS modeling for HLB and healthy trees showing total samples, correct classifications (June 14, 2007) Healthy 1 Healthy 2 HLB 1 HLB 2 NE SE T NE SE T NE SE T NE SE T Calibration 5 5 5 4 5 6 6 5 9 5 5 5 Validation 5 5 5 3 5 5 5 5 10 5 5 5 Total 10 10 10 7 10 11 11 10 19 10 10 10 Correct (full range) 2 3 4 3 3 5 4 3 8 3 3 5 Correct (narrow range) 3 4 4 3 3 5 1 3 6 4 2 3 Figure 4 3. Percentage of correct classification of HLB and healthy trees in full range (400 2450 nm) and narrow NIR range (400 900 nm) on June 14, 2007. PLS modeling for Eureka lemon, mandarin, Madam Vinous and Sunchusha citrus types are give n in Table 4 3. Figure 4 4 shows the percentage of spectra classified correctly with PLS modeling on Aug 2, 2007. PLS classified 78.3%, 100% in full range and 78.6%, 54.3% in narrow range respectively, for healthy and HLB trees. PLS modeling for HLB and he althy trees from the greenhouse in natural light (Aug 3, 2007) with 0 10 20 30 40 50 60 70 80 90 Classification accuracy (%) Full range Narrow range Healthy HLB

PAGE 85

85 various citrus is shown in Table 4 4. Full range and narrow NIR ranges were almost the same in identifying healthy and HLB infected trees except healthy trees of calamondin and sour orange In the full range, PLS classified 79.5% healthy and 86.1% HLB trees correctly (Figure 4 5), while in the narrow range, it has classified 74.5% of healthy and 79.8% of HLB trees correctly. Table 4 3 PLS modeling for HLB and healthy trees showing total samples, correct classifications in a greenhouse with artificial light (Aug 2, 2007). Variety HLB Healthy Eureka lemon Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct (Full) 4 5 Correct (Narrow) 0 5 Mandarin Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct (Full) 4 2 Correct (Narrow) 4 4 Madam Vinous Calibration 3 5 Validation 2 5 Total no of spectra 5 10 Correct (Full) 2 5 Correct (Narrow) 2 5 S unchusha Calibration 3 3 Validation 2 3 Total no of spectra 5 6 Correct (Full) 2 3 Correct (Narrow) 2 3

PAGE 86

86 Figure 4 4. Percentage of correct classification of HLB and healthy trees in full range (400 2450 nm) and narrow NIR range (400 900 nm) on Aug 2, 2007. 0 20 40 60 80 100 Classification accuracy (%) Full range Narrow range Healthy HLB

PAGE 87

87 Table 4 4. PLS modeling for HLB and healthy trees showing total samples, correct classifications in a greenhouse with natural light (Aug 3, 2007) Variety HLB Healthy Calamondin Calibration 5 5 Validation 4 5 Total no of spectra 9 10 Correct (Full) 3 4 Correct (Narrow) 4 0 Duncan grapefruit Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct (Full) 5 5 Correct (Narrow) 5 5 Madam Vinous Calibration 4 5 Validation 5 5 Total no of spectra 9 10 Correct (Full) 5 4 Correct (Narrow) 5 4 Maxican lime Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct (Full) 5 5 Correct (Narrow) 5 5 Trifoliate Orange Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct (Full) 5 5 Correct (Narrow) 5 5 Sweet lime Calibration 5 5 Validation 10 10 Total no of spectra 5 5 Correct (Full) 5 5 Correct (Narrow) 5 5

PAGE 88

88 Table 4 4. Continued Variety HLB Healthy Valencia Calibration 5 4 Validation 5 6 Total no of spectra 10 10 Correct (Full) 4 1 Correct (Narrow) 4 3 Sunchusha Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct(Full) 5 5 Correct (Narrow) 5 5 Sour orange Calibration 5 5 Validation 5 5 Total no of spectra 10 10 Correct(Full) 3 5 Correct (Narrow) 5 0 Figure 4 5. Percentage of correct classification of HLB and healthy trees in full range (400 2450 nm) and narrow NIR range (400 900 nm) on Aug 3, 2007. 0 20 40 60 80 100 Classification accuracy (%) Full range Narrow range Healthy HLB

PAGE 89

89 The results of discriminant analysis are shown in Table 4 5. Overall, it shows that discriminant analysis can correctly classify the HLB spectra. It is more accurate at the full range than the narrow range. Table 4 5. Number of misclassified spectra in di scriminant analysis. Values given in the parenthesis are percentage of misclassified spectra in full range and narrow range. HLB Healthy Date (2007) Total no of spectra No of misclassified spectra Total no of spectra No of misclassified spectra Full range Narrow range Full range Narrow range June 13 49 0 (0) 10 (20.4) 64 0 (0) 5 (7.8) June 14 70 3 (4.3) 18 (25.7) 58 0 (0) 18 (31.0) Aug 2 40 0 (0) 0 (0) 40 0 (0) 2 (5.0) Aug 3 158 16 (10.1) 54 (34.2) 158 14 (8.9) 53 (33.5) Figure 4 6 shows the canonical plot of the points and the multivariate means of healthy and HLB infected trees in the full spectral range for Aug 2, 2007. Each multivariate mean is a labeled circle. The size of circle corresponds to a 95% confidence limit for the mean. Groups that are significantly different tend to have non intersecting circles. Discriminant analysis showed that it has good potential to classify HLB.

PAGE 90

90 Figure 4 6. Canoniocal plot shows the points and the multivariate means of HLB and healthy trees in full range for Aug 2, 2007. The reason behind misclassifying healthy and HLB infected tress might be the degree of infection in HLB infected trees. Some HLB infected trees show s evere infection all over the tree while some trees show symptoms in only a small portion of the tree. Thus we need to collect data very precisely with all the information about its degree of infection. If the miss classification at narrow range can be tol erated, i t is quite possible to develop a low cost rugged sensor to collect reflectance spectra at the narrow range. Such a sensor can be used in assisting the scouting process. However, a sensor that can cover the full range would have a significantly hig her cost Conclusions Partial least squares (PLS) modeling and discriminant analysis techniques identified HLB under field conditions and in a greenhouse with artificial lights. Results supports that these techniques have the potential to discriminate H LB for different types of citrus. Overall, the full range of data gave more accurate results compared to narrow range with both techniques. However, the narrow range (400 nm to 900 nm) data gave better results with PLS modeling. In contrast, discriminant a nalysis was better overall

PAGE 91

91 using the full spectral range. It seems that the narrow range can produce very good results if the HLB symptoms are visible, but a major goal is to detect HLB before visible symptoms appear.

PAGE 92

92 CHAPTER 5 IDENTIFICATION OF CI TRUS GREENING (HLB) INFEC TED CITRUS TREES USI NG SPECTROSCOPY AND STA TISTICAL CLASSIFICAT ION Introduction One of the most important diseases found in citrus in Florida is citrus greening, also known as Huanglongbing or HLB. It is a systemic bacterial disease tr ansmitted by Asian citrus psyllids ( Diaphorina citri ) The disease causes substantial economic losses to the citrus industry by shortening the life span of infected trees. In the infected citrus orchards, trees are decimated and the productive duration of fruit bearing is reduced. As there is no cure reported for citrus greening so far, the growers have to remove the infected trees. The elimination and removal of infected trees due to citrus canker and greening diseases contributed to the gross loss of 19,9 18 acres on Florida (NASS, 2009) Muraro (2007) reported that the total grower costs of fresh fruit increased from $ 1,657 to $ 2,283 per acre due to HLB. Yellow angular blotching has been considered a symptom specific to the HLB disease and consists of blotches of yellow on dark greenish grey leaves. By the time these symptoms are apparent, a plant can already be severely affected. Takushi et al. (2007) reported that starch content of HLB affected leaves can be 20 times higher than leaves from healthy leaves. Etxeberria et al. (2009) studied anatomical distribution of abnormally high levels of starch in HLB infected valencia orange trees. They reported phloem collapse in HLB infecte d leaves in addition to starch accumulation. They found multiple starch grains per chloroplast in HLB infected leaf palisade cells whereas healthy leaf chloroplasts include a small number of lipid inclusions and occasional smaller starch grains. They furth er reported that HLB infected leaves have corky texture due to the thicker photosynthetic cell walls.

PAGE 93

93 Accurate diagnosis of HLB is essential before applying control strategies like tree removal to prevent a major outbreak. HLB diagnosis is difficult base d on field observations as the symptoms bear resemblance to nutrient deficiencies such as zinc deficiency (Etxeberria et al.2007). Electron microscopy and bioassay can be used to diagnose HLB but are time consuming and cannot be done under field conditions (Chung and Brlansky, 2005). Molecular methods like real time polymerase chain reaction (PCR) based assays are used to detect the presence of HLB. However, identification of plants as suspect by foliar and fruit symptoms is required by a trained scouting c rew prior to real time PCR assays. The current methods being expensive, time consuming and tedious necessitate developing a rapid and reliable method to identify HLB infected trees from healthy trees. Visible (VIS) and near infrared (NIR) spectrum of a lea f contains information on plant pigment concentration, leaf cellular structure, and leaf moisture content (Borengasser et al. 2001) Previous studies have shown that VIS NIR spectroscopy has the potential to iden tify plant anomaly due to disease or malnutrition (Bravo et al. 2004; Zhang et al. 2002; Zhang et al. 2005) Smith et al. (2005) studied the plant stress caused by elevated levels of natural gas in the soil, dilute herbicide solution, and extreme shade. They found that the red edge position was strongly correlated with chlorophyll content across all the treatments The rat io of reflectance centered on the wavelengths 670 and 560 nm was used to detect increases in red pigmentation in gas and herbicide stressed leaves. Stress due to extreme shade could be distinguished from the stress caused by natural gas and herbicide by a nalyzing the change in spectral features. Liu et al. (2007) characterized and estimated rice brown spot disease severity

PAGE 94

94 using stepwise regression, principal component regression, and partial least squares regression With PCA they predicted disease severity with root mean square errors (RMSE) of 16.3% and 13.9 % for the training and testing dataset while with PLS with for training a nd testing dataset, respectively. They concluded that it was feasible to estimate the disease severity using hyperspectral reflectance from the leaves. Lee et al. (2008) used aerial hyperspectral imaging to detect HLB They used spectral angle mapping (SAM) and spectral feature fitting (SFF) methods. They reported that it was difficult to obtain good results with SAM and SFF because of the positioning errors in GPS ground truthing and aerial imaging, and the spectral s imilarity between non symptomatic HLB infected trees and healthy trees. Delalieux et al. (2007) used hyperspectral imaging and parametric approaches such as logistic regression, partial least squares, discriminant analysis and tree based modeling to detect biotic stress ( Venturia inqequalis ) in apple trees. A fast method for detecting HLB in the field will assist growers to better manage and control the disease resulting in significant production and economical ben efits The long term goal of this study is to develop a ground based method to detect HLB at early stages of development in the orchard. The specific objective of this study was to investigate the possibility of identifying HLB infected trees using VIS NIR spectroscopy and to determine the best classification techniques. Materials and Methods Field experiments A total of 1,239 spectra were collected from 135 (80 HLB, 55 healthy) Valencia orange trees. Table 5 1 shows detailed information on the data used in this study. The

PAGE 95

95 age of all the healthy and HLB infected trees, used in this study, were 15 20 years old. S pectral data were collected using two portable spectroradiometers. The first series of data were collected with a FieldSpec 3 spectroradiometer (Analytical Spectral Devices (ASD), Boulder, CO) The reflectance data from 350 nm to 2 500 nm were collected u sing the spectroradiometer and transmitted wirelessly to a laptop computer The equipment was set up to collect 10 scans average the scans and represent it as a single observation Bare fiber optic cable with 25 field of view (FOV) was used for data coll ection. The ASD spectroradiometer has a circular field of view. The integration time was optimized using the optimiz ation options within the software. Optimization values depend on the response to light in a particular spectral region. The bare fiber optic cable was placed at a distance of approximately 50 to 80 cm from the tree canopy, where sensor can see canopy clearly. During the data collection, the area scanned by the spectroradiometer was approximately 385 to 988 cm2. The second series of data were c ollected using a SVC HR 1024 portable spectroradiometer (Spectra Vista Corporation, Poughkeepsie, New York). Spectral range of SVC is 350 nm 2,500 nm with a spectral Table 5 1. Spectral data from healthy and HLB infected trees used in this study Locations in FL Data collection date No. of trees No. of Spectra Type of Spectroradiometer HLB Healthy HLB Healthy Lake Alfred June 13, 2007 2 2 119 122 ASD Lake Alfred June 25, 2007 4 4 90 60 ASD Lake Placid Feb 22, 2007 4 4 24 24 ASD Immokalee Jan 16, 2009 10 5 100 50 SpectraVista Immokalee Feb 5, 2009 10 5 100 50 SpectraVista Immokalee Feb 27, 2009 10 5 100 50 SpectraVista Southern Garden Mar 24, 2009 40 30 200 150 SpectraVista

PAGE 96

96 resolution of 3.5 nm (350 1000 nm), 9.5 nm (1000 1,850 nm), and 6.5 nm (1,850 2,500 nm) which was similar to the ASD spectroradiometer The SVC HR 1024 spectroradiometer communicated with a handheld PDA through a Bluetooth. All the data were collected with 4 FOV. The FOV of the Spectra V ista spectroradiometer was rectangular which cover ed an area of 26.6 x 9.8 cm from a distance of 665 cm. A minimum integration time of one millisecond was used In the field condition scan time was set at 4 s Laboratory tests were co nducted with various objects to compare the spectral signatures of ASD and SVC HR 1024 spectroradiometers. Similar results were obtained from both spectroradiometer. Dark current measurements were automatically taken immediately prior to the reference or t arget scans using the ASD and SVC spectroradiometers. The spectral data were collected from sunlit canopy between 11:00 am and 2:00 pm with no additional source of light Spectral reflectance was collected from tree canopy from two sides of the tree to com pensate for the effect of shade. Both spectroradiomters were handheld and trees were scanned by their sides. R eference readings of white panel were collected every 10 min to reduce the error in the reflectance due to the atmospheric changes. Both spectro radiometers provides relative reflectance values of the tree canopy based on the reference value. D ata A nalysis Spectral pretreatment and feature selection The Multiplicative Scatter Correction (MSC) was performed on the NIR portion of the spectra. Then, the local average of the spectrum, and its first and second derivatives were calculated at selected wavelengths along the spectra. To ensure complete use of the information in the spectra, it was decided to compute these values

PAGE 97

97 at every 25 nm along the entire spectrum, except in the noisy regions. Therefore, the local mean, and the first and the second derivatives were calculated at 25 nm intervals from 375 nm to 1,325 nm, from 1,500 nm to 1,750 nm, and from 2,050 to 2,300 nm. At each of these 61 points, the local mean was computed by averaging the spectral reflectance values for that wavelength and its four neighbors. For example, the local mean at 375 nm was computed by averaging from 373 nm to 377 nm. In order to avoid noise amplification, which results from differentiation, the Savitzky Golay method (Orfanidis 1996) was used to compute the first and second derivat ives. A quadratic polynomial and a window size of 21 were used. As three values were computed for each of the 61 points mentioned above (i.e. the mean, first and second derivatives), each spectrum was represented by a feature vector of 183 elements. Becaus e the number of features was significantly high and many of these features could be correlated, Principal Component Analysis (PCA) was performed to reduce the number of features (Figure 5 1). The 183 spectral elements were reduced to 25 principal component s (PCs) that accounted for more than to above 90% variance within the data. From the Fig. 5 1, it can be seen that the first 25 PCs contribute to about 90% variance. The data analysis described in this paper was performed in Matlab (Mathworks Inc., Natick, MA).

PAGE 98

98 Figure 5 1 Variance of data explained by principal component analysis Classification The principal components (25) were used as the input features for the classification algorithm. The output classes obtained from the classification models were data were used for training, while 25% of the data were used for testing the classification algorithms. Weighted K Nearest Neighbors (KNN) Weighted k Nearest Neighbors (KNN) is an instance based classification algorithm. These types of algorithms do not develop an explicit function or model for predicting the target classes, instead all the training samples are stored and computation is delayed until a new unknown sampl e must be classified. For a new sample, the Euclidian distance between its feature vector and the feature vectors of the training samples is computed. The computed distances are then used to find the k training samples that are closest to the unknown sampl e and a prediction is made

PAGE 99

99 based on these nearest neighbors. For a classification problem, this can be done simply by a majority vote. A more sophisticated approach used in this study, was to give a different weight to the contribution of each of the K nea rest neighbors in the prediction. Commonly, a weight that is inversely proportional to the square of the Euclidian distance is used (Mitchell 1997) : = arg max ( = 1 ( ) ) = 1 ( ) 2 = 1 In these equations, indicates the predicted class for the new unknown feature vector denotes t he th element of the th feature vector, are the weights, and the function is defined as follows: = 1 = 0 data. Logistic Regression (LR) Logistic regression provides a model for the probability of occurrence of an event by fitting the data to the logistic curve. As shown in Fig. 5 2 this curve maps the entire real axis onto the interval [0, 1], making it ideal for modeling the probability of an event. When used in classification, the variable is usually defined as a linear combination of the features (Witten and Frank 2005) : = = 1 1 + (4) (1) (2) (3)

PAGE 100

100 Figure 5 2. The logistic curve describing logistic regression model In this equation, is the parameter vector, is the feature vector, and = 1 if and = 0 if A batch gradient descent approach was used for training the model, as shown in the following pseudocode: + 1 = + = 1 ( ) This iteration continued for all j till the gradient converges to local minima. Here, N is the number of features, is the learning rate, and is the true label (either 1 or 0). A sufficiently small will ensur e that the global optimum will be achieved, although the computation time will increase by decreasing the value of In this study a value of was used. Support Vector Machines (SVM) Support Vector Machine s (SVM) is one of the most successful machine learning algorithms that is widely used in various fields. The basic SVM solves a classification problem with only two target classes. However, it can be generalized to solve problems (5)

PAGE 101

101 that involve more than two classes. For a linearly separable data, such as the data that is shown in Fig. 5 3 one can draw many different hyperplanes that can separate the data. The idea in SVM, however, is to find a hyperplane that separates the data with the largest possible mar gin as shown in Fig. 5 3 b. SVM uses numeric labels 1 and 1 to identify the two classes. In this paper, the following notation will be used to show the SVM classifier (Webb, 2002) : = = ( + 0 ) A B Figure 5 3 A ) An example of a linearly separable set of data, and B) the maximum margin classifier for this data set The parameters and define the decision boundary (i.e., the separating hyperplane). The int ercept is a scalar, while and the feature vector are dimensional vectors. In this equation, if and if The equation defining the separating hyperplane is as follows: + 0 = 0 (7) (6)

PAGE 102

102 The geometric margin of the th example from the separating hyperplane can be calculated using the following equation: = + 0 Here, a positive would mean that and are of the same sign, or equivalently, the th example is correctly classified. The geometric margin of the classifier with respect to the set of all training samples, is defined as the minimum of the geometric margins for each sample: = min = 1 : Therefore, the goal of SVM is to find a separating hyperplane that maximizes In mathematical terms, SVM will pose the following optimization problem: Such that ( + 0 ) = 1 = 1 The s econd constraint is a non convex constraint and this optimization problem cannot be directly solved. It can be shown (Gunn 1998) that this optimization problem can be replaced by the following equivalent problem: ( + 0 ) = 1 (8) (9) (10) (11)

PAGE 103

103 In the above equation, is called the functional margin and is shown by The difference between functional margin and the geometric margin is that, multiplying and by a constant will not change the geometric margin but will change the functional margin. Functional and geometric ma rgins are related as follows: = Equation 1 1 means that the optimal margin classifier sought by SVM can be found by minimizing the norm of w under the constraint that the functional margin for all examples is at least equal to 1. It is easy to see that the minimum margin always belongs to the points that are on the edge of the widest strip that separates the data as shown in Fig. 5 3 b. Only f or these poi nts ; for the remaining points, very small fraction of the total number of points, resulting in huge reduction in the computational cost. Although the optimization problem in Eq. 1 1 can be solved by commercially available quadratic programming code, an easier equivalent problem can be found by using the method of Lagrange multipliers (S trang, 1991) Using this method, the following dual optimization problem will be obtained: max = 1 1 2 = 1 0 = 1 = 0 = 1 (12) (13)

PAGE 104

104 In this equation, s are the Lagrange multipliers and indicates the inner product between the th and th feature vectors. The interesting point is that all of the Lagrange multipliers are zero except for the ones that correspond to the support vectors. Once the above optimization problem is solved, and can be fou nd using the following formulas: = = 1 0 = = 1 + = 1 2 It is important to note that the new optimization problem in Eq. 13 is only in terms of the inner products of the feature vector s. Moreover, with the new definition of in terms of Lagrange multipliers (Eq. 1 4 ), for a new example the prediction of the SVM classifier can be written as follows: = = + 0 = 1 In other words, the entire algorithm can be written in terms of the inner product of feature vectors. In many practical applications, the data are not linearly separable or, even if it is, the optimal margin classifier described so far may not be the best choice because it can be very sensitive to outliers. Therefore, a modified version of SVM problem is defined as follows (Webb, 2002 ) : 0 1 2 2 + = 1 (14) (15) (16)

PAGE 105

105 + 0 1 0 = 1 With this definition, some of the examples are allowed to fall on the wrong side of the separating hyperplane. However, for each such example, a cost, will be considered. Th e parameter determines the importance that we attach to these errors. Once again, the method of Lagrange multipliers can convert this problem into an easier one. The resulting problem will be as follows: 1 2 ( ) = 1 = 1 0 = 1 = 0 = 1 This is the same as the previous optimization problem in Eq. 12, except that the first constraint has changed from to Although the SVM algorithm described so far is already a very powerful method, significant improvements would result by introducing kernels. The idea is to map the feature vector into a higher dimensional space using a mapping function As mentioned before, the entire SVM algorithm can be written in terms of the inner products of the feature vectors. After mapping feature vectors using all instances of the inner products of the feature vectors will be replaced by the inner product of the corresponding mapped feature vectors, This inner product is called a kernel: = ( ) (17) (18)

PAGE 106

106 The idea of kernels significantly improves the SVM technique in terms of both its accuracy and the scope of problems that it can solve. Various different kernels have been introduced and used in practice (Ch erkassky and Mulier 2007) Following the suggestion of (Hsu et al. 2008) a Gaussian kernel was used in this study: = 2 Therefore, the SVM problem that was solved in this study was of the following form: 1 2 ( ) = 1 = 1 0 = 1 = 0 = 1 : = 2 To solve this problem, the sequential minimal optimization (SMO) algorithm was used. This algorithm solves the above optimization problem by a coordinate ascent approach. Because of the second constraint in the problem, i.e. it is not possible to change only one of the Lagrange multipliers. The SMO algorithm optimizes the function with respect to two Lagrange multipliers simultaneously. The advantage of this approach is that the innermost loop is very fast compared to other algorithms (Platt, 1998) (19) (20)

PAGE 107

107 The parameters and in the problem are unknown. Therefore, the first step in solving the problem is to find the optimum values for these parameters. The grid search procedure suggeste d by Hsu et al. (2008) was used for this purpose. First, the approximate values of the optimum and were found on a coarse grid with and After finding the approximate values for the best and a finer grid was defined to search for more accurate values of the optimum and Reduci ng the C lassification E rror by U sing M ultiple M easurements The goal of the present study was to develop classification algorithms that can be used for automatic detection of HLB infected citrus trees in the field. The preliminary results indicated that the classification accuracy from the statistical methods to classify HLB infected trees were low when data with single measurements were considered for analysis. This could be due to variability in experimental conditions such as the amount of sunlight and th e orientation of the leaves with respect to the sensor. Therefore, it was hypothesized that high detection accuracy can be achieved through multiple measurements on a single tree. In this study, we investigated the accuracy of classification methods (KNN, LR, and SVM methods) using spectral data from one, three, or five measurements from different canopy areas of the same tree. When more than one measurement was presented to the classifier, a simple method based on majority voting can be used to determine the classification group. The tree is labeled as HLB infected tree when more than half of the measurements from that tree are classified as HLB infected. An improved approach is to consider the confidence of the classifier for its prediction on each of the measurements. For example, if a

PAGE 108

108 confidence, it would be more reasonable to classify the t was based on the majority with confidence of the classifier. For the KNN algorithm, the following equation was used to estimate the confidence of each prediction: = w i is a lready defined in Eq. 2. For the LR method, Here w i is for those among the K nearest neighbors that have the same class as the predicted class and w i is for all the K nearest neighbors. The value of the function in Eq. 4 can be used to calculate a measure of confidence in prediction. The closer is to either 0 or 1, the more confident is the prediction. Therefore, the value was used to evaluate the confidence of the predictions by LR method: = 0 5 For the SVM classifier, the value of the function in Eq. 15 can be used as a measure of confidence in pred iction. The larger the absolute value of this function, the more confident the prediction. Therefore, was used to evaluate the confidence of predictions by the SVM method: (22) (23) (21)

PAGE 109

109 Results and Discussion Figure 5 4 shows samples of the collected spectra from both healthy and HLB trees. Each observation is an average of ten measurements. Noise was observed in bands 1350 to1500 nm, 1750 to 1950 nm and bands after 2350 nm due to the presence of atmospheric moisture. Figure 5 4 Representative spectroradiometer spectra: the spectra from two healthy and two HLB infected trees For the SVM method, the first step was to find the optimum values for and As mentioned before, this was performed by a gird search. Figure 5 5 shows typical contour graphs that were used to find the optimum parameter values. First, the classification error was evaluated on a coarse grid (Figure 5 5a) to fi nd approximate estimates for and Then, a finer grid was defined (Figure 5 5b) and used to determine accurate values of and that minimized the classification error. In the

PAGE 110

110 specific case shown in the Table 5 2, the lowest classification error was approximately 18% which was obtained for and Figure 5 5. Contour plots of classification error for finding the optimum values for the parameters C and Analysis showed that MSC correction of the spectra did not improve the classification accuracy. In other words, when MSC was removed as a pretreatment procedu re described in the data analysis section of the paper, the classification error did not improve for any of the three classification techniques. Table 5 2 shows the classification error for each of the three classification techniques without MSC correction The table shows the percent of misclassified citrus trees when one, three, or five spectral measurements from each tree were used for classification. Table 5 2. Average classification error for three classification techniques. Classification method Error with one spectrum Error with three spectra Error with five spectra Weighted KNN 23% 11 % 6.5 % Logistic regression 35% 2 3 % 1 9 % SVM 18% 6 .2 % 3 .0 % It can be seen from Table 5 2 that the classification error with a single spectrum was relatively large. However, when three or five spectra from the same tree were used, the classification error significantly decreased. The SVM method demonstrated lower

PAGE 111

111 c lassification errors than other two methods, especially with five spectra. The classifiers obtained by weighted KNN, logistic regression, and SVM algorithms can be easily programmed on a microcontroller. The three algorithms were also very different in ter ms of computation time. For SVM based classification, the most time consuming step was to find the optimum values for and The grid search method for optimal value selection used in this study took several hours on a PC. Although there are faster methods for finding and (Hsu et al. 2008) thi s step is time consuming. The batch gradient descent algorithm described previously for logistic regression method took approximately 12 s to complete on a PC with a 4.8 GHz processor and 512 MB of RAM. As mentioned before, the KNN method does not develop any model from the data. Therefore, for KNN algorithm there is no computation until a new spectrum is to be classified. Once the model (i.e., the classifier) is obtained, however, the computation time for classification of a new spectrum is very fast for S VM and logistic regression. On the same PC as mentioned above, the time to make a prediction based on three spectra from the same tree was only 0.06 ms for logistic regression classifier, whereas for weighted KNN and SVM methods, this time was 5.5 ms and 5 .0 ms, respectively. Therefore, the logistic regression method was computationally much faster. As the KNN and SVM based algorithms yielded better classification accuracy than logistic regression, it would be considered as a more preferable algorithm for H LB detection in citrus groves. Conclusion The goal of this study was to develop a technique for rapid detection of HLB infected citrus trees. Canopy reflectance spectra w ere measured from the infected and

PAGE 112

112 healthy trees using a spectroradiometer and three common classification algorithms were used to classify the infected trees from the healthy ones. The results indicated that a single measurement was insufficient for accurate detection of the infected trees. The classification error was between 18% and 35% using a spectrum. However using multiple spectral measurements from a single tree, the classification accuracy increased significantly. SVM method showed an accuracy of higher than 97% when it was provided with five spectra from the same tree. Under real field conditions, varying sunlight and other environmental factors can produce noise that might reduce the classification accuracy. Under these conditions, multiple measurements will be necessary to ensure acceptable classification accuracy.

PAGE 113

113 C HAPTER 6 AN ACTIVE OPTIC SENS OR FOR DETECTION OF HUANLONGBING (HLB) DISEASE Introduction Florida accounts for 70 percent of total U.S. citrus production and produced 7,236 tons of citrus from 224,358 hectare (554,400 acre) of bearing orchards in 2006 200 7 (FASS, 2008) Huanglongbing (HLB), also known as greening, a systemic bacterial disease transmitted by the Asian citrus psyllid ( Diaphorina citri ), is considered one of the most devastating citrus diseases in the world. Since HLB is a relatively new disease in the USA, very little published information is available on its dynamics, epidemiology, and molecular characteristics. Figures 6 1 and 6 2 show leaves of HLB infected and healthy tree, respectively. Figure 6 1. HLB symptomatic

PAGE 114

114 Figure 6 2. Hea lthy leaves Reflectance spectra of vegetation, measured in the visible and infrared regions, contain information on plant pigment concentration, leaf cellular structure, and leaf moisture content ( Borengasser et al. 2001). A multi band sensor measures ref lected radiation at specific wavebands. Malthus and Madeira (1993) studied the spectral reflectance of field bean leaves attacked by Botrytis fabae They found the most significant changes in spectral reflectance due to Botrytis fabae were flattening of th e response in the visible region and a decrease in the near infrared reflectance, around 800 nm. Correlation between percentage infection and reflectance in the visible region (peaks at 525 nm and 589 nm) were higher in the first order derivative spectra t han for the original reflectance spectra (zero order). The study indicated the potential for using spectral information for disease detection. Such techniques can be used to identify HLB in citrus trees. However, very little is known about the spectral cha racteristics of HLB

PAGE 115

115 infected leaves in the visible and near infrared (NIR) regions. This band specific information could discriminate HLB infection in the citrus grove. Currently, the detection of HLB relies on scouting groves for visible symptoms and foll owing up with off site diagnosis of the disease using the polymerase chain reaction (PCR) technique. HLB infection increases the amount of starch accumulation in leaves (Schneider, 1968). Therefore, on immersing the HLB infected leaves in iodine solution f or one or two minutes, the leaves stain very dark grey to black along cut surfaces while healthy citrus leaves show no or very little staining (Etxeberria et al. 2007). However, this procedure is too slow for testing of every tree. Infected trees are remo ved and insecticides are sprayed to control the population of Asian citrus psyllids. Scouting is laborious, time consuming, often subjective, and prone to errors. Knowledgeable growers in Brazil estimate that at least 50% of infected trees with visible sym ptoms go undetected by trained scouts. Similar results, even if using an observation platform, have been reported in Florida (Futch et al. 2009). Rapid, early, and accurate diagnosis, especially at the orchard level, is essential to eliminate the disease at early stages of infection. Non uniform distribution of disease organisms complicates the tests that are based on detecting the causal agent or visual symptoms. In the case of HLB, it has been shown that high incidence of Candidatus Liberibacter asiaticu s (CaLas) in psyllids can be found in an area well before symptomatic plants are found (Manjunath et al. 2008). Presently, there is a need for rapid detection techniques to prevent the spread of disease. The present work evaluates a spectroscopic method f or early detection of HLB infected trees. The specific

PAGE 116

116 objective of this study was to develop and evaluate a multi band active optic sensor for detecting HLB infected citrus trees under field conditions. Material and Methods Data C ollection with M ulti B and S ensor The sensor was composed of four narrow band (active optic) light sources with four different wavelengths; two in the visible region (at 570 nm and 670 nm) and two in the near infrared region (at 870 nm and 970 nm), with accuracy of 1% reflectance. All the four bands have epoxy lens type illuminator manufactures by Marubeni America Corporation (Santa Clara, CA 95054). These illuminators are wide viewing and extremely high output power illuminators assembled with a total of 60 high efficiency aluminum gallium arsenide (AlGaAs) diode chips, mounted on metal stem TO 66 with aluminum nitride (AlN) ceramics and covered with double coated clear silicone and epoxy resin. Sampling frequency was set at 10 Hz and that was also the averaging time. It was not inf luenced by ambient conditions. All the four detectors were arranged to view same area of target. Vibrating leaves or moving while sampling, the ambient conditions may vary and have very low frequency content. Sunlight of course even with clouds will have only a fraction of 1 Hz. The ambient conditions then are essentially negligible in their effect. Only the original transmitted light signal is passed for measurement. Design of this sensor eliminates the need for dark operation by using modulated illumi nation. The sensor was calibrated by measuring irradiance over a This material has a reflectance of approximately 0.98 in the four selected bands. The sensor also incorporates compens ation for temperature and variations in the supply voltage. These effects are measured by detecting the level of light output from the LED illuminators and

PAGE 117

117 the reflected irradiance is adjusted accordingly. The sensor automatically generated the values of reflectance for each of the four wavelengths. The sensor was interfaced with a hand held personal digital assistant (PDA) through a serial port. The PDA saved the reflectance data and the internal temperature of the unit at the time of reading. A similar s ensor was used in plant physiology studies at International Maize and Wheat Improvement Center (CIMMYT) in Mexico Stone et al. (2010) detected nitrogen status in winter wheat using similar sensor.The general design of this sensor is identical to the Gree nseeker manufactured by the NTech division of Trimble Navigation Inc. (Stone et al., 2003). Figures 6 3 and 6 4 depicts the four band sensor used in this study The spectral data from the multi band sensor were collected in a grove at Fort Basinger (lon gitude 08220.5248 W, 2721.9617 N), Immokalee (longitude 08126.5901 W, 2628.0339 N), and a third location (longitude 8126.5915 W, 2628.0241 N) near Clewiston in Florida. Valencia orange trees were measured in Immokalee and Clewiston, and Mid sweet orange tr ees were evaluated in Fort Basinger. A total of 10 trees (5 HLB, 5 healthy) in Fort Basinger, 10 trees in Immokalee (10 HLB), and 58 trees (37 HLB, 21 healthy) near Clewiston in Florida were evaluated in the data analysis. Since there were no healthy trees of Valencia available in the grove at Immokalee, 10 healthy trees of Valencia near Lake Alfred were evaluated. Ten readings from both HLB symptomatic and healthy tree leaves were collected from different locations in the same tree. For the HLB infected tr ees the data collected from the same branch that was confirmed HLB positive by the PCR test. These data were collected in Fort Basinger during fall 2008 and in Immokalee and Clewiston during spring 2009.

PAGE 118

118 The four band sensor was build from Applied Technolo gy (Stillwater, OK). The sensor was composed of four narrow band (active optic) light sources with four different wavelengths: two in the visible region (at 570 nm and 670 nm) and two in the near infrared region (at 870 nm and 970 nm). The band width for e ach spectral band was about 50 nm. After lighting the target, the reflected light was captured by a receiver located in the center of the device. The sensor automatically generated the values of reflectance for each of the four wavelengths. The sensor was interfaced with a hand held personal digital assistant (PDA) through a serial port. The PDA saved the reflectance data and the internal temperature of the unit at the time of data collection. Figures 6 3 and 6 4 depict the multi band sensor used in this st udy. Measurements of reflectance were made at a distance of about one meter between the sensor and the target with the sunlight intensity below 900 lux. Based on the preliminarily tests, it was determined that these were the best conditions for acquiring s pectral readings with the multi band sensor.

PAGE 119

119 Figure 6 3. Multi band active optic sensor Figure 6 4. Field measurements using the four band sensor

PAGE 120

120 Data A nalysis Each measurement was composed of four reflectance values at 570, 670, 870, and 970 nm. Using these values, 11 different vegetati on indices were computed. Table 6 1 shows the details of the vegetati on indices computed in this study. Table 6 1. List of vegetation indices used in analysis. Vegetation Index (VI) Equation Normalized Difference Vegetation Index (NDVI 1 ), Rouse et al (1974) 870 = 870 670 870 + 670 NDVI 2 970 = 970 670 970 + 670 Simple Ratio Index (SR 1 ), Rouse et al (1974) 870 = 870 670 Simple Ratio Index (SR 2 ) 970 = 970 670 Modified Triangular Vegetation Index (MTVI 1 ), Haboudane et al. (2004) 1 = 1 2 [ 1 2 870 570 2 5 670 570 ] Modified Triangular Vegetation Index (MTVI 2 ), Haboudane et al. (2004) 2 = 1 5 [ 1 2 870 570 2 5 670 570 ] ( 2 870 + 1 ) 2 6 870 5 670 0 5 Renormalized Difference Vegetation Index (RDVI), Rougean and Breon (1995) = ( 870 670 ) ( 870 + 670 ) Greenness Index (G) = 550 670 Triangular Veg. Index (TVI), Broge and Leblanc (2000) = 0 5 [ 120 870 570 200 ( 670 570 ) Modified Chlorophyll Absorption in Reflectance Index (MCARI 1 ), Haboudane et al. (2004) 1 = 1 2 [ 2 5 870 670 1 3 ( 870 570 ) ] Structure Intensive Pigment Index (SIPI), Zacro Tejada (2000) = ( 870 570 ) ( 870 + 670 ) Outliers in the data were detected and removed prior to the classification of diseased and healthy trees. This was done by performing a principal component analysis (PCA) on the data and plotting the first four principal components. Among a

PAGE 121

121 total of 1552 m easurements, 20 measurements were recognized as outliers (larger than 3 standard deviation) from PCA analysis and removed. After the removal of outlier, the four reflectance values (raw data) and the vegetation indices were used for the classification. Fiv e different classification techniques were applied to the data. In the following paragraphs, these techniques are briefly described. Decision T rees The iterative dichotomiser 3 (ID3) algorithm (Coppin, 2004) was used to build decision trees. The number of decision tree layers was seven. This number was selected based on some preliminary analysis to find the optimum number of layers for best results, i.e. to avoid over fitting. No pruning was performed on the trees. Because decision trees are inherently unst able classifiers, it was decided to develop an ensemble of decision trees using a stacking scheme (Polikar, 2006). More specifically, first 20 decision trees were built, and then another decision tree was used to learn the output pattern of these decision trees. k N earest N eighbors (KNN) A weighed k nearest neighbor scheme was used, with weights inversely proportional to the square of Euclidian distance (Fukunaga, 1990). In our analysis, k was chosen as 25 based on the preliminary analysis. Logistic R egres sion In this method, the goal is to adjust the parameters ( ) of the logistic curve in order to best fit the curve to the training data (Larose, 2006). = 1 1 + (1)

PAGE 122

122 A batch gradient descent method was u sed to find the optimum values for the parameter Neural N etworks A feed forward neural network with a single hidden layer containing 20 neurons and 2 output neurons (equal to the number of classes, i.e. healthy and HLB infected) were used. Sigmoid tran sfer functions were used in the hidden and output layers. The scaled conjugate gradient method was used to train the network (Haykin, 1998). Support V ector M achines (SVM) This technique aims to find the hyperplane that separates the data with the largest p ossible margin. In this study a modified SVM method (Webb 2002) with a Gaussian kernel was used. This will lead to an optimization problem of the following form: 1 2 = 1 ( ) = 1 (2) i = 0 = 1 (3) Where, = 2 (4) On solving this optimization problem, the optimal margin classifier will have the following form: = + 0 = 1 (5) where: = = 1 (6) 0 = max = 1 + min = 1 2 (7)

PAGE 123

123 In equation 5, f(t) =1 if and f(t) = 1 if t < 0 the two classes that the classifier separates. In this study, we used the sequential minimal optimization (SMO) algorithm (Platt 1998) to solve this problem. Results and Discussion Table 6 2 shows the results for the f ive classification methods used in this study. For all the models, 75% of the data was used for training, while 25% of the data was used for testing. (Haykin, 1998). The Table 6 2 shows the classification error for each of the classification methods when o nly one measurement was presented to the classifier in the testing phase. The misclassification errors were between 17% and 40%. Because the classification errors were high, it was decided to evaluate the performance of the classifiers with more than one m easurement as input in the testing phase. Table 6 2 shows the classification error when three and five measurements from the same tree were used as an input to the classifier. These multiple measurements were taken from the different locations of the same decreased significantly with multiple measurements. This is due to large variability in the field measurements caused by environmental factors (such as the orientation of the leaves with respect to the sensor or the wind) and by the human operator (such as the non constant distance between the sensor and the leaves). Using multiple measurements eliminates these sources of noise and allows higher classification accuracies. Decision trees, SVM, and KNN achieved an accuracy of higher than 95% with five measurements from each tree.

PAGE 124

124 Table 6 2. Average classification error for different classification techniques. Neural networks Logistic regression k nearest neighbors Support vector machines Decision t rees Error with one set of measurements ( % ) 25 40 18 17 18 Error with three sets of measurements ( % ) 16 35 8.5 7.5 8.5 Error with five sets of measurements ( % ) 10 32 4.5 3.5 4.5 A closer examination of the decision tree classifiers reveals important information regarding the power of each of the vegetati on indices in separating the healthy and infected leaves. Investigation of ten different decision trees showed that the top level (i.e. root) node always tested on the RDVI index. This means that if we seek to separate the data into healthy and HLB infected classes by testing on only one variable, RDVI would be the best choice. This result is in accordance to Roujean and Breon (1995) findings. They observed NDVI was less affected by spectral and view ge ometry and DVI was less affected by soil background. Therefore, they combined these two indices and introduced RDVI which minimizes the effect of soil background as well as the view geometry. Other vegetati on indices that were frequently used by the decisi on trees, especially in the upper layers, include NVDI 970 SR 970 MTVI 2 and MCARI 1 On the other hand, SR 870 and SIPI were not used much in the decision trees, which indicate that these indices contain little or no information regarding HLB infection. Among the original reflectance values, the two values in the NIR region (i.e. 870 and 970 nm) were used much more frequently t han the ones in the visible range (i.e. 570 and 670 nm). This signifies that the canopy reflectance in the NIR region contains more information with respect to HLB infection compared to the visible range.

PAGE 125

125 Conclusions The results indicated that the multi ba nd optic sensor used in this study has a very good potential for detecting HLB infected citrus trees under field conditions. However, to obtain high classification accuracy, it is necessary to acquire multiple measurements from a single tree. The sensor ca n be integrated with the scouting practice, to improve the effectiveness in HLB disease detection. Obtaining multiple readings using this sensor is easy and fast, and can be performed by a human or an automated vehicle system. The measurements in this stud y were performed on Valencia and mid sweet orange cultivars. It would be interesting and useful to know the performance of the sensor and the classification algorithms on other orange cultivars or other citrus types.

PAGE 126

126 CHAPTER 7 APPLICATION OF HYPERSPECTRAL IMAGIN G FOR THE DETECTION OF HLB IN THE FIELD Introduction Citrus is one of the most important agricultural products in Florida as it is the largest citrus producing state in United States. Citrus production being a multi million dollar ind (NASS 2010). But recently, it has been threatened by Huanglongbing (HLB), a devastating and rapidly spreading disease of citrus. It was first confirmed in Florida in August 2005 (Halbert, 200 5; Bouffard, 2006). Asian Citrus Psyllid, Diaphorina citri is the vector of citrus greening or HLB. The bacteria are restricted to the sieve tubes of infected plants, and are acquired and transmitted by nymphs and adults of Asian citrus psyllid during fe eding (Garnier and Bov, 1983). Psyllids prefer feeding and breeding on younger leaves (Halbert and Manjunath, 2004) so younger trees are at higher risk of infection as they produce newer leaves and flushes throughout the year. Symptoms of HLB infected cit rus include a blotchy mottle or asymmetrical chlorosis, and yellowing of leaf veins due to inefficient production of chlorophyll (Brlansky et al., 2007). The fruits from HLB infected trees fall prematurely. Hyperspectral imagery offers a better solution f or the early detection plant diseases. In the hyperspectral imaging, the spectral reflectance of each pixel is acquired for a range of wavelengths in the electromagnetic spectra. Application of hyperspectral images in NIR VIS region (450 930 nm) was inves tigated to detect citrus canker and other damages in ruby red grapefruit. The authors reported 96% classification accuracy

PAGE 127

127 in determining disease and dam a ged fruits. Early detection of yellow rust disease in winter wheat was investigated by Bravo et al. (2 003) by using hyperspectral imaging. The authors reported 92 98% classification accuracy in discriminating diseased plants. Application of hyperspectral imaging was used by Safri and Hamdan (2009) to detect ganoderma basal stem rot disease in oil palm tree s. They used various vegetati on indices and red edge techniques to identify the infected trees. Conventional spectrometers or spectrophotometers measure optical spectr a from a specific field of view that is often restricted spatially Therefore, spectral d ata collection from tree crops such as citrus would require measurement of the spectrum at several spatial locations. This problem can be overcome using a hyper spectral camera. The hyper spectral camera scans a scene one line at a time and disperses light to its spectral components in each pixel in the line. Hyperspectral camera capture s the line image of the tree and disperse s it to a spectrum. Thus, each image frame includes the line pixels in one direction (spatial axis) and spectral distribution (light intensities in spectral elements) in another dimension (spectral axis), as illustrated in Figure 7 1

PAGE 128

128 Figure 7 1. Schematic representation of principle behind hyperspectral imaging (Nagoaka et al. 2007) Material and Methods Data collection Hyperspectral i mages were collected from the grove at Fort Basinger ( 27 20 81 10 Southwest Florida Research and Education Center (SWFREC) at Immokalee ( 2 6 2 81 and a grove near C lewiston ( 2 6 2 81 Florida. Figure 7 2 shows the location of the study areas. Valencia was evaluated in this study. These data were collected in Fort Basinger during F all 2008 and in Immokalee and Clewiston during S pring 2009.

PAGE 129

129 Figure 7 2. Data collection sites in Florida Images were collected with a Specim hyperspectral camera (Autovision Inc., Los Angles, CA, USA) (Figures 7 3 and 7 4) having a spectral range from 306.5 nm to 1067.1 nm with 2.7 nm spectral resolution. SpectralCube spectral imaging software (version 2.7) provided by Specim, Ltd. and AutoVision Inc. was used for capturing hyperspectral images. The windows based SpectralCube application conducts data collections by creating and storing sequences of spectra l images into files that can be used for data analysis. Spectral cube has several controls like video control to start or pause spectral video, data cube setup to open/close the data recording control window, color view to open/close the scrolling color im age window, spectral view to open/close the spectral plot window, band selection to activate band selection option, GPS information etc. to facilitate image capturing. The camera consists of 135 bands.

PAGE 130

130 Figure 7 3. Hyperspectral camera (Dimension (16x3x6 inches (length x width x depth)) Figure 7 4. Hyperspectral image acquisition in field

PAGE 131

131 A total of 40 images 13 images (8 HLB, 5 Healthy) in Fort Basinger, ten images (HLB) in Immokalee the and 20 images (10 HLB, 7 Healthy) at the third location were evaluated in this stud y. There was no healthy tree of Valencia available in Immokalee grove. Generally, this study is comprised of two important parts. The first part includes pre processing and second includes processing. The h yperspe ctral images were Figure 7 5. Flow chart of methodology Imported raw images in ENVI 4.5 Dark subtraction Noisy band removal Build mask to remove background and unwanted objects Applied mask s on the images and obtained final image Computed vegetati on indices and statistical analysis Flat field calibration

PAGE 132

132 imported in remote sensing software ENVI 4.5 ( ITT Visual Information Solutions, Boulder, Colorado). Figure 7 5 shows the overall flow of the process that has been involved in this study. Preprocessing Preprocessing involved several steps Raw image acquired by hyperspectral camera is shown in Figure 7 6. Since noisy data was acquired between 306 nm to 420 nm, and bands after 870 nm, these bands were eliminated from the data and a total of 80 bands were used in the analysis. Now dark subtraction was applied by subtracting minimum band value. In the next step, image was calibrated with flat field method. Flat field method normalize ima ge with the area of known reflectance. In each image capturing spectralon white reference was included. By selecting the pixels of spectralon white panel in the image, hyperspectral images were calibrated. In the next step, background objects such as soi l, sky, grasses etc. were removed before data extraction. For isolating citrus leafy area, two masks were created. First mask was created with the help of unsupervised classification. Unsupervised classification classified images into various categories. T he mask based on the unwanted objects were created and applied on the calibrated image. Now still there are few objects remain in the image needs to remove before final processing.

PAGE 133

133 Figure 7 6. Raw image acquired from hyperspectral camera The region of interests (ROI) was selected for remaining unwanted objects and second mask was created (Figure 7 7). This mask was applied on the image obtained after applying first maskFinally image contains only leafy area of citrus is ready to analyze. ( Figure 7 8)

PAGE 134

134 Figure 7 7. Final mask for removing background as sky, soil, grass etc.

PAGE 135

135 Figure 7 8. Processed image used for data analysis Processing Processing part involved calculation of vegetation indices. List of all the vegetation indices, given in table 6 1, were evaluated in this study. Statistics for all individual bands computed. In this study band 970nm were replaced by 800 nm because of the noise Results and Discussion Statistical analysis was p erformed to see the difference between vegetati on indices of HLB infected tree and healthy trees. ANOVA were performed to find out whether the calculated vegetation indices were significantly different from one another among two class (healthy and HLB). Am ong the estimated vegetati on indices ( Table 6

PAGE 136

136 1 ), few vegetati on indices were found to be statistically significant to differentiate between HLB infected trees and healthy trees. Normalized d ifference v egetation i ndex (NDVI 1 & NDVI 2 ) simple ratio index (SR 1 & SR 2 ), m odified t riangular v egetation i ndex (MTVI 2 ) r enormalized d ifference v egetation i ndex (RDVI) m odified c hlorophyll a bsorption in r eflectance i ndex (MCARI 1 ) and s tructure i ntensive p igment i ndex (SIPI) were found significant from each other and could be utilized to discriminate HLB tress from healthy trees. Table 7 1. Means of vegetation indices of HLB infected and healthy trees showing statistically significant differen ce at =0.05 (Same letter in ro w shows no significant difference between healthy and HLB samples) Vegetati on indices HLB Healthy Normalized Difference Vegetation Index (NDVI 2 ) 0 .4 7 a 0.7 1 b Normalized Difference Vegetation Index ( NDVI 1 ) 0.49 a 0.72 b Simple Ratio Index (SR 2 ) 3. 75 a 6. 44 b Simple Ratio Index (SR 1 ) 3.97 a 6.80 a Modified Triangular Vegetation Index (MTVI 1 ) 0.83 a 0.94 a Modified Triangular Vegetation Index (MTVI 2 ) 0.50 a 0.72 b Renormalized Difference Vegetation Index (RDVI) 0.4 5 a 0.6 3 b Greenness Index (G) 1.73 a 1.99 a Triangular Vegetation Index (TVI) 32.16 a 37.00 a Modified Chlorophyll Absorption in Reflectance Index (MCARI 1 ) 0.83 a 0.94 a Structure Intensive Pigment Index (SIPI) 0.34 a 0.58 b

PAGE 137

137 Figure 7 9. Vegetati on indices of healthy and HLB trees Both tests concluded that m odified t riangular v egetation i ndex (MTVI 1 ) g reenness i ndex (G) t riangular v egetation i ndex (TVI) and m odified c hlorophyll a bsorption in r eflectance i ndex (MCARI 1 ) are not significant to discriminate HLB among healthy trees. All these i ndices suggested that reflectance at 870, 800, 670 and 570 nm are very critical. It seems that G index in insignificant to discriminate HLB trees from healthy trees because of the absence of NIR band. Though 870, 670 and 570 nm are critical wavebands thei r combination in vegetation indices are also very critical. It seems MTVI 1 TVI and MCARI 1 fail to discriminate HLB trees from healthy trees because of the difference of reflective values narrows the mean difference of HLB and healthy trees. Conclusions 40 Hyperspectral images (28 HLB, 12 Healthy) were evaluated for detection of HLB trees in field condition. They demonstrated good potential to discriminate HLB infected and healthy trees Since the sizes of hyperspectral images are large, 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 NDVI800 NDVI870 MTVI1 MTVI2 RDVI SIPI Vegetation Index Healthy HLB

PAGE 138

138 preprocessing consu mes longer time. Software of hyperspectral camera is also not very user friendly and have communication issues with computer. Normalized d ifference v egetation i ndex (NDVI 1 & NDVI 2 ) simple ratio index (SR 1 & SR 2 ), m odified t riangular v egetation i ndex (MTVI 2 ) r enormalized d ifference v egetation i ndex (RDVI) m odified c hlorophyll a bsorption in r eflectance i ndex (MCARI 1 ) and s tructure i ntensive p igment i ndex (SIPI) showed good potential to discriminate HLB infected trees. Future studies involve the evalua tion of the imaging and optical sensor for discriminating nutrient deficient tree and trees infected with other disease s The measurements in this study were performed on one cultivar of orange. It w ill be necessary to assess the hyperspectral images and t he classification algorithms for their performance with respect to other orange cultivars and citrus trees.

PAGE 139

139 C HAPTER 8 SUMMARY AND RECOMMEN DATIONS The goal of this research was to develop a technique for rapid detection of HLB infected trees in the field condition. In this study c anopy reflectance spectra w ere measured on infected and healthy trees using FieldSpec 3 spectroradiometer (Analytical Spectral Devices, Boulder, CO), SVC HR 1024 portable spectroradiometer (Spectra Vista Corporation, Po ughkeepsie, New York) multiband sensor (Applied Technology Stillwater, OK) and Specim hyperspectral camera (Autovision Inc., Los Angles, CA, USA) Various classification algorithms like, KNN, logistic regression, support vector machine, neural network and decision tree were used to classify the infected trees from the healthy ones. Chapter two describes the brief overview of all suspected diseases similar to HLB and nutrient deficiencies that may be confused with HLB. A very careful and rigorous inspectio n with trained people is required in the grove to identify HLB. Citrus trees also need to supply proper fertilizer and water for proper growth and to avoid nutritional deficiencies. This chapter explains chlorosis, types of chlorosis, causes of chlorosis, various mechanism of chlorosis due to diseases and nutritional deficiencies. It also discusses the diseases in citrus caused by vectors, virus, pathogens and post harvest decays. Chapter three identifies the critical wavelength that may be helpful in devel oping low cost sensor. Discriminability, derivative analysis and spectral ratio analysis were performed before applying machine learning techniques. Results were promising and supports our hypothesis that spectroscopy can detect HLB.

PAGE 140

140 Chapter four reports the partial least squares modeling and discriminant analysis to identify HLB in the field with ambient light and in greenhouse with artificial light. Results s howed that these techniques are promising in HLB detection for various varieties of citrus Overa ll, the full range of data gave more accurate results compared to narrow range with both techniques. However, the narrow range (400 nm to 900 nm) data gave better results with PLS modeling. In contrast, discriminant analysis was better in overall us e of th e full spectral range. It seems that the narrow range can produce very good results if the HLB symptoms are visible, but a major goal is to detect HLB before visible symptoms appear. Chapter five discusses application of visible NIR spectroscopy in HLB de tection. The goal of this study was to develop a technique for rapid detection of HLB infected citrus trees. Canopy reflectance spectra w ere measured on infected and healthy trees using a SpectraVista spectroradiometer Three machine learning techniques (K NN, logistic regression and support vector machine) were used to classify the infected trees from the healthy ones. The results concluded that a single measurement was insufficient for accurate detection of the infected trees. The classification error was between 18% and 35% using a single spectrum. However using multiple spectral measurements from a single tree, the classification accuracy increased significantly. SVM method showed an accuracy of higher than 95% when it was provided with five spectra from the same tree. Under real field conditions, varying sunlight and other environmental factors can produce noise that might reduce the classification accuracies. Under these conditions, multiple measurements will be necessary to ensure acceptable classifica tion accuracy.

PAGE 141

141 Chapter six illustrates the potential of multiband sensor to detect HLB in the field. The results indicated that the multi band optic sensor has a very good potential for detecting HLB infected citrus trees under field conditions. However, to achieve high classification accuracy, it requires several measurements from a single tree. The sensor can be incorporated with the scouting practice, to i ncrease the efficacy in HLB disease detection. Collect ing multiple readings using multi band sensor is easy and fast, and can be used by a human or an automated vehicle system. The results indicated that the multi band optic sensor used in this study has a very good potential for detecting HLB infected citrus trees under field conditions. However, to ob tain high classification accuracy, it is necessary to acquire multiple measurements from a single tree. The sensor can be integrated with the scouting practice, to improve the effectiveness in HLB disease detection. The measurements in this study were perf ormed on v alencia and mid sweet orange cultivars. It would be interesting and useful to know the performance of the sensor and the classification algorithms on other orange cultivars or other citrus types with similar diseases and nutrient deficiencies sym ptoms Chapter seven concluded the application of hyperspectral imaging to detect HLB in the field conditions. 40 hyperspectral images (28 HLB, 12 Healthy) were used in this study. They demonstrated good potential to discriminate HLB infected and healthy t rees Hyperspectral images require good storage space in computer and good processor to play with hyperspectral images. Normalized d ifference v egetation i ndex (NDVI 1 & NDVI 2 ) simple ratio index (SR 1 & SR 2 ), m odified t riangular v egetation i ndex (MTVI 2 ) r enormalized d ifference v egetation i ndex (RDVI) m odified c hlorophyll a bsorption in r eflectance i ndex (MCARI 1 ) and s tructure i ntensive p igment i ndex (SIPI)

PAGE 142

142 showed good potential to discriminate HLB infected trees. Future studies involve the evaluation of t he imaging and optical sensor for discriminating nutrient deficient tree s and trees infected with other disease s Several other classification algorithms can be used in future to see their effectiveness. The measurements in this study were performed on one cultivar of citrus It w ill be necessary to assess the hyperspectral images and the classification algorithms for their performance with respect to other citrus varieties.

PAGE 143

143 REFERENCE LIST Abadia, J. 1992. Leaf responses to Fe Deficiency: A Review. Journal of Plant Nutrition. 15(10): 1699 1713. Achor, D. S. and L. G. Albrigo. 2005. Biuret toxicity symptoms in citrus leaves mimics cell senescence rather than nutritional deficiency chlorosis. Journal of American Society for Horticultural Science. 130(5): 667 673. Anderson, J. M. and B. Anderson. 1982. The architecture of photosynthetic membranes: lateral and transverse organization. Trends Biochem. Sci. 7: 288 292. Blasco, J., N. Alexios, J. Gomez and E. Molto. 2007. Citrus sorting by identification of the most common defects using multispectral computer vision. Journal of Food Engineering. 83: 384 393. Boochs, F., G. Kupfer, K. Dockter and W. Kuhbauch. 1990. Shape of the red edge as vitalit y indicator for trees. International Journal of Remote Sensing. 11: 1741 1753. Borel, C. C. and A. W. Gerstl. 1994. Are leaf chemistry signatures preserved at the canopy level? IEEE. 996 998. Borengasser, M., T. R. Gottwald and T. Riley. 2001. Spectral Reflectrance of citrus canker. Proceedings of Florida State Horticulture Society. 114: 77 79. Bouffard, K. 2006. Greening found in 10 counties. Citrus Industry. 87(1): 5 26. Bravo, C., D. Mosho u, J. West, A. McCartney and H. Ramon. 2003. Early disease detection in wheat fields using spectral reflectance. Biosystems Engineering. 84(2): 137 145. Bravo, C., M. Dimitrios, R. Oberti, J. West, M. Alastair, B. Luigi and H. Ramon. 2004. Foliar disease detection in the field using optical sensor fusion. Agricultural Engineering Internatinal: the CIGR Journal of Scientific Research and Development VI. Brlansky, R. H., K. R. Chung and M. E. Rogers. 2007. Florida Citrus Management Guide: Huanglongbing (Cit rus Greening). University of Florida. IFAS Extension. PP 225. Butler, W. L. and D. W. Hopkins. 1970. Higher derivatives analysis of complex absorption spectra. Photochemistry and Photobiology. 122:439 450. Bove, J. M. 2006. Huanglongbing: a destructive, newly emerging century old disease of citrus. Journal of Plant Pathology 88(1):7 37.

PAGE 144

144 Carter, G. A. and L. Estep. 2002. General spectral characteristics of leaf reflectance response to plant stress and their manifestation at the landscape scale. In ed. Ano nymous Carter, G. A. and A. K. Knapp. 2001. Leaf optical properties in higher plants: Linking spectral characteristics to stress and chlorophyll concentration. American Journal of Botany. 88(4): 677 684. Carter, G. A. and D. C. McCain. 1993. Relation ship of leaf spectral reflectance to chloroplast water content determined using NMR microscopy. Remote Sensing of Environment. 46:305 310. Cherkassky, V. and F. Mulier. 2007. Learning from Data Concepts, Theory, and Methods. Second ed. Hoboken, NJ: John Wiley & Sons, Inc. Chung, K. R. and R. H. Brlansky. 2005. Citrus Diseases Exotic to Florida: Huanglongbing (Citrus Greening). University of Florida. IFAS Extension. PP 210. Cibula, W. G. and G. A. Carter. 1992. Identification of a far red reflectance res ponse to ectomycohizae in slash pine. International Journal of Remote Sensing. 13239 247. Coppin, B. 2004. Artificial Intelligence Illuminated. Sudbury, MA: Jones and Bartlett Publishers, Inc. Costa, G., M. Noferini, G. Fiori and F. Spinelli. 2007. Innov ative application of non destructive techniques for fruit quality and disease diagnosis. Proceedings Vith IS on Kiwifruit, Acta Hort iculture 753 Curran, P. J. 1989. Remote Sensing of Foliar Chemistry. Remote Sensing of Environment. 30:271 278. da Graca J. V. 1991. Citrus greening disease. Phytopathology. 29:109 136. Delwiche, S. R. and R. A. Grayboscht. 2002. Identification of waxy wheat by Near infrared reflectance spectroscopy. Journal of Cereal science. 35:29 38. Demetriades Shah, T. H., M. D. Ste ven and J. A. Clark. 1990. High resolution derivative spectra in remote sensing. Remote Sensing of Environment .33:55 64. Dobrowski, S. Z., J. C. Pushnik, T. Zacro and S. L. Ustin. 2005. Simple reflectance indices track heat and water stress induced chang es in steady state chlorophyll flurescence at canopy scale. Remote Sensing of Environment. 97:403 414. Duda, R. O., P. E. Hart and D. G. Stork. 2000. Pattern Classification. Second edition ed. New York, N.Y: John Wiley and Sons.

PAGE 145

145 Elmasry, G., N. Wang, C. Vigneault, J. Qiao and A. Elsayed. 2008. Early detection of apple bruises on different background colors using hyperspectral imaging. LWT 41:337 345. Erwin, D. C. and O. K. Ribeiro. 1996. Phytophthora Diseases Worldwide. APS Press. St. Paul, Minnesota. Etxeberria, E., P. Gonzalez and Dawson, W. and Spann, T. 2007. An Iodine based starch test to assist in selecting leaves for HLB testing. University of Florida. IFAS Extension Etxeberria, E., P. Gonzalez, D. Achor and L. G. Albrigo. 2009. Anatomical dist ribution of abnormally high levels of starch in HLB affected Valencia oranges trees. Physiological and Molecular Plant Pathology doi:10.1016/j.pmpp.2009.09.004. FASS. 2008. Florida Agriculture Statistics Survey. Fouche, P. S. 1999. Detecting nitrogen de ficiency on irrigated cash crops using remote sensing methods. South African Journal of Plant and Soil. 1659 63. Fukunaga, K. 1990. Introduction to statistical pattern recognition, second edition. San Diego, CA, Academic Press. Futch, S. H., S. Weingarten and M. Irey. 2009. Determining greening infection levels using multiple survey methods in Florida citrus. Proceedings of Florida State Horticulture Society 122. Gaffney, J. J. 1972. Reflectance properties of citrus fruits. Tran sactions of the ASABE: 310 314. Gamon, J. A., C. B. Field, W. Bilger, O. Bjrkman, A. L. Fredeen and J. Peuelas. 1990. Remote sensing of the xanthophyll cycle and chlorophyll fluorescence in sunflower leaves and canopies. Oecologia. 85(1): 71 76. Garnie r, M. and J. M. Bov. 1983. Transmission of the organisms associated with citrus greening disease from sweet orange to periwinkle by dodder. Phytopathology. 73:1358 1363. Garnier, M., E. S. Jagoueix, P. R. Cronje, H. F. Le Roux and J. M. Bove. 2000. Genom ic characterization of a Liberibacter present in an ornamental rutaceous tree, Calodendrum capense in the Western Cape province of South Africa. Proposal of 'Candidatus Liberibacter africanus subsp. capensis'. Int. J. System. Evol. Microbiol. 50: 2119 212 5. Gates, D. M., H. J. keegan, J. C. Schleter and V. R. Weider. 1965. Spectral properties of plants. Applied optics. 4(1): 11 20.

PAGE 146

146 Gepstein, S. 1988. Photosynthesis. Senescence and aging in plants. 85 109. Gitelson, A. A. and M. N. Merzylayak. 1997. Remo te estimation of chlorophyll content in higher tree leaves. International Journal of Remote Sensing. 18: 2691 2697. Gomez, A. H., Y. He and A. G. Pereira. 2006. Nondestructive measurement of acidity, soluble solids and firmness of Satsuma mandarin using V IS/NIR spectroscopy techniques. Journal of Food Engineering. 77: 313 319. Graeff, S., J. Link and W. Claupein. 2006. Identification of powdery mildew ( Erysiphe graminis sp. tritici ) and take all disease ( Gaeumannomyces graminis sp. tritici ) in wheat ( Trit icum aestivum L.) by means of leaf reflectance measurements. Central European Journal of Biology. 1: 275 288. Green, D. E., L. L. Burpee and K. L. Stevenson. 1998. Canopy reflectance as a measure of disease in Tall Fescue. Crop science. 38:1603 1613. Guo, T., L. Guo, X. Wang and M. Li. 2009. Application of NIR spectroscopy inclassification of plant species. First International Workshop on Education Technology and Computer Science 3: 879 883 Halbert, S. E. and K. L. Manjunath. 2004. Asian citrus psyllids (Sternorrhyncha:Phyllidae) and greening disease of citrus: A literature review and assessment of risk in Florida. Florida Entomologist. 87(3): 330 353. Hall, J. D., R. Barr, A. H. Al Abbas and F. L. Crane. 1972. The ultrastructure of chloroplasts in mine ral efficient maize leaves. Plant Physiology. 50: 404 409. Haykin, S. 1998. A Comprehensive Foundation. second edition. Upper Saddle River, NJ, Prentice Hall. Horler, D. N. H., M. Dockray and J. Barber. 1983. The red edge of plant leaf reflectance. Inter national Journal of Remote Sensing. 4: 273 288. Hsu, C. W., C. C. Chang and C. J. Lin. 2008. A practical guide to support vector classification. Technical Report. Huang, J. F. and A. Apan. 2006. Detection of Sclerotinia rot disease on Celery Using hypers pectral Data and partial Least Squares regression. Journal of Spatial Science. 51(2): 129 142. Kane, K. E. and W. S. Lee. 2006. Spectral sensing of different citrus varieties for precision agriculture. In Proceeding of ASABE annual meeting, July 9 12, Po rtland, OR ASABE paper no: 061065

PAGE 147

147 Kim, M. S., Y. R. Chen and P. M. Mehl. 2001. Hyperspectral reflectance and fluorescence imaging system for food quality and safety. Transactions of the ASABE. 44(3): 721 729. Kobayashi, T., E. Kanda, K. Ishiguro and Y. Torigoe. 2000. Phytopathology. 91(3): 316 323. Larose, D. T. 2006. Data mining methods and models. Hoboken, NJ: John Wiley & Sons, Inc. Larsolle, A. and H. H. Muhammed. 2007. Measuring crop status using multivariate analysis of hyperspectral field reflect ance with application to disease severity and plant density. Precision Agriculture. 8: 37 47. Lee, W. S., M. R. Ehsani and L. G. Albrigo. 2008. Citrus greening (Huanglongbing) detection using aerial hyperspectral imaging. In proceeding of 9th Internationa l conference on Precision Agriculture July 20 23, Denver, Colorado, USA Liu, Z., J., J. S. Huang, R. Tao, W. Zhou and L. Zhang. 2007. Characterizing and estimating rice brown spot disease severity using stepwise regression, principal component regressi on and partial least square regression. Journal of Zhejiang University Science B. 8(10): 738 744. Lopez, M. M., E. Bertolini, A. Olmos, P. Caruso, M. T. Gorris, P. Llop, R. Penyalvey and M. Cambra. 2003. Innovative tools for detection of plant pathogenic viruses and bacteria. Internatonal Microbiology. 60: 233 243. Lorenzen, B. and A. Jensen. 1989. Changes in leaf spectral properties induced in barley by cereal powdery mildew. Remote Sensing of Environment. 27(2): 201 209. Lu, R. 2003. Detection of bruis es on apples using near infrared hyperspectral imaging. Transactions of ASAE. 46(2): 1 8. Mahesh, S., A. Manickavasagan, D. S. Jayas, J. Paliwal and N. D. G. White. 2008. Feasibility of near infrared hyperspectral imaging to differentiate Canadian wheat classess. Biosystems Engineering 101: 50 57. Malthus, T. J. and A. C. Madeira. 1993. High r esolution spectroradiometry: spectral reflectance of field bean leaves infected by Botrytis fabae Remote Sensing of Environment. 45(1): 107 116. Manjunath, K. L., S. Halbert, C. Ramadugu, S. Webb and R. F. Lee. 2008. Detection of Candidatus Liberibacter asiaticus in Diaphorina citri and its importance in the management of citrus huanglongbing disease. Phytopathology. 98: 387 396.

PAGE 148

148 Matile, P. 1992. Chloroplasts senescence. 413 525. In: B. R. Baker and H. Thomas (eds.). Crop Photosynthesis: Spatial and tem poral determinants Elsevier, Amsterdam, the Netherlands. Mitchell, T. M. 1997. Machine Learning. First ed. New York, NY: McGraw Hill. Moshou, D., C. Bravo, S. Wahlen, J. West, A. McCartney, J. D. Baerdemaeker and H. Ramon. 2006. Simultaneous identificati on of plant stresses and diseases in arable crops using proximal optical sensing and self organising maps. Precision Agriculture. 7: 149 164. Muraro, R. P. 2007. Summary of 2006 07 citrus budget for the central Florida (Ridge) production region. Nagoak a, T., N. S. Eikje, A. Nakamura, K. Aizawa, Y. Kiyohara, F. Ichikawa, T. Yamazaki, M. Doi, K. Nakamura, S. Otsubo and T. Sota. 2007. Inspection of skin hemodynamics with hyperspectral camera. In Proceedings of the 29th Annual International Conference of th e IEEE EMBS, August 23 26, Cit Internationale, Lyon, France NASS. 2009. National Agriculture Statistics Survey. Nicolai, B. M., E. Lotze, A. Peirs, N. Scheerlinck and K. I. Theron. 2006. Non destructive measurement of bitter pit in apple fruit using NIR hyperspectral imaging. Postharvest Biology and Technology. 40:1 6. Oberbacker, M. F. 1954. A chlorosis of citrus produced by biuret as an impurity in urea. In Proceedings of Florida State Horticultural Society 67 Orfanidis, S. J. 1996. Introduction to S ignal Processing. Englewood Cliffs, NJ: Prentice Hall. Platt, J. 1998. Fast Training of Support Vector Machines using Sequential Minimal Optimization. In Advances in Kernel Methods Support Vector Learning eds. B. Scholkopf, C. Burges, and A. Smola, 41 65. Polikar, R. 2006. Ensemble based systems in decision making. IEEE Circuits and Systems Magazine. 21 44. Polischuk, V. P., T. M. Shadchina, T. I. Kompanetz, I. G. Budzanivskaya and A. A. Sozinov. 1997. Changes in reflectance spectrum characteristic of nicotiana debneyi plant under the influence of viral infection. Archives of Phytopathology and Plant Protection 31:115 119. Pontius, J., R. Hallett and M. Martin. 2005. Near Infrared Spectroscopy: Indices comparison and algorithm development. Applied Spe ctroscopy. 59(6): 836 843.

PAGE 149

149 Ramon, H., J. Anthonis, E. Vrindts, R. Delen and J. Reumers. 2002. Development of a weed activated spraying machine for targeted application of herbicides. In: Aspects of Applied Biology, International Advances in Pesticide Appl ication, Wellesbourne, UK: Assoc. ,Applied. Biology. 66: 147 162 Reddy, G. S., C. L. N. Rao, L. Venkatratnam and P. V. K. Rao. 2001. Influence of plant pigments on spectral reflectance of maize, groundnut and soybean grown in semi arid environments. International Journal of Remote Sensing. 122(17): 3337 3380. Riedell, W. E. and T. M. Blackmer. 1999. Leaf Reflectance Spectra of Cereal Aphid Damaged Wheat. Crop science. 39:1835 1840. Roggo, Y., L. Duponchel and J. P. Huvenne. 2003. Application to qual itative analysis of sugar beet by near infrared spectroscopy. Analytica Chimica Acta 477:187 200. Roujean, J. L. and F. M. Breon. 1995. Estimating PAR absorbed by vegetation from bidirectional reflectance measurements. Remote Sensing of Environment .51(3) : 375 384. Rundquist, D., L. Han, J. Schalles and J. Peake. 1996. Remote measurement of algal chlorophyll in surface waters: the case for the first derivative of reflectance near 690 nm. Phogrammetric Engineering and Remote Sensing. 62(2): 195 200. Salgi o, P., M. L'Hospital, D. Lafleche, G. Dupont, J. M. Bove, J. G. Tully and E. A. Freundt. 1973. Spiroplasma citri gen. and n.: a mycoplasma like organism associated with 'stubborn' disease of citrus. International Journal of Systematic Bacteriology 23(3):1 91 204. Sankaran, S., A. Mishra, R. Ehsani and C. Dima. 2010. A review of advanced techniques for detecting plant diseases. Computer and Electronics in Agriculture 72(1): 1 13. Savitzky, A. and M. J. E. Golay. 1964. Smoothing and differentiation of data by simplified least square procedure. 36(8): 1627 1639. Schneider, H. 1968. Anatomy of greening diseased sweet orange shoots. Phytopathology 58: 1155 1160. Schubert, T. S., T. R. Gottwald, S. A. Rizvi, J. H. Graham, X. Sun and W. N. Dixon. 2001. Meeting the challenge of Eradicting citrus canker in Florida Again. Plant diseas.e 85(4): 340 356. Shackelford, S. D., T. L. Wheeler and M. Koohmaraie. 2004. Development of optimal protocol for visible and near infrared reflectance spectroscopic evaluation of mea t quality. Meat Science. 68: 371 381.

PAGE 150

150 Sighicelli, M., F. Colaco, A. Lai and S. Patsaeva. 2009. Monitoring post harvest orange fruit disease by fluressence and reflectance hyperspectral imaging. In Proc. Ist IS on Hort in Europe, Acta Hort., ISHS. Smith, K. L., M. D. Steven and J. J. Collis. 2005. Plant spectral responses to gas leaks and other stresses. International Journal of Remote Sensing. 26(18): 4067 4081. Stall, R. E., J. W. Miller, G. M. Marco and B. I. Echenique. 1981. Timing of sprays to control cancrosis of grapefruit in Argentina. Proceedings of International Society of Citriculture. 1:414 417. Stewart, I. and C. D. Leonardo. 1952. The cause of yellow tipping in citrus leaves. I n Proceedings of Florida State Horticultural Society. 25 Stone, M. L., J. B. Solie, R. W. Whitney, W. R. Raun and H. L. Lees. 2010 Sensors for Detection of Nitrogen in Winter Wheat. Stone, M. L., J. B. Solie, W. R. Raun, R. W. Whitney, S. L. Taylor and J. D. Ringer. 1996. Use of spectral radiance for correctin g in season fertilizer nitrogen deficiencies in winter wheat. Transactions of the ASAE. 39(5): 1623 1631. Strang, G. 1991. Calculus. Wellesley, MA: Wellesley Cambridge Press. Sundaram, J. and C. V. Kandala. 2009. Application of near infrared spectroscopy to peanut grading and quality analysis: overview. Sens. & Instrumen. Food Qual. 3: 156 164. Takushi, T., T. Toyozato, S. Kawano, S. Taba, A. Ooshiro and M. Numazawa. 2007. Starch method for simple, rapid diagnosis of citrus Huanglongbing using iodine to detect high accumulation of starch in citrus leaves. Japanese Journal ofF Phytopathology 73: 3 8. Tallada, J. G., M. Nagata and T. Kobayashi. 2006. Detection of Bruises in Strawberries by hyperspectral imaging. In Proceeding of ASABE annual meeting, July 9 12, Portland, OR ASABE paper no: 06 3014. Terry, N. 1980. Limiting factors in photosynthesis I. Use of iron stress to control photochemical capacity in vivo. Plant Physiology. 65:114 120. Thomas, H. and J. L. Stoddart. 1980. Leaf Senescence. Annual Revi ew of Plant Physiology 3:183 111. Thomas, J. R. and G. F. Oerther. 1972. Estimating nitrogen content of sweet pepper leaves by reflectance measurements. Agronomy Journal. 83: 926 928.

PAGE 151

151 Thompson, J. E., G. Paliyath, J. H. Brown and C. L. Duxbury. 1987. In: Plant Senescence: Its Biochemistry and Physiology, eds.,Thomson, W. W., Nothnagel, E. A. & Huffaker, R. C. The American Society of Plant Physiologists, Rockville, MD 146 155. Timmer, L. W. and S. E. Zitko. 1997. Evaluation of fungicides for control of Al ternaria brown spot and citrus scab. In Proceedings of Florida State Horticultural Society 71 76. Timmer, L. W. and S. E. Zitko. 1996. Evaluation of copper fungicides and rates of metallic copper for control of melanose on grapefruit in Florida. Plant Disease 80166 169. Timmer, L. W., S. M. Garnsey and J. H. Graham. 2000. Compendium of Citrus Disease. Second edition ed. APS Press. St. Paul, Minnesota. Vesk, M., J. V. Possingham and F. V. Mercer. 1966. The effect of mineral nutrient deficiencies on the structure of the leaf cells on tomato, spinach and maize. Australian Journal of Botany. 141 18. Wang D., M. S. Ram and F. E. Dowell. 2002. Classification of damaged soybean seeds using near infrared spectroscopy. Transactions of ASAE. 45(6): 1943 1948. Wang, W., C. Thai, C. Li, R. Gitaitis and E. W. Tollner. 2009. Detecting of sour skin diseases in Vid alia sweet onions using near infrared hyperspectral imaging. In Proceeding of ASABE annual meeting, Ju ne 21 24 Reno NV, USA ASABE paper no: 096364. Webb, A. R. Statistical Pattern Recognition. Second ed. Chichester, England: John Wiley & Sons Ltd. Wes t, J. S., C. Bravo, R. Oberti, D. Lemaire, D. Moshou and H. A. McCartney. 2003. The potential of optical canopy measurement for targeted control of field crop disease. Annual Review of Phytopathology 41:593 614. Whiteside, J. O. 1983. Timing of spray trea tments for citrus greasy spot control. Proceedings of Florida State Horticultural Society 96:17 21. Whiteside, J. O. 1976. A newly recorded Altenaria induced brown spot disease on Dancy tangerines in Florida. Plant disease Reporter 60:326 329. Witten, I. H. and E. Frank. 2005. Data Mining Practical Machine Learning Tools and Techniques. Second ed. San Francisco, CA: Morgan Kaufmann Publishers. Wittenbach, V. A. 1982. Effect of pod removal on leaf senescence in soybeans. Plant Physiology 70: 1544 1548.

PAGE 152

152 Woolhouse, H. W. 1984. The biochemistry and regulation of senescence in chloroplast. Canadian Journal of Botany 62: 2934 2942. Wu, D., L. Feng, C. Zhang and Y. He. 2008. Early detection of Botrytis Cinerea on eggplant leaves based on visible and near infr ared spectroscopy. Transactions of the ASABE 51(3): 1133 1139. Xing, J. and J. D. Baerdemaeker. 2005. Bruise detection on 'Jonagold' apples using hyperspectral imaging. Postharvest Biology and Technology 37:152 162. Xing, J., C. Bravo, P. T. Jancsok, H. Ramon and J. D. Baerdemaeker. 2005. Detecting bruises on golden delicious apples using hyperspectral imaging with multiple wavebands. Biosystems Engineering. 90(1): 27 36. Yates, J. D., S. H. Futch and T. M. Spann. Sc outing for Citrus Greening. University of Florida. IFAS Extension HS 1147:1 2. Zacro tejada, P. J., S. L. Ustin and M. L. Whiting. 2005. Temporal and spatial relationship between within field yield variabilty in cotton and high spatial hyperspectral remot e sensing imagery. Agronomy Journal. 97(3): 641 653. Zhang, C., Y. Shen, J. Chen, P. Xiao and J. Bao. 2008. Nondestructive prediction of total phenolics flavonoid contants and antioxident capacity of rice grain using near infrared spectroscopy. Journal of Agricultural and Food Chemistry. 56: 8268 8272. Zhang, L. and G. W. Small. 2002. Calibration Standardization: Algorithm for Partial Least Squares Regression: Application to the Determination of Physiological Levels of Glucose By Near Infrared Spectroscop y. Analytical chemistry. 7:44097 4108. Zhang, M., Z. Qin and X. Liu. 2005. Remote Sensed Spectral Imagery to detect late blight in field tomatoes. Precision Agriculture. 6:489 508. Zhang, M., X. liu and M. O'Nell. 2002. Spectral discrimination of phytoph thora infests infection on tomatos based on principal component and cluster analysis. International Journal of Remote Sensing. 23(6): 1095 1107. Zhao, D., K. R. Reddy, V. G. Kakani, J. J. Read and G. A. Carter. 2003. Corn ( Zea mays L. ) growth, leaf pigmen t concentration, photosynthesis, and leaf hyperspectral reflectance properties affected by nitrogen supply. Plant and Soil 257:205 217.

PAGE 153

153 BIOGRAPHICAL SKETCH Ashish Ratn Mishra was born in Allahabad city in the state of Uttar Pradesh, India, in 1980. He graduated from Allahabad Agricultural Institute Deemed University, moved to University of Arkansas, Fayetteville, Arkansas, United States to pursue his gradu ate studies, in 2003. He graduated with a Master of Science degree in agricultural and biological engineering in 2005. To continue his higher education he joined the Ph.D. eng ineering, where he specialized in hyperspectral imaging, spectroscopy and geographical information systems (GIS).