Citation
Spectral-Based Nitrogen Sensing for Citrus

Material Information

Title:
Spectral-Based Nitrogen Sensing for Citrus
Creator:
MIN, MIN
Copyright Date:
2008

Subjects

Subjects / Keywords:
Calibration ( jstor )
Datasets ( jstor )
Leaves ( jstor )
Moisture content ( jstor )
Nitrogen ( jstor )
Pixels ( jstor )
Reflectance ( jstor )
Spectral correlation ( jstor )
Spectral reflectance ( jstor )
Wavelengths ( jstor )
City of Gainesville ( local )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Min Min. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
5/31/2008
Resource Identifier:
496180310 ( OCLC )

Downloads

This item is only available as the following downloads:


Full Text

PAGE 1

SPECTRAL-BASED NITROGEN SENSING FOR CITRUS By MIN MIN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2006

PAGE 2

Copyright 2006 by Min Min

PAGE 3

I would like to dedicate this work to my parents for their limitless support.

PAGE 4

iv ACKNOWLEDGMENTS First and foremost I would like to express sincere gratitude to my major advisor, Dr. Won Suk “Daniel” Lee, for his consis tent guidance, encouragement and support throughout this research project. His motiv ation, dedication, and enthusiasm towards research were essential for successful comple tion of this body of work. Without his help, it would not have been possible to complete this research work. I would also give my thanks to my supe rvisory committee, Dr. Thomas F. Burks, Dr. Jonathan D. Jordan, Dr. Arnold Schumann, and Dr. Huikai Xie, for their guidance and suggestions to complete this work. Special thanks go to Dr. Ismail Bogrekci for his inspiring and helpful discussion. His guidance and assistance have been very helpful for this project. I also wish to thank the staff of th e Agricultural and Biological Engineering Department, especially Mr. Harmon Pearson fo r his assistance in leaf sampling and other lab support, and Mr. Michael Zingaro for hi s help in fabrica ting sensor parts. Finally, I would like to thank my family and friends for their continued support and care throughout this m ilestone in my life.

PAGE 5

v TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iv LIST OF TABLES............................................................................................................vii LIST OF FIGURES...........................................................................................................ix ABSTRACT.....................................................................................................................xi ii CHAPTER 1 INTRODUCTION........................................................................................................1 Citrus in Florida............................................................................................................2 Nitrogen Consumption for Citrus and Environmental Issues.......................................3 Precision Agriculture for Citrus....................................................................................5 2 OBJECTIVES...............................................................................................................8 3 DETERMINATION OF SIGNIFICANT WAVELENGTHS AND PREDICTION OF NITROGEN CONTENT FOR CITRUS................................................................9 Introduction................................................................................................................... 9 Objectives...................................................................................................................11 Materials and Methods...............................................................................................11 Leaf Sampling and Reflectance Measurement....................................................11 Normalization of the Reflectance Spectra...........................................................13 Determination of Important Wavelengths...........................................................16 Correlation coefficient spectrum..................................................................16 Standard deviation spectrum........................................................................17 Stepwise multiple linear regression (SMLR)...............................................17 Partial least squares (PLS) regression..........................................................18 Results and Discussion...............................................................................................20 Chemical Analyses of Sample Leaves................................................................20 Wavelength Selection and Calib ration Models Development.............................21 Correlation coefficient spectra.....................................................................21 Standard deviation spectra...........................................................................22 Stepwise multiple linear regression (SMLR)...............................................24 Partial least squares (PLS) regression..........................................................25

PAGE 6

vi Summary.....................................................................................................................30 4 WATER EFFECT ON NITROGEN PREDICTION FOR CITRUS LEAF...............32 Introduction.................................................................................................................32 Objectives...................................................................................................................34 Materials and Methods...............................................................................................34 Results and Discussion...............................................................................................35 Water Contents Distribution................................................................................35 Water Effect on Spectra Characteristics..............................................................35 Water Effect on Nitrogen Prediction...................................................................37 Summary.....................................................................................................................42 5 NITROGEN SENSING SYS TEM FOR CITRUS LEAF..........................................43 Introduction.................................................................................................................43 Objective.....................................................................................................................4 5 Design Criteria............................................................................................................45 Wavelength Selection..........................................................................................45 No Moving Part...................................................................................................45 Single Leaf Detection..........................................................................................47 Diffuse Reflectance Measurement......................................................................48 Sensor Structure and Component...............................................................................49 Data Acquisition System............................................................................................52 Sensor Testing Results in Lab Environment..............................................................55 Wavelength Calibration.......................................................................................55 Dark Current and Signal Output..........................................................................61 Noise and Signal to Noise Ratio..........................................................................63 Linearity..............................................................................................................65 Stability................................................................................................................69 Leaf Measurement...............................................................................................70 Development of Calibration Models..........................................................................72 Leaf Sampling and Spectra Measurement...........................................................72 Statistical Results of Data Set Measured by Spectrophotometer........................73 Statistical Results of Data Set Measured by Nitrogen Sensor.............................75 Discussion...................................................................................................................77 Summary.....................................................................................................................80 6 FUTURE WORK........................................................................................................81 LIST OF REFERENCES...................................................................................................82 BIOGRAPHICAL SKETCH.............................................................................................86

PAGE 7

vii LIST OF TABLES Table page 3-1 Results of N concentration analysis of the samples from five different N application rates for data set 2002............................................................................21 3-2 Results of N concentration analysis of the samples from five different N application rates for data set 2003............................................................................21 3-3 SMLR analysis result for data sets 2002 and 2003..................................................26 3-4 Percent variation explained by the seve n factors in the PLS regression for data set 2002....................................................................................................................27 3-5 Percent variation explained by the nine factors in the PLS re gression for data set 2003..........................................................................................................................2 7 3-6 Results for the data sets 2002 and 2003 using PLS regression................................29 4-1 Results of N concentration and water c ontent analysis of th e samples from the blocks with five different N application rates in 2002.............................................37 4-2 Results of N concentration analysis a nd water content of the samples from the blocks with five different N application rates in 2003.............................................37 4-3 Prediction of water content using absorbance spectra with PLS and SMLR procedures for data set 2002.....................................................................................38 4-4 Prediction of water content using absorbance spectra with PLS and SMLR procedures for data set 2003.....................................................................................38 4-5 N prediction by PLS using two predic tors (water and N contents) and one predictor (Dry basis N or we t basis N) for data set 2002.........................................41 4-6 N prediction by PLS using two predic tors (water and N contents) and one predictor (Dry basis N or we t basis N) for data set 2003.........................................41 4-7 SMLR analysis resu lts of dataset 2002....................................................................42 4-8 SMLR analysis resu lts of dataset 2003....................................................................42 5-1 Predicted wavelengths for the VIS detector using pixel number.............................59

PAGE 8

viii 5-2 Predicted wavelengths for the NIR detector using pixel number.............................60 5-3 Results of N concentration analysis of the samples from five different N application rates for samples collected in 2005.......................................................73 5-4 SMLR and PLS analysis result for data sets 2005 measured by Cary 500..............75 5-5 SMLR and PLS analysis results for da ta sets 2005N measured by N sensor..........77

PAGE 9

ix LIST OF FIGURES Figure page 1-1 The basic structure of a chlorophyll.........................................................................1 3-1 Spectral response of a citrus leaf, with and without a Teflon sheet.......................13 3-2 Normalized spectrum of the same leaf shown in figure 3-1 compared to the original spectrum...................................................................................................15 3-3 Absorbance spectra of two di fferent citrus leaf samples.......................................16 3-4 Distribution of N concentration for year 2002 and 2003. N ranges were 19.933.8 g kg-1 for data set 2002, and 24.9-35.3 g kg-1 for data set 2003....................22 3-5 Correlation coefficients between abso rbance at each wavelength and leaf N concentration of the calibra tion data set for year 2002..........................................23 3-6 Correlation coefficients between abso rbance at each wavelength and leaf N concentration of the calibra tion data set for year 2003..........................................23 3-7 Standard deviation of absorbance for calibration data set of year 2002................24 3-8 Standard deviation of absorbance for calibration data set of year 2003................24 3-9 N concentration prediction using SMLR for data set 2002. This method generated R2 = 0.816 and RMSD = 1.32 g kg-1 for the validation data set............26 3-10 N concentration prediction using SMLR for data set 2003. This method generated R2 = 0.550 and RMSD = 1.31 g kg-1 for the validation data set............26 3-11 B coefficient determined from the tr aining data set using PLS regression with seven factors for data set 2002...............................................................................27 3-12 B coefficient determined from the trai ning data set using PLS with nine factors for data set 2003.....................................................................................................28 3-13 N concentration prediction using PLS for data set 2002. This method generated R2 = 0.828 and RMSD = 1.22 g kg-1 for the validation data set............................29 3-14 N concentration prediction using PLS for data set 2003. This method generated R2 = 0.597 and RMSD = 1.20 g kg-1 for the validation data set............................30

PAGE 10

x 4-1 Distribution of water contents for da ta set 2002. Four outli ers (circled) were removed for data analysis......................................................................................36 4-2 Distribution of water contents fo r data set 2003. One outlier (circled) was removed for data analysis......................................................................................36 4-3 Reflectance spectra of fresh, dried sa mple and their difference spectrum with water content of 0.607 g g-1...................................................................................38 4-4 Correlation coefficient between absorbance and N (g kg-1) of dried samples of training data sets of 2002 and 2003.......................................................................40 4-5 Correlation coefficient between absorbance and N (g kg-1) of fresh samples of training data sets of 2002 and 2003.......................................................................40 5-1 Selected wavelengths range for N de tection. Chlorophyll related wavelengths are concentrated in the visible range, and protein relate d wavelengths are concentrated in the near-infrared range . ................................................................46 5-2 A spectrometer built with a linear variable filter (LVF) and a detector array to avoid moving parts.................................................................................................47 5-3 Model for diffuse reflectance measurement..........................................................49 5-4 Design scheme of the N sensing system................................................................51 5-5 A picture of the actual N sensor.............................................................................51 5-6 Light intensity of the light source. It covers wavelength range of 400-2600 nm, and has peak intensity at 950 nm...........................................................................52 5-7 Spectral sensitivity for (a) VIS det ector S8377-256Q and (b) NIS detector G9208-256W..........................................................................................................52 5-8 Timing chart for the VIS detector..........................................................................53 5-9 The LabView data acquisition bl ock diagram for VIS detector............................54 5-10 Data acquisition interfac e for the VIS detector......................................................54 5-11 Data acquisition interface for the NI R detector developed by Hamamatsu, Inc...55 5-12 Wavelength calibration for the VIS detector using Cary 500................................56 5-13 Pixel response for different wa velength for the VIS detector................................57 5-14 Relationship between monochrome light and pixel number. Linearity of the VIS detector is 0.999..............................................................................................57

PAGE 11

xi 5-15 Errors between predicted wavelengths and actual wavelengths. The error is less than 2 nm................................................................................................................57 5-16 Wavelength calibration for the NIR de tector. Pixel respons es were plotted according to different wavelength.........................................................................58 5-17 Relationship between monochrome light and pixel number for the NIR detector. The linearity is 0.999..............................................................................58 5-18 Errors between predicted wavelengths and actual wavelengths. The error is less than 5 nm................................................................................................................58 5-19 Dark voltage output of the VIS dete ctor using 130 ms integration time...............61 5-20 Dark voltage output of the NIR dete ctor using 100 ms integration time...............62 5-21 Signal output of leaf measurement by th e VIS detector subtracting the dark.......62 5-22 Signal output of leaf measurement by th e NIR detector subtracting the dark.......62 5-23 Noise for the VIS detector calculated us ing ten measurements of a single leaf....64 5-24 A reflectance spectrum of a leaf measurement......................................................64 5-25 Noise for the NIR detector calculated us ing ten measurements of a single leaf....64 5-26 Signal to noise ratio for the VIS detector...............................................................65 5-27 Signal to noise ratio for the NIR detector..............................................................65 5-28 Voltage outputs for one leaf measured by the VIS detector with eight different integration times.....................................................................................................67 5-29 Linearity for the VIS detector for leaf measurement.............................................67 5-30 Regression line for pixel 70. The r is 0.999...........................................................67 5-31.Residual for predicted voltage output s for pixel 70 of the VIS detector...................68 5-32 Voltage outputs for white reference (PTF E) measured by the VIS detector with five different in tegration times...............................................................................68 5-33 Linearity for the VIS detector for white reference measurement..........................68 5-34 Voltage outputs for one leaf measured by the NIR detector with seven different integration times.....................................................................................................69 5-35 Linearity for the NIR detector................................................................................69

PAGE 12

xii 5-36 Stability test of the VIS detector during two hours period. Ratios were obtained by dividing each successive meas urement to the first scan...................................70 5-37 Stability test of the NIR detect or during two hours period. Ratios were obtained by dividing each successive m easurement to the first scan....................70 5-38 Reflectance spectrum of a citrus l eaf measured by the VIS detector....................71 5-39 Reflectance spectrum of a citrus l eaf measured by the NIR detector....................71 5-40 Actual citrus leaf reflectance spec trum measured by the spectrophotometer........72 5-41 Flow chart for N sensing system design which was completed in two flow cycles......................................................................................................................73 5-42 Correlation coefficient between ab sorbance of data set 2005 obtained by spectrophotometer..................................................................................................74 5-43 Standard deviation of absorb ance of data set 2005 obtained by spectrophotometer..................................................................................................75 5-44 N prediction for data set 2005 usi ng PLS procedure. Five samples had N prediction error larger than 2 g kg-1.......................................................................75 5-45 Correlation coefficient spectrum between the 1st derivative data set and the N concentrations........................................................................................................76 5-46 Standard deviation of the 1st derivative data set. .................................................76 5-47 N prediction for data set 2005N using PLS procedure. Fifteen percent of the samples had prediction N error larger than 2 g kg-1...............................................77 5-48 Picture of a citrus le af with uneven cuticle............................................................78 5-49 Standard deviation of repeated meas urement of a single leaf at different positions using the VIS detector............................................................................79 5-50 Standard deviation of repeated meas urement of a single leaf at different positions using the NIR detector............................................................................79 5-51 Voltage separation for th e VIS detector is 0.2 V...................................................79 5-52 Voltage separation for th e NIR detector is 0.14 V.................................................80

PAGE 13

xiii Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy SPECTRAL-BASED NITROGEN SENSING FOR CITRUS By Min Min May 2006 Chair: Won Suk “Daniel” Lee Major Department: Agricultur al and Biological Engineering Citrus is the most important agricultural crop in Florida. Nitrogen is the essential nutrient for growing crops and is also a con cern in maintaining a healthy environment. Heavy reliance on agricultural chemicals and low efficiency of fertilizers have raised many environmental and economic concerns. In this study, a nitrogen sensing system for citrus was developed. First the significant wavelengths were determined by correlation coefficient spectrum, standard deviation sp ectrum, stepwise multiple linear regression (SMLR) and partial least squa res (PLS) regression methods . The results showed that there has been a very good relationship between leaf spectra and their nitrogen concentrations. The standard error of prediction (SEP) and root mean square difference (RMSD) values were 1.2 g kg-1 and 1.2 g kg-1 for validation data sets for years 2002 and 2003. The most important wavelength is around 710 nm which was identified by all statistical methods. Some other wave lengths (448, 669, 1377, 1773, and 2231 nm) were identified by both SMLR and PLS as signi ficant wavelengths for nitrogen detection. Water did not show significant effect on nitrogen prediction.

PAGE 14

xiv Based on selected wavelengths, the sens ing system was designed to measure reflectance in wavelength ranges of 680-950 nm and 1400-2450 nm. A reflectance housing was designed to block environmenta l noise and to ensu re single leaf measurement. A halogen light source, two detector arrays, linear variable filters and a data acquisition board with 16-bit analog to digital convert (ADC) were used. The designed system had no moving parts, so it w ould be robust and resistant to vibration. The test results showed that the sensing system had good signal to noise ratio and linearity. The developed calibration m odel had 75% unknown samples which had predicted N concentrations with less than 2 g kg-1 error.

PAGE 15

1 CHAPTER 1 INTRODUCTION Visible and near-infrared spectroscopy (VNIS) is a nondestructive method which can rapidly analyze the chemical composition of materials with little sample preparation. VNIS has been largely used to detect plant st atus such as moisture content and nutrient stress by utilizing visible and near infrared spectral response from plant samples. According to the Beer-Lambert Law (Williams and Norris, 2001), the concentration of an absorber is directly proportional to the sample absorbance, and constituent of interest will absorb more light at particular wave lengths with higher concentrations. Among agricultural chemicals, nitrogen (N) is the most important and essential nutrient for growing crops and is also the most important nutrient element for maintaining a healthy environment. Nitrogen is an integral part of chlorophyll, which is th e primary absorber of light energy needed for photosynthesis. If N is used properly in conjunction with other necessary soil nutrients, it can spee d crop growth and increase yields. Figure 1-1. The basic structure of a chlorophyll. (http://www.chm.bris.ac.uk/motm/ch lorophyll/chlorophyll_h.htm)

PAGE 16

2 Variable fertilizer application is largely ba sed on either soil sampling or plant tissue analyses that are expensive, sl ow, and labor intensive. Rapid changes of nutrients in leaf or soil also require a real-time nutrient sens or. VNIS is a promising technology to realize nitrogen sensing in a fast and convenient way. Once nutrient variation maps are created ba sed on the nutrient sensor measurement, the results will be used to decide fertilizati on application rates at different locations of groves/trees. A dry fertilizer spreader truck could be utiliz ed to apply fertilizers more precisely. Precise placement of fertilizers will save much expense, time and labor, and prevent ground water contaminations from excessive fertilization. Citrus in Florida In Florida, citrus is the most important agricultural crop. According to the United States Department of Agricultu re (USDA) National Agricultural Statistics Service, in 2002 the total area of citrus in Florid a was 322,657 acres, and citrus production accounted for 74 percent of the total U.S. produc tion; California totaled 23 percent, while Texas and Arizona produced the remaining 3 percent. The economic value of FloridaÂ’s total production of 251 million boxes was $815 m illion (Florida Agricultural Statistics Service [FASS], 2005). Among citrus, the orange is a favorite fruit. It has consistently ranked as the third most consumed fresh fruit behind bananas a nd apples, and it ranks number one as a juice (Pollack et al., 2003). Among FloridaÂ’s tota l citrus production area, 81.4 percent was committed for orange production, followed by grapefruit production (13.2 percent); lastly, specialty fruit (e.g., ta ngerines, tangelos, lemons, limes, etc.) made up the final 5.4 percent (Annamalai, 2004).

PAGE 17

3 Fruit quality factors which ar e important to the Florida ci trus industry include juice content, juice solids and acid c oncentration, as well as matur ity date, fruit size, shape, color, and rind thickness (Tucker et al., 1995). They indicated that: Fruit quality is affected by variety, rootstock, climate, soil, pests, irrigation, and nutrition. The effects of irriga tion and nutrition on fruit quali ty could be substantial. The most important management practices in fluencing fruit quality are N, P, and K nutrition, and irrigation. Matu re trees on an adequate nutrition and irrigation management program produce fruit of accep table quality and show little response to minor changes in management programs. Excessive or deficient levels of either water or most nutrients, however, resu lt in poor fruit quality. For nitrogen, the effects of higher application level would Increase juice volume, total soluble solids (TSS), acid content, and juice color. Increase soluble solids per box and per acre. However, excessive N, particularly with inadequate moisture , can result in lowe r fruit production and result in lower TSS per acre. Decreases fruit size, weight, and peel thickness. In excessively fertilized young trees peel thickness can increase. Increases green fruit at harvest. The amount of green in the peel at harvest is primarily controlled by climate, but high N level may delay color break and increase re-greening of Valencia. Increases incidence of creasing and s cab and decreases incidence of peel blemishes such as wind scar, mite russeting, a nd rind plugging. Reduces stem end rot incidence and green mold or fruit in storage. Nitrogen Consumption for Citrus and Environmental Issues Nitrogen is an essential nut rient for citrus tree growth and fruit production. The average N usage of orange in 1999 was 240 kg ha-1 per crop year, 200 kg ha-1 per crop year in 1995 (FASS, 2005). In a mature grove, nitrogen used for leaf growth or taken up by a cover crop is largely recycled as leaves drop or the cover crop dies, the vegetative material decomposes, and mineralization re leases the N for reuse by the tree. This recycled N supplies most of the continuing need for new leaves, and relatively little

PAGE 18

4 fertilizer N is needed for growth. Repl acement of the N removed by fruit harvest becomes the main N requirement in a mature grove. Tucker et al. (1995) described that: A crop of 600 boxes of oranges per acre re moves approximately 75 lb of N from the grove. If a mature grove is re ceiving 200 lb of N per acre annually, approximately 125 lb of N per acre remains to be accounted for after crop removal. The fate of this 125 lb N is incompletely understood. Some may be lost by volatilization or deni trification, although denitrifi cation in vulnerable soils is generally considered to be minimal. In controlled leaching studies, approximately 40 percent of the N applied to the soil is not recovered even when water is supplied soon after fertilizer application. Al though unknown mechanisms may partially reduce N levels, substantial amounts of N applied in fertilizer are subject to leaching as indicated by elevated N levels in groundwater in some groves. Clearly, excess N application should be avoided on vulnerable soils where the potential for leaching exists. The presence of fertilizer and agricultural chemicals in groundwater has become a serious problem in Florida due to extremel y sandy soils, frequently shallow groundwater tables, frequent and intense rainfall, and heavy reliance on groundwater as a domestic and municipal drinking water source (Alva, 1997) . Many soils commonly used for citrus production in the Ridge area in Florida are partic ularly subject to leaching and referred to as vulnerable soils. These soils are well dr ained with low organic matter content and provide ideal conditions for leaching of many soil applied chemicals and nutrients including nitrogen fertilizers. Loss of N by den itrification is minimal for these soils due to their well aerated condition. In Florida, over 50 percent of the tota l fresh water comes from groundwater, and over 90 percent of the public rely on groundwater supplies fo r their drinking water. The quality of this water is an important cons ideration, since it may co me in contact with chemicals or heavy metals prior to returning to the surficial aquifer or flowing off site. Nitrate-N levels in groundwater above th e maximum contamination limit (MCL) have been observed in a number of shallow drinki ng water wells adjacent to citrus production

PAGE 19

5 areas of Central Florida. Th e National Drinking Water Stan dard for nitrate-N in the United States is 10 ppm. A nitrogen best ma nagement practice (BMP) bill was passed by the Florida Agriculture and Consumer Servi ces (DACS) “to devel op fertilizer BMPs designed to meet groundwater standards. Th ese BMPs are not mandatory, but if the grower implements the BMPs, the landow ner or lessee will not be subject to administrative penalties if nitrate gr oundwater standards are violated.” Reduction in groundwater nitrates will requi re higher N use efficiency by citrus which may require a combination of lower rates, more precise placements, split applications, controlled-release formulations, better irrigation management, supplemental foliar N sprays, and other improved practices. Fertilizer rate will probably have to be reduced on site where ground water nitrate al ready exceeds the MCL. Rates must be matched closely to requirements. Young trees under good management and groves producing fewer than 500 boxes per acre require N rates below 200 lb. Although sampling and analysis to obtain leaf N c ontent are time consuming and costly, leaf analyses should be included as a management tool at problem loca tions. Fertilizer rates should be gradually reduced in groves with existing nitrate levels above the MCL if leaf N is above the recommended range. Precision Agriculture for Citrus Precision agriculture is a management philosophy that responds to spatial variability found on agricultural landscapes . It means managing each crop production input, fertilizer, limestone, herbicide, insec ticide, and seed on a si te-specific basis to reduce waste, increase profits, and maintain the quality of the environment (Morgan and Ess, 2003). Current whole-field management a pproaches ignore variability and seek to apply crop production inputs in a uniform manne r. With such an approach, there was

PAGE 20

6 obviously the likelihood of over a pplication and under applica tion of inputs in a single field. Economical and environmental issues ar e the most important factors affecting the transition from whole-field to management s ite-specific crop manage ment due to in-field variability. Steps involv ing precision agriculture include determining yield variability in a field, determining its cause, deciding on possible solutions based on economic justification, implementing new techniques and repeating the procedure in a cyclic approach. Precision agriculture techniques could be used to improve economic and environmental sustainability in crop production. Global positioning system (GPS), geographic information system (GIS), remote sensing (RS), variable rate technology (VRT), yield mapping, and advances in sensor and information technology have enabled the farmer to visualize the entire field in a way that could help manage the agricultural opera tions efficiently and improve overall productivity. With precision agriculture tec hnologies, the farmer could effectively manage the crop throughout its life cycle, st arting from preparing soil, sowing seeds, applying fertilizers/pesticides and finally estimating yield during harvesting based on each individual plant, thus reducing the waste of resources due to in-field variability. Among precision agriculture technologies, id entifying in-field variability is important for implementing site-specific crop management on a specific field. For fertilizer management, leaf analyses pr ovide necessary information for fertilizer decisions. Combining nutrient information wi th soil type, tree position, age and size, variety and rootstock, a prescr iption map could be develope d for fertilizer practice. Commercial variable rate technology (VRT) fertilizer spreaders for citrus are currently being implemented in Florida groves. Usi ng a variable nitrogen application rate

PAGE 21

7 according to a prescription map would im prove profitability and reduce nitrate contamination of groundwater. Nutrient management largely relies on so il and leaf sampling. For citrus, nitrogen contents are stable in 4-to-6 month spri ng flush which is used for leaf sampling. Traditional leaf sampling is an expensive, labor intensive and time consuming work which generally takes several days for dryi ng and chemical analysis procedures. For highly efficient crop management, sensor-based variable rate fertilizer application would be more feasible to save cost, time and labor . Rapid changes of nutrients in leaf or soil also require a real-time sensor.

PAGE 22

8 CHAPTER 2 OBJECTIVES Three main objectives and sub-objectives of this research are outlined as follows: 1. To determine important wavelengths in the electromagnetic spectrum to assess N status of citrus leaves using reflectan ce spectroscopy. The sub-objectives are to Identify significant wavelengths in laboratory environments Develop a calibration model for nitrogen prediction 2. To study water effect on nitrogen prediction 3. To develop an in-field spectral-based nitrogen sensor that could detect N concentration in citrus leav es. The sub-objectives are to Determine design criteria for an N sensing system Fabricate the sensing system Design an algorithm to predict N content of unknown samples Test the N sensing system in a laboratory environment.

PAGE 23

9 CHAPTER 3 DETERMINATION OF SIGNIFICANT W AVELENGTHS AND PREDICTION OF NITROGEN CONTENT FOR CITRUS This chapter discusses the statistical methods for wavelength selection. General information of nitrogen relative wavelengths which were identified by previous research is presented in the text. Then the chap ter describes four methods for wavelength selection. Results based on two years of data sets are presented in this chapter. Introduction Two main N sources, proteins and chlorophy ll, generally exist in green leaves. Proteins, which are the primary nitroge nous compound in leaves, typically hold 70–80% of all N. Spectral bands for N related to protein absorptions at 2054 nm and 2172 nm were due to N in the molecular structure, in particular to C–N and N–H bonds (Kokaly , 2001). Chlorophyll holds an additional 5–10% of N. Chlorophylls exhibit strong absorption in the visible region arising fr om conjugated carbon–carbon single and double bonds of the porphyrin ring and the magnesium (Mg) ion. The infrared spectra of chlorophylls show strong abso rption due to C–H bonds in the phytol tail of the molecule (Katz et al., 1966). Chlorophyll a absorbs light at wavelengths of 430 nm and 660 nm and chlorophyll b absorbs at 450 nm and 650 nm (Farabee, 2001). The successful prediction of nutrient conten t is largely based on identification of wavelengths that significantly rela te to various nutrients. This is one of the main issues of spectral-based sensing technology, because th e success of a calibration model heavily depends on selected wavelengths. Various math ematical and statistical analysis methods

PAGE 24

10 have been used in setting up linear and nonlinear calibrati on models for N prediction. Thomas and Oerther (1972) found a non-linear relationship between reflectance at 550 nm and leaf N content of sweet pepper leaves with a correlation coefficient of -0.93. The near-infrared/red reflectan ce ratio (760 to 900 nm/630 to 690 nm) was reported to differentiate N treatment better than singl e-band reflectance measures for a corn canopy (Walburg et al., 1982). Card et al. (1988) found that N in dried and ground tree leaves could be determined accurately from reflectan ce with a laboratory spectrometer. Stepwise multiple linear regression (SMLR) was used to select 580 nm and 480 nm for total N prediction, and R2 was 0.90. Blackmer et al. (1994) f ound that reflectance near 550 nm measured on detached maize leaves obtained from a field used for an N fertilization experiment were able to separate N treatme nts. In addition to 550 nm, Blackmer et al. (1996) also selected 450, 630, 690, 710, and 760 nm for N estimation of corn canopies. Yoder and Pettigrew-Crosby (1995) found that th e short-wave infrared bands were best predictors for N, while visible bands were best for chlorophyll in fresh bigleaf maple leaves. Sui et al. (1998) devel oped a spectral reflectance sens or to detect N status of cotton plants using the four spectral bands of blue, green, red, and near-infrared. They reported that preliminary test results for di agnosing N status in cotton were promising. Lee et al. (1999) found that SPAD (Soil a nd Plant Analyzer De velopment, Minolta, Inc.) readings, based on transmittance at 659 and 940 nm, were well correlated with N content in corn ear leaves (R2 = 0.962). They developed prediction models by partial least squares (PLS) regression, principal compone nt regression (PCR), and multiple linear regression (MLR). The results showed that models built by PLS and PCR were better than models from MLR, and that the standard errors of prediction (SEP) for ear leaf N

PAGE 25

11 were 1.6 g kg-1, 1.5 g kg-1, and 2.0 g kg-1 for PLS, PCR, and MLR, respectively. Jensen (2000) reported that chlorophyll a absorbed light at wavele ngths of 430 and 660 nm, and chlorophyll b absorbed light at 450 and 650 nm. Studies by Carter and Knapp (2001) showed good linear relationships of reflectance, transmittance, and absorbance near 700 nm to total leaf chlorophyll concentration in senescing leaves of five different species (sweetgum, red maple, wild grape, switchcan e, and longleaf pine). They reported that these wavelengths near 700 nm were crucia l for estimating leaf chlorophyll stress. Tumbo et al. (2002) used a back-propagation neural network model for corn N prediction in field conditions. The model used 201 spectral bands as inpu t, covering the range from 407 to 940 nm, and results proved that the ne ural network model could considerably reduce interfering effects of cloud cover and solar angle. The model showed good correlation between predicted a nd actual chlorophyll meter re adings of the training set (R2 = 0.91). A good relationship was also found in the validation data set (R2 = 0.74). Objectives The objectives of this chapter were to determine important wavelengths for predicting the N status of ci trus trees using reflectance spectroscopy in a laboratory setting, to develop N prediction models usi ng stepwise multiple lin ear regression (SMLR) and partial least squares (PLS ) analysis, to compare the performance of SMLR and PLS regression in wavelength selection a nd calibration model development. Materials and Methods Leaf Sampling and Reflectance Measurement During the July of 2002, 1000 spring flush or ange leaves were collected from an experimental orange grove located near Lake Alfred, Florida. The Valencia trees were planted in March 1986 at a 7.6 3.8 m spacing with 346 trees ha-1. The experimental

PAGE 26

12 plots consisted of single rows of four adjacent trees randomly located in the grove. There were seven replications of five treatment rates of N fertilizer: 0 kg ha-1 (N1), 112 kg ha-1 (N2), 168 kg ha-1 (N3), 224 kg ha-1 (N4), and 280 kg ha-1 (N5). The N fertilizer was split into four parts and equally applied in February, April, May and October. In July of 2003, another 1000 spring flush orange leaves were collected from Gapway Groves, Ft. Meade, Florida. The e xperiment was a randomized complete block design. Each plot had 36 trees (6 6), with the middle 16 trees (4 4) measured. The citrus leaf samples were obtained from five citrus blocks with N application rates of 156.8 (G1), 201.6 (G2), 246.4 (G3), 291.2 (G4), and 336 kg ha-1 (G5). The variety was Hamlin. Data sets collected from years 2002 and 2003 were defined as data set 2002 and data set 2003. Two hundred leaves were randomly picked from trees of each N treatment. Collected leaves were kept in plastic bags and stored in refrigerator. Due to sample requirements for laboratory N analysis, whic h needed 0.25 g of dried and ground leaves, five leaves from the same N treatment were combined to make one composite sample. Thus, there were 40 samples in each treatm ent and 200 samples in all for each year. The leaves were cleaned to remove dust firs t. Then diffuse reflectance of each leaf was measured by a spectrophotometer (Cary 500, Varian, Inc.) with an integrating sphere (DRA-CA-5500, Labsphere, Inc.) from 400 to 2500 nm with a 1 nm increment. The diameter of the sample measurement port was 38 mm, and the coating material inside the integrating sphere was white polytetrafl uoroethylene (PTFE). UV and mercury lamps were used as light sources. The FWHM (full width at half maximum) of the spectrophotometer was 2 nm. A 50 mm diameter PTFE disk was used to obtain the optical reference standard for the system each day before spectral measurement of the

PAGE 27

13 leaf samples. There were a total of 1000 spectra (i.e., one per leaf), and the five spectra of one sample were averaged to relate N concentration to reflectance. After reflectance measurement, leaves were dried in an oven at 103°C for 24 h and were then ground (ASAE Standards, 2002). Chem ical analysis of act ual N concentration was conducted by AOAC Of ficial Method 990.03 ( AOAC, 1995), which had a repeatability of 99%. Normalization of the Reflectance Spectra Because 31 of the 1000 citrus leaves from 2002 were too small to cover the entire sample port in the spectrophotometer, a white 3.175 mm thick Teflon sheet with a 25 mm diameter hole in the middle was used duri ng the reflectance measurement process to avoid unwanted background exposure to the spectrophotometer and to hold these 31 leaves in the proper position. However, use of the Teflon sheet increased the reflectance due to the white area exposed to the spect rophotometer. Figure 3-1 shows reflectance spectra of the same leaf with and without th e Teflon sheet. The Teflon sheet affected the spectrum; therefore, data obtai ned with the Teflon sheet could not be used directly for analysis. 0 10 20 30 40 50 60 70 80 90 100 400700100013001600190022002500Wavelength (nm)Reflectance (%) leaf scanned without Teflon sheet leaf scanned with Teflon sheet Figure 3-1. Spectral response of a citrus leaf, with and without a Teflon sheet.

PAGE 28

14 In order to remove the effect of the Teflon sheet, an experiment was conducted. Another calibration set of twen ty citrus leaves was obtained from an experimental citrus grove at the University of Florida and was used to find relationships between spectra obtained with and without the Teflon sheet. A linear relationship was obtained for each wavelength between reflectance with the Teflon sheet ( R ') and reflectance without the Teflon sheet ( R ). Spectra measured with Teflon sheet were normalized with equation 3-1: R = a + bR ' (3-1) The coefficients ( a and b ) were obtained by linear regression analysis. From the 20 calibration spectra, the mean absolute error between the original spectra without the Teflon sheet and the normalized spect ra was calculated by equation 3-2: 2071 20 | ' | MAE20 1 2485 415 ij ij ijR R (3-2) where MAE = mean absolute error (%) R 'ij = predicted reflectance in the i th spectrum at wavelength j (%) Rij = actual reflectance in the i th spectrum at wavelength j (%). Figure 3-2 shows the effects of normaliz ing an example spectrum. Over all wavelengths of the 20 calibration spectra, the mean absolute error was 0.6%. A total of 31 leaves were measured with the Teflon sheet. They were smoothed using a binomial method with 15 wavelengths for reducing spectral measurement error and were normalized with equation 3-1. As a result, 14 wavelengths at each end were truncated after smoothing, and the wavelength range was changed to 415 to 2485 nm. The original and smoothed spectra were very similar; thus it appeared that th ere was no significant

PAGE 29

15 information loss in the smoothing process. The other 969 spectra were also smoothed with the same 15-wavelength process. 0 20 40 60 80 100 400 700 1000 1300 1600 1900 2200 2500 Wavelength (nm)Reflectance (%) leaf scanned without Teflon sheet normalized spectrum Figure 3-2. Normalized spectrum of the same leaf shown in figure 3-1 compared to the original spectrum. The spectra of five leaves were aver aged due to minimum sample weight requirements in chemical analysis. Then, according to the Beer-Lambert law, which states that the concentration of an absorb er is directly propor tional to the sample absorbance (Williams and Norris, 2001), spectral reflectance was converted to absorbance using equation 3-3 to obtain a better relationship be tween spectra and N concentration: R A 1 log (3-3) where A is absorbance, and R is reflectance. Figure 3-3 shows the absorbance of two different citrus leaf samples. It illustrates that the leaves strongly abso rb blue (450 nm) and red (680 nm ) light in the visible range (Farabee, 2001) and reflect green light ( 550 nm) to a different extent depending on N concentration. Sample 5N31, with 32.3 g kg-1 N, absorbed more green light than 1N04,

PAGE 30

16 which had N concentration of 22.3 g kg-1. Two water absorption bands were found at 1450 and 1940 nm. 0 0.2 0.4 0.6 0.8 1 1.2 40055070085010001150130014501600175019002050220023502500 Wavelength (nm)Absorbance (Abs) 5N31, N=32.3 g/kg 1N04, N=22.3 g/kg 550 660 450 1450 1940 710 Wate r Figure 3-3. Absorbance spectra of two differe nt citrus leaf samples: 1N04 (red line) received 0 kg N ha-1 (N1) and its leaf N concentration was 22.3 g kg-1, and 5N31 (blue line) received 280 kg N ha-1 (N5) and its N concentration was 32.3 g kg-1. Determination of Important Wavelengths The data set for each year was separated in to training and validation data sets. The training data set included 100 samples (20 samples from each N treatment), which was half of the total of 200 samples. The remaining 100 samples were used as the validation data set. Four methods, correlation coeffici ent (r) spectrum, sta ndard deviation (std) spectrum, SMLR, and PLS regression, were used for wavelength selection. Correlation coefficient spectrum The simplest method was to compute corre lation coefficients between absorbance at each wavelength and the actual N concen tration of the samples. The correlation coefficient spectrum provided a picture of the relationship between absorbance and N concentration. Wavelength regions showing hi gh correlation are regions that should be

PAGE 31

17 selected, and regions showing low or no co rrelation should be ignored. The SAS CORR procedure (SAS, 1990) was used to cal culate correlation coefficients ( r ). Standard deviation spectrum Different from correlation coefficient sp ectrum, standard deviation spectrum only examines information in the spectra themselves . This is an unsupervised analysis method which demonstrates variability of absorb ance without considering N concentrations. Wavelengths showing larger standard deviation may indicate more important N absorption bands. Some other components in the samples, for example water content, may cause larger standard deviations also. Stepwise multiple linear regression (SMLR) Stepwise multiple linear regression (SMLR) is an improved version of forward regression that permits re-examination at every step of the variables incorporated in the model in previous steps. Each forward selection step, with a significance level ( ) of 0.5, can be followed by one or more backward e limination steps with a significance level ( ) of 0.1. The stepwise selection process terminat es if no further variable can be added to the model or if the variable just entered into the model is the only variable removed in the subsequent backward elimination. SMLR was reported to have a good ability for wavelength selection by Card et al. (1988) . In SMLR, overfitting could be a problem because too many wavelengths might be select ed by the stepwise procedure. In this study, the number of wavelengths for the ca libration model was selected based on the model that yielded the best R2 for the validation data set. The stepwise option in the SAS REG procedure was used to select wavelengths by SMLR.

PAGE 32

18 Partial least squares (PLS) regression Partial least squares (PLS) regression has been used as a powerful tool in chemometrics and other fields (Wold et al ., 1983). Frank and Frei dman (1993) described PLS as a method of creating new explanat ory variables that maximize the squared covariance between themselves and the res ponse variable. Techniques implemented in the PLS procedures work by extracting successive linear combinations of the predictors, optimally explaining response variation and pr edictor variation. PLS has been described as a two-step method where the first step re duces matrix dimensions and the second step identifies latent structure models in the da ta matrix (Lingaerde and Christophersen, 2000; Helland, 2001). In contrast to PCR, which chooses factors that explain the maximum variance in predictor variables without c onsidering the response variables, the PLS method balances the two objectives, seeking th e factors that explain both response and predictor variations (SAS, 1990). The predicted residual sum of squares (PRESS) statistic in PLS measures how well the regression equation fits the data set. An optimal number of factors is generally obtained when PRESS is minimized (S undberg, 1999), and a smaller PRESS value indicates a better model prediction. However, selecting the number of factors where the absolute minimum PRESS exists may not be the best choice. By using the crossvalidation option (CVTEST) in SAS PLS, a st atistical comparison can be performed to test the significance of differences in the PR ESS value at each number of factors, thus determining how many factors should be selected for a calibration model. In PLS, X-loadings represent the common variations in the spectral data, and Xweights represent the changes in the sp ectra that correspond to the regression constituents. High X-loadings and X-weight s are usually used to identify important

PAGE 33

19 wavelengths (Esbensen , 2002). However, since X-loadings and X-weights do not directly reflect the relationship betw een predictors and responses , the B-matrix from the traditional regression equation, Y = XB , is used to give the accumulated picture of the most important wavelengths. Wavelengths with a high B value contribute more to a calibration model, and could be considered to be important wavelengths. A B-matrix can be calculated from the PLS loadings and weights: ' ) ' ( B1q w p w (3-4) where w is the X-weight, p is the X-loading, and q is the Y-weight. The coefficient of determination (R2) between predicted N concentration and true concentration, standard error of calibration (SEC), standard e rror of prediction (SEP), and root mean square difference (RMSD) were used to evaluate reliability of the calibration model (ASTM, 1997). SEC, SEP, and RMSD were determined by the following equations: n i ie p n kg g1 2 11 1 ) ( SEC (3-5) n i ie e n kg g1 2 1) ( 1 1 ) ( SEP (3-6) n i ie n kg g1 2 11 ) ( RMSD (3-7) where n = number of samples p = number of independent va riables in calibration model

PAGE 34

20 ei= difference between actual N concentrati on and predicted N c oncentration in the i th sample e= mean of ei. Results and Discussion Chemical Analyses of Sample Leaves The results of N concentration analysis of the Valencia and Hamlin leaves are shown in tables 3-1 and 3-2. The trend in N c oncentration was similar to the trend in the actual N application rates applie d to the groves where the samples were collected for data set 2002. Mean testing using Tukey's grouping sh ows that the average N concentrations of the samples from the five application rate s were significantly different from each other with the exception of N2 and N3 (table 3-1) . For data set 2003, a m eans test using Tukey grouping analysis shows that different N application rates did not separate N concentrations in citrus leav es of each grove very well. M eans of N concentration of the groves G2, G3, G4 and G5 were grouped to re present the same N level. Only grove G1 was significantly different from other groves. Due to higher fertili zer application rates (160 to 330 kg ha-1) that were used, linear responses of leaf nutrient were not as obvious as that when using lower fertiliz er application rates (0 to 160 kg ha-1). Figure 3-4 shows the N concentration distribution for data sets 2002 and 2003. Since the experimental grove in year 2003 received a higher N fertili zer application rate, the N response in the leaves was higher than data set 2002. Table 32 lists the N analysis result for data set 2003. N concentration ranges were 19.9 g kg-1 to 33.8 g kg-1 for 2002 and 24.9 g kg-1 to 35.3 g kg-1 for 2003. According to the chemical usage informati on for oranges in Florida in 1999 (FASS, 2005), N fertilizer was generally applied four or five times per year, with a total average

PAGE 35

21 application rate of 242 kg ha-1. The leaf N concentration rang e of the sample leaves was similar to the generally expected N conten t range of an orange grove, about 25 g kg-1 to 30 g kg-1 (Hanlon et al., 1995). Table 3-1. Results of N concentration analys is of the samples from five different N application rates for data set 2002. N treatment Value N1 N2 N3 N4 N5 Actual N application (kg ha-1) 0 112 168 224 280 Number of samples 40 40 40 40 40 Average N (g kg-1)* 23.2 A 28.1 B 28.3 B 29.7 C 31.0 D Standard deviation (g kg-1) 1.4 1.0 1.3 1.4 1.4 N Range (g kg-1) 19.9-25.7 25.9-30.224.8-31.5 25.6-33.7 27.8-33.8 Average Water content (g g-1) 0.622 0.627 0.627 0.612 0.624 * Means within a row followed by the same letter are not signif icantly different ( P> 0.05). Table 3-2. Results of N concentration analys is of the samples from five different N application rates for data set 2003. N treatment Value G1 G2 G3 G4 G5 Actual N application (kg ha-1) 156.8 201.6 246.4 291.2 336.0 Number of samples 40 40 40 40 40 Average N (g kg-1)* 29.6 A 30.8 B 30.9 B 31.4 B 31.6 B N Range (g kg-1) 26.232.3 26.935.3 25.733.1 24.9-35.2 27.933.9 Standard deviation (g kg-1) 1.3 1.8 1.4 2.1 1.4 Average Water content (g g-1) 0.606 0.615 0.607 0.604 0.615 * Means within a row followed by the same letter are not signif icantly different ( P> 0.05). Wavelength Selection and Calibration Models Development Correlation coefficient spectra Figure 3-5 shows the correlation coefficient spectrum between absorbance and actual N concentration of the calibration samp les. Absorbances in some regions were highly correlated with the N concentration. The wavelength range near 550 nm, a wellknown N absorption band (Thomas and Oerthe r, 1972; Blackmer et al., 1994), showed a peak correlation coefficient of 0.51. The highe st correlation coeffici ent in the visible range (0.56) was seen at 707 nm. This is consis tent with results by Blackmer et al. (1996)

PAGE 36

22 and Carter and Knapp (2001). They reporte d that wavelengths near 700 nm were significant for detecting chlo rophyll, which is also cl osely correlated with N concentration in green leaves . In the near-infrared range, large regions from 1364 to 1897 nm and from 1995 to 2485 nm showed a nega tive correlation coefficient with | r | larger than 0.5. For data set 2003, since the N range (24.9-35.3 g kg-1) was not as wide as N range in dataset 2002 (19.9-33.8 g kg-1), we got very different result for data 2003. Wavelength range from 450 nm to 1350 nm had higher correla tion coefficient with r larger than 0.3, especially for wavelength 730 nm (figure 36), which had highest r = 0.44. However, for wavelength range of 1450 2500 nm, the correlat ion coefficients were very low, which were just above or below zero. 0 10 20 30 40 50 60 70 80 90 100 19.1-20.020.1-22.022.1-24.024.1-26.026.1-28.028.1-30.030.1-32.032.1-34.034.1-36.0 N concentration (g/kg)Number of samples 2003 2002 Figure 3-4. Distribution of N concentration for year 2002 and 2003. N ranges were 19.933.8 g kg-1 for data set 2002, and 24.9-35.3 g kg-1 for data set 2003. Standard deviation spectra Figures 3-7 and 3-8 show the standard de viation spectra for data sets 2002 and 2003. Peaks near 580, 702, 1460 and 1904 nm matche d with peaks in r spectra. In the visible region, the standard deviations at 580 nm and 703 nm of data set 2002 is 0.037, which were higher than the std values for data set 2003 (std = 0.19). This is because N

PAGE 37

23 range in data set 2002 was wider than that of data set 2003. In the near-infrared region, standard deviations were si milar for both data sets. -1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 400700100013001600190022002500 Wavelength (nm)Correlation coefficient (r) 707 553 1364 1897 1995 Figure 3-5. Correlation coefficients between ab sorbance at each wavelength and leaf N concentration of the calibra tion data set for year 2002. -0.1 0 0.1 0.2 0.3 0.4 0.5 40055070085010001150130014501600175019002050220023502500Wavelength (nm)Correlation coefficient (r) 730 494 630 Figure 3-6. Correlation coefficients between ab sorbance at each wavelength and leaf N concentration of the calibra tion data set for year 2003.

PAGE 38

24 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 400700100013001600190022002500 Wavelength (nm)Standard deviation (g/kg) 580 702 1460 1904 Figure 3-7. Standard deviation of absorb ance for calibration data set of year 2002. 0 0.05 0.1 0.15 0.2 0.25 400700100013001600190022002500Wavelength (nm)Standard deviation (g/kg) 560 704 1445 1890 2020 2389 Figure 3-8. Standard deviation of absorb ance for calibration data set of year 2003. Stepwise multiple linear regression (SMLR) Table 3-3 lists the analysis results ba sed on the SMLR procedures. The results showed a good relationship betw een the actual N concentrat ion and the predicted value for both training and validation data sets. Ten wavelengths we re selected for data set 2002, eight wavelengths for data set 2003. The R2 for the validation data set was 0.816 and RMSD was 1.32 g kg-1 for data set 2002, 0.550 and 1.31 g kg-1 for data set 2003. The R2 value for the data set 2002 was higher than 20 03 due to wide spread N range for data

PAGE 39

25 set 2002. However, high R2 does not necessarily mean high SEP and RMSD value. Same SEP or RMSD values coul d have very different R2 values. Comparing the SEP and RMSD value for both data sets, it was found th at two data sets had very similar SEP and RMSD values which indicate the equal accuracy for predicted N. Figures 3-9 and 3-10 illustrate the regression relationship between predicted N and actual N. The wavelengths of 448 and 719 nm selected from data set 2002, 537 and 561 nm selected from data set 2003 are related to chlorophyll, which has absorbance peaks at 430, 450, 650, and 660 nm. Wavelengths of 719 nm from data set 2002 correspond to a peak at 707 nm in the correlation coefficien t spectrum, indicating a higher importance of this wavelength than 550 nm for N detection. Wavelengths of 2101 and 2231 nm from data set 2002, 2202 nm from data set 2003 are related to protein. Partial least squares (PLS) regression In PLS, modeling is not complete until the appropriate number of factors is chosen. For data set 2002, the smallest error was found at the seventh factor with a PRESS value of 0.44 ( P >0.1). These seven factors accounted for 98.8% of the model variation and 88.9% of the variation of dependent variables (table 3-4). For data set 2003, the smallest error was found at the ninth fact or with a PRESS value of 0.74 ( P >0.1) which is larger than data set 2002. For data set 2002, the B coefficients calcu lated by the PLS procedure with seven factors is shown in figure 3-11. Wavelengths of 447, 676, 724, 775, 1160, 1376, 1490, and 2231 nm had |B| > 0.2 and therefore were c onsidered to contribute more information to the calibration model.

PAGE 40

26 Table 3-3. SMLR analysis resu lt for data sets 2002 and 2003. 20 25 30 35 20253035Actual N (g/kg)Predicted N (g/kg) Figure 3-9. N concentration prediction us ing SMLR for data set 2002. This method generated R2 = 0.816 and RMSD = 1.32 g kg-1 for the validation data set. 20 25 30 35 20253035 Actual N (g/kg)Predicted N (g/kg) Figure 3-10. N concentration prediction us ing SMLR for data set 2003. This method generated R2 = 0.550 and RMSD = 1.31 g kg-1 for the validation data set. RMSD ( g kg-1) R2 Year Selected wavelengths (nm) SEC ( g kg-1) SEP ( g kg-1) CalibrationValidation Calibration Validation 2002 448, 669, 719, 1377, 1773, 1793, 1834, 2000, 2101, 2231 0.94 1.26 0.89 1.32 0.916 0.816 2003 537, 561,1653, 1683, 1378, 1618, 2202, 1864 1.02 1.15 0.98 1.31 0.700 0.550 SEP = 1.26 g kg-1 RMSD = 1.32 g kg-1 R2 = 0.816

PAGE 41

27 Table 3-4. Percent variation explained by the se ven factors in the PLS regression for data set 2002. Model Effects Dependent Variables No. of Extracted Factors Current(%) Total(%) Current(%)Total(%) Root Mean PRESS 1 51.7 51.7 52.1 52.1 0.77 2 13.7 65.4 7.5 59.6 0.74 3 7.4 72.7 5.8 65.4 0.69 4 23.0 95.7 1.7 67.1 0.68 5 2.1 97.9 7.0 74.2 0.61 6 0.6 98.4 11.6 85.8 0.47 7 0.4 98.8 2.6 88.9 0.44 Table 3-5. Percent variation explained by the ni ne factors in the PLS regression for data set 2003. Model Effects Dependent Variables No. of Extracted Factors Current(%) Total(%) Current(%)Total(%) Root Mean PRESS 1 53.9 53.9 14.5 14.5 1.07 2 36.1 89.9 7.1 21.6 1.01 3 5.4 95.3 11.8 33.5 0.98 4 1.9 97.3 16.1 49.6 0.91 5 0.7 98.0 7.9 57.5 0.85 6 0.7 98.8 5.3 62.7 0.78 7 0.1 98.9 7.8 70.5 0.76 8 0.3 99.3 2.4 72.9 0.75 9 0.2 99.5 1.7 74.7 0.74 -0.4 -0.2 0 0.2 0.4 0.6 0.8 400700100013001600190022002500 Wavelength (nm)B coefficient 447 724 553 1376 1160 2231 775 1490 676 1773 Figure 3-11. B coefficient determined from the training data set using PLS regression with seven factors for data set 2002.

PAGE 42

28 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 400700100013001600190022002500Wavelength (nm)B coefficient 443 721 828 1135 1638 2033 2469 Figure 3-12. B coefficient determined from the training data set using PLS with nine factors for data set 2003. Figure 3-12 shows B coefficients for data set 2003. Due to narrow N range, there is no clear structure shown in the B coefficient spectrum. Only a peak at wavelength of 721 nm matched with the similar wavele ngth selected from other methods. Wavelengths at 448, 669, 719, 1377, 1773, and 2231 nm determined from data set 2002 matched with the peaks in the B-matrix ve ry well, an additiona l indication of their possible importance. In figure 3-11, two peak s at 447 and 676 nm in the B coefficient spectrum, matched the chlorophyll abso rption bands. Wavelengths of 1377, 1773, and 2231 nm have not been reported in previous research, which mostly concentrated on wavelengths in the 400 to 1100 nm range. Table 3-6 shows the results obtained with the PLS proce dure for data sets 2002 and 2003. Using the full spectrum, results for both da ta sets are very si milar to the results generated by SMLR procedure. The data se t 2002 had an RMSD for the validation data set of 1.22 g kg-1, R2 of 0.828 and SEP of 1.20 g kg-1. After removing three outliers, the data set 2003 had an RMSD for th e validation data set of 1.20 g kg-1, R2 of 0.597 and

PAGE 43

29 SEP of 1.19 g kg-1 which even had better N prediction accuracy. The predicted N results were shown in figures 3-13 and 3-14. Comparing the SMLR and PLS procedures , both worked very well in this study. Min and Lee (2005) indicated that SMLR wo rked especially well for low collinearity data set. They averaged absorbance value of every 20 wavelengths into one variable in data set 2002. The SMLR results base d on this averaged data set had R2 = 0.839 and RMSD = 1.22 g kg-1. The PLS procedure appears superi or to the SMLR procedure for full-spectrum analysis due to its ability to compress data. However, difficulties with the PLS method include a complex algorithm, whic h could be difficult to understand, and the large number of wavelengths that we re used for the calibration model. Table 3-6. Results for the data se ts 2002 and 2003 using PLS regression. 20 25 30 35 20253035Actual N (g/kg) Predicted N (g/kg) Figure 3-13. N concentration prediction using PLS for data set 2002. This method generated R2 = 0.828 and RMSD = 1.22 g kg-1 for the validation data set. RMSD ( g kg-1) R2 Year Factors SEC ( g kg-1) SEP ( g kg-1) CalibrationValidation Calibration Validation 2002 7 1.06 1.20 1.02 1.22 0.889 0.828 2003 9 0.94 1.19 0.90 1.20 0.747 0.597 SEP = 1.20 g kg-1 RMSD = 1.22 g kg-1 R2 = 0.828

PAGE 44

30 20 25 30 35 20253035 Actual N (g/kg)Predicted N (g/kg) Figure 3-14. N concentration prediction using PLS for data set 2003. This method generated R2 = 0.597 and RMSD = 1.20 g kg-1 for the validation data set. Summary This chapter was conducted as a prelimin ary step toward developing a real-time spectral N sensor for citrus trees. Diffuse re flectance of leaf samples was measured from 400 to 2500 nm using a spectrophotometer in a laboratory environment. A correlation coefficient spectrum, standard deviat ion spectrum, SMLR, and PLS regression procedures were used to determine im portant wavelengths. Both SMLR and PLS regression procedures yielded good results. Wa velengths of 560 and 710 nm selected by correlation coefficient spectra and sta ndard deviation spectra, 448, 669, 719, 537, and 561 nm selected by SMLR, and 724 nm from B coefficient spectra were related to chlorophyll absorption bands. Wavelengths arou nd 710 nm had been selected by most of the statistical methods, which had been c onsidered as the important wavelength for separating different N concentrations. Wave lengths selected by SMLR in the near infrared range, e.g., 2101, 2231, and 2202 nm might relate to protein, a nd could also have potential use in N detection. Due to the ability of PLS to reduce col linearity in datasets, calibration models using PLS produced better results than those using SMLR for full SEP = 1.19 g kg-1 RMSD = 1.20 g kg-1 R2 = 0.597

PAGE 45

31 spectrum analysis. The RMSD values produced by PLS procedure were 1.20 g kg-1 for the validation dataset of both years, wh ile the RMSD values produced by SMLR procedure were 1.30 g kg-1 for the validation dataset of both years. Strong relationship between leaves spectra and their N concentrations was revealed (SEP=1.19 g kg-1 and RMSD=1.20 g kg-1).

PAGE 46

32 CHAPTER 4 WATER EFFECT ON NITROGEN PREDICTION FOR CITRUS LEAF Among all the required constituents in a plant, water is the most important substance necessary for plant growth. By comp aring the results from data sets based on dry basis N and wet basis N, the effect of wate r on nitrogen prediction is discussed in this chapter. Introduction Water is a good absorber of middle-in frared energy which strongly affects reflectance spectra of the leaf. As the water co ntent of the leaves decreases, reflectance in the middle-infrared region increases substa ntially (Thomas et al ., 1971; Tucker, 1980; Carter, 1991; Jensen, 2000). A fresh ci trus leaf consists of 0.55-0.65 g g-1 water. Water is a good solvent and medium in which many metabolic reactions occur. Additionally, water itself is a reactant in a number of metabolic reactions. For example, in photosynthesis the hydrogen atoms from th e water molecule are incorporated into organic compounds and oxygen atoms are releas ed as oxygen gas. Studies of cellular water in plant cell showed that a large portion of the water can be considered “free” and mobile, which is quite similar in nature to a salt solution. On the other hand, a small fraction of the water in plant cells is held very tenaciously by plant cell constituents, through dipolar and hydrogen bond forces. Moreove r, water in the membrane interfacial region has such a high density that water mo lecules are so closely appressed to one another as to make this frac tion of cell water almost semi crystalline in structure (Noggle and Fritz, 1976).

PAGE 47

33 The effect of water content on leaf spect ral reflectance could be considered as primary and secondary effects (Carter, 1991) . The primary effect of water content on reflectance depends solely on the absorption pr operty of water, while secondary effect depends on other substances change which was caused by change of water content. Curcio and Petty (1951) indicated that wa ter absorbs strongly at wavelengths from approximately 1,300 nm to 2,500 nm. As the water content of leaves decreases, reflectance in the 1300-2500 nm region in creases substantially. Reflectance at wavelengths of 1450 nm and 1940 nm were sensitive to leaf water condition changes. The relationship between reflectance of citrus leaf at 1450 nm and re lative turgidity were statistically significant, r = -0.95 (Thomas, et al. 1971). Diffu se reflectance of the nearinfrared energy from plant leaves is due to the internal scattering at the cell wall-air interfaces within the l eaf (Gausmann et al., 1969). Scatter effect by water molecules is far less significant than diffuse reflectance. As the water content of leaves decreases, reflectance also increases in the range of 400-1300 nm (Woolley, 1971; Bowman, 1989). The secondary effects of water content that occur on reflectance can not be explained solely by the ra diative properties of water. Other substances in the leaf, such as pigmen ts, can influence absorption of water content. The secondary effects of water content on reflectance were largely wavelengthindependent and resulted from decreased ab sorption by pigments. In the visible range, water largely transmits light. Reflectance spec trum increases in the 400 to 700 nm region did not result from decreased absorption by water. Rather, they corresponded to absorption of chlorophyll (Carter, 1991).

PAGE 48

34 Objectives Since water largely affects le af spectral characteristics, it may have potential effect on N prediction. A recent study by Min et al. ( 2005) showed that water content had a good relationship with N content in Chinese cabbage leaf (r = 0.76). The objective of this research is to study the relati onship between water and N conten t in citrus leaf, and also to discover how water aff ects the relationship between spectral reflectance and N concentration of citrus leaf. Materials and Methods Leaves with same amount of N could ha ve different water contents. High water contents in leaves tend to decrease reflectance . Variation of water co ntent in leaves would affect accuracy for N prediction. Difference sp ectra between the dried and fresh sample can be obtained by subtracting dried sample sp ectra from fresh sample spectra. Spectra of fresh leaves and difference spectra were studie d for their relationship with water content. However, it is much more difficult to normalize water content in spectra than to predict water content using spectra because each spectrum contains 2500 wavelengths while there is only a water predictor. Since N concentration is reported base d on dried and ground citrus leaves, and reflectance spectra are measured from fresh leaves, an alternative way of normalizing water content could be developed by convert ing N concentration based on weight of dried leaves into N concentration of fres h leaves. Eq. 4-1 desc ribes the relationship between N concentrations of dried leaf (Nd, g kg-1) and those of fresh leaf (Nf, g kg-1). Then Nf includes information on both N c oncentration and water content. d f d fN W W N (4-1)

PAGE 49

35 where Nd = dry basis N concentration of a dried leaf sample (g kg-1), Nf = wet basis N concentration of a fresh leaf sample (g kg-1), Wd = Weight of a dried leaf sample (g), and Wf = Weight of a fresh leaf sample (g) Data set of 2002 and 2003, and statistical methods of correlation coefficient (r) spectrum, partial least squares (PLS) regres sion, and stepwise multiple linear regression (SMLR) were used for data analysis. In PL S, two constituents, N and water content, could be used as predictors, and this method was named as PLS II method. Compared with just using N contents as predictors, named as PLS I method, we can see how water would affect the analysis result. Results and Discussion Water Contents Distribution Water content distributions for both year s were shown in figures 4-1 and 4-2. Tables 4-1 and 4-2 list the simple statistic results of water contents of the datasets of 2002 and 2003. Average water content in the 2002 dataset was 0.62 g g-1, and the average water content in the 2003 dataset was 0.61 g g-1, which are very close to each other. Four outliers in data set 2002 and one outlier in data set 2003 were removed for water prediction later on in the data analysis due to their value were beyond 99.7% (z>3) confidence interval. Water Effect on Spectra Characteristics Figure 4-3 shows reflectance of a sample l eaf measured at fresh and dried status with a water content of 0.607 g g-1. Difference between the fr esh and dried spectra was calculated and also plotted in figure 4-3. Reflectance of the dried sample is much higher

PAGE 50

36 than the fresh sample in the middle-infr ared range (1300 to 2500 nm). The largest reflectance differences in response to decr eased water content occurred throughout the infrared range, with smaller differences occurring in the visible range. Negative reflectance differences were observed n ear 760 nm. In figure 4-3, the maximum reflectance difference occurred at waveleng ths of 1415 and 1889 nm which were in the wavelength ranges of 1410 to 1420 nm and 1880 to 1890 nm which were indicated by Carter (1991). This shows that primary effect s of water content on leaf reflectance were far more prominent than secondary effects. S econdary effects that re sult from changes in absorption by pigments with changes in wate r content may be nearly as prominent as some primary effects. 0.4 0.5 0.6 0.7 0.8 020406080100120140160180200Sample numberWater content (g/g) Figure 4-1. Distribution of wate r contents for data set 2002. F our outliers (circled) were removed for data analysis. 0.4 0.5 0.6 0.7 020406080100120140160180200 Sample numberWater content (g/g) Figure 4-2. Distribution of water contents for data set 2003. One outlier (circled) was removed for data analysis.

PAGE 51

37 Table 4-1. Results of N concentration and wate r content analysis of the samples from the blocks with five different N application rates in 2002. N treatment Value N1 N2 N3 N4 N5 Actual N application (kg ha-1) 0 112 168 224 280 N Range (g kg-1) 19.9-25.7 25.9-30.2 24.831.5 25.6-33.7 27.8-33.8 Wet N range (g kg-1) 7.8-10.2 8.8-11.2 8.911.7 10.3-13.3 10.6-12.6 Water content range (g g-1) 0.588-0.650 0.442-0.759 0.5830.667 0.555-0.636 0.593-0.650 Average Water content (g g-1) 0.622 0.627 0.627 0.612 0.624 Table 4-2. Results of N concentration analysis and water content of the samples from the blocks with five different N application rates in 2003. N treatment Value G1 G2 G3 G4 G5 Actual N application (kg ha-1) 156.8 201.6 246.4 291.2 336.0 N Range (g kg-1) 26.2-32.3 26.9-35.3 25.733.1 24.9-35.2 27.9-33.9 Wet N range (g kg-1) 10.2-13.9 9.5-15.1 10.814.3 9.6-14.0 9.8-13.2 Water content range (g g-1) 0.568-0.635 0.507-0.681 0.5570.637 0.567-0.657 0.554-0.666 Average Water content (g g-1) 0.606 0.615 0.607 0.604 0.615 Tables 4-3 and 4-4 show prediction resu lts by PLS and SMLR calibration model on water contents. Good relationships between spectra and water were found in both data sets. The R2 values for validation data set 2002 was 0.76 by SMLR method and 0.74 for validation data set of 2003. Most of the prediction values were smaller than the actual water contents for data set 2002, which caused the SEP values to be no larger than the SEC values in table 4-3. Water Effect on Nitrogen Prediction The correlation coefficients between wate r content and N concentration were 0.07 for the data set from 2002 and -0.14 for the data set from 2003. No direct relationship between leaf water content and N concentrati on was shown by the analysis results. This indicates that leaves with the same N concen trations would have di fferent water contents

PAGE 52

38 that reflect different spectra. The variability of water cont ent did introduce error to the N prediction procedure. -10 0 10 20 30 40 50 60 70 80 400700100013001600190022002500 Wavelength (nm)Reflectance (%) Dry Fresh Difference Water content 0.607g/g Figure 4-3. Reflectance spectra of fresh, dried sample and th eir difference spectrum with water content of 0.607 g g-1. Table 4-3. Prediction of water content usi ng absorbance spectra with PLS and SMLR procedures for data set 2002. R2 SEC (g kg-1)SEP (g kg-1) RMSD (g kg-1) Method Factors / Wavelengths CalibrationValidationCalibrationValidation CalibrationValidation PLS 6 factors 0.78 0.73 7.7 7.7 7.4 8.8 SMLR 2305, 1885, 1234, 1704, 740, 1393 0.80 0.76 7.4 7.3 7.2 8.7 Table 4-4. Prediction of water content usi ng absorbance spectra with PLS and SMLR procedures for data set 2003. R2 SEC (g kg-1) SEP (g kg-1) RMSD (g kg-1) Method Factors / Wavelengths CalibrationValidationCalibration Validation CalibrationValidation PLS 5 factors 0.85 0.66 13.8 17.8 17.9 31.2 SMLR 1379, 1705, 1575 0.86 0.74 11.5 15.5 12.8 24.2 Converting dry basis N concentration to wet basis N concentration or including water content as a predictor in the PLS procedure would add water information to the calibration model development. Figures 44 and 4-5 show spectra of correlation coefficients for both data sets, whic h used dry basis N concentration (Nd ) and wet basis N concentration (Nf ) as predictors. Figure 4-4 used dry basis N concentration, while

PAGE 53

39 figure 4-5 used wet basis N concentration. In fi gure 4-4, correlation co efficient spectra of both years were very different, although p eaks of the two spectra were matched at wavelengths around 710 and 1940 nm, but spectru m of data set 2003 was shifted up from the spectrum of data set 2002 in the near -infrared range. By converting dry basis N concentration to wet basis N concentration, in figure 4-5, spectra of both years have very similar shape with matching wavelengths around 550 nm, 730 nm, 1450 nm and 1890 nm. It may indicate that using wet basis N con centration is better th an using dry basis N concentration. The PLS analysis results for both years are listed in tables 4-5 and 4-6. Compared with only using dry basis N concentration as predictor, results by the PLS procedure based on wet basis N concentration, and usi ng N and water contents as predictors indicated good predictio n performances. For data set 2002 (t able 4-5), results were very similar for the three methods. PLS I and PLS II had almost the same results by yielding the same RMSD, SEC and SEP value. The re sults based on wet N concentration had a different scale that could not compare di rectly with the other two methods. The R2 for the validation data set on wet basis N concentration was lower than the other two methods. The wet basis N range was from 7.8 g kg-1 to 13.3 g kg-1. The RMSD for wet basis was 0.5 g kg-1, which took 8.6% of the total wet basi s N range, while the RMSD for dry basis was 1.2 g kg-1 which also took 8.6% of the total dr y basis N range. This shows that the three methods produced very similar results.

PAGE 54

40 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 400700100013001600190022002500 Wavelength (nm)Correlation coefficient 2002 2003 710 Figure 4-4. Correlation coefficient between absorbance and N (g kg-1) of dried samples of training data sets of 2002 and 2003. -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 400700100013001600190022002500 Wavelength (nm)Correlation coefficient 2002 2003 550 730 1890 1450 Figure 4-5. Correlation coefficient between absorbance and N (g kg-1) of fresh samples of training data sets of 2002 and 2003. The R2 value of PLS results for data se t 2003 (0.49, 0.59 and 0.66, respectively) are not as good as results of the 2002 year (0.83, 0.83 and 0.80, respectively) partly because of smaller N range for data set 2003, which was 24.9-35.3 g kg-1. However, the RMSD and SEP for dry basis N pr ediction were both 1.2 g kg-1, which were similar to data set 2002, the RMSD and SEP for wet basis N were both 0.5 g kg-1 which were the same as data set 2002. The R2 for validation data set on wet ba sis was 0.66, which was better than

PAGE 55

41 the other two methods. Also using RMSD va lue divided by N range, the RMSD for wet basis was 0.5 g kg-1, which took 8.8% of the total wet basis N range, while the RMSD for dry basis was 1.2 g kg-1, which also took 11.5% of the to tal dry basis N range. Since the smaller value of RMSD indicate better results, PLS results based on wet N were superior to dry basis N and PLS II methods for data set 2003. Table 4-5. N prediction by PLS using two pr edictors (water and N contents) and one predictor (Dry basis N or we t basis N) for data set 2002. R2 SEC (g kg-1)SEP (g kg-1) RMSD (g kg-1) Predictors Factors CalibrationValidation Calibration Validation Calibration Validation Water & Dry N 8 0.88 0.83 1.1 1.2 1.0 1.2 Dry N 7 0.89 0.83 1.1 1.2 1.0 1.2 Wet N 8 0.91 0.80 0.4 0.5 0.4 0.5 Table 4-6. N prediction by PLS using two pr edictors (water and N contents) and one predictor (Dry basis N or we t basis N) for data set 2003. R2 SEC (g kg-1)SEP (g kg-1) RMSD (g kg-1) Predictors Factors CalibrationValidation Calibration Validation Calibration Validation Water & Dry N 6 0.59 0.49 1.2 1.3 1.1 1.2 Dry N 8 0.75 0.59 0.9 1.2 0.9 1.2 Wet N 6 0.57 0.66 0.6 0.5 0.6 0.5 Analysis results by SMLR were listed in tables 4-7 and 4-8. The best result was shown by data set 2002 using wet basis N, and the RMSD, R2, and SEP were 0.3 g kg-1, 0.91, and 0.3 g kg-1, respectively. These were better than dry basis SMLR and PLS analysis (table 4-7). The SMLR prediction on dry basis is not as good as results of PLS; the RMSD is 1.5 g kg-1 for the validation 2002 data set by dry basis, while 1.2 g kg-1 for PLS analysis. For data set 2003, R2 for validation data set by wet basis N prediction was better than dry basis N prediction. However the RMSD was 0.6 g kg-1 which is worse than PLS

PAGE 56

42 prediction for data set 2002. By removing scale difference, wet basis N prediction of data set is not as good as dry basis N prediction. Table 4-7. SMLR analysis results of da taset 2002. R2 SEC (g kg-1)SEP (g kg-1) RMSD (g kg-1) Predictor Wavelengths CalibrationValidationCalibration Validation Calibration Validation Dry N 1773, 719, 1810, 2231, 2101 0.86 0.74 1.2 1.4 1.1 1.5 Wet N 1489, 2233, 1954, 2102, 729 0.90 0.91 0.4 0.3 0.4 0.3 Table 4-8. SMLR analysis results of dataset 2003. R2 SEC (g kg-1)SEP (g kg-1) RMSD (g kg-1) Predictor Wavelengths CalibrationValidationCalibrationValidation CalibrationValidation Dry N 561, 537, 1653, 1683, 1378, 1618, 2202, 1864 0.70 0.55 1.0 1.2 1.0 1.2 Wet N 2085, 2230, 706, 1949, 2214, 1479, 1465, 2424 0.62 0.62 0.6 0.6 0.6 0.6 Summary In this chapter, we studied the effect of water on the relationship between spectral reflectance and N concentration. Two sets of ci trus leaf were collected from different years and different varieties. Correlation coefficient spectrum, PLS, and SMLR were used to study water content effect on N sensing. In the PLS method, two constituents, water, and N contents were used as predicto rs. Instead of normalizing spectra for water effect, wet basis N concentrati ons of fresh leaf samples were converted from dry basis N concentrations of dried leaves based on wei ght. The results showed that using wet basis N concentration as a predictor could generate better result s by SMLR for data set 2002. Also correlation coefficient spectra based on wet N had similar shape for both data sets. However, for PLS analysis results using wate r and N contents, these two predictors did not show better performance than ju st using N content as a predictor.

PAGE 57

43 CHAPTER 5 NITROGEN SENSING SYS TEM FOR CITRUS LEAF Based on selected wavelengths, design of a sensing system is discussed in this chapter. Four design criteria are described followed by a general in troduction. Then the sensor testing results give more details for the sensor performance. Introduction Traditional leaf sampling is an expensive, slow, and labor intensive procedure. For highly efficient crop management, it would be desirable to obtain leaf nutrition by real time sensing. Spectral-based sensing methods have been largely used to detect plant status such as moisture content and nutrient stress by utilizing visi ble and near infrared spectral response from plant leaves or canopi es (Card et al. (19 88), Blackmer et al. (1994), and Yoder and Pettigrew-Crosby (1995)). There have been many investigations to implement an N sensing system for real -time field use. Sudduth and Hummel (1993) developed a portable spectrophotometer for so il organic matter analysis by using a circular variable filter (CVF ), a mono-chromator, a lead su lfide (PbS) photodetector, fiber optics, a lamp, an aluminum housing, a amplif ier and a A/D converter, etc. The sensing system had a spectral response range of 1650-2650 nm and bandwidth of 55 nm. The system stability of the ratio to the first m easurement was within 2%. Stone et al. (1996) developed a sensor system for N and weed detection using phot odiode detectors and interference filters (red and NIR). They tested the sensor with winter wheat and reported that total forage N uptake was highly co rrelated with NDVI (Normalized Difference Vegetation Index). Sui and Thomasson (2004) developed a spectral reflectance sensor to

PAGE 58

44 detect N status of cotton plants with four spectral bands of blue, green, red, and NIR light. The N status of the cotton plants were able to be accurately divided into two categories using neural netw ork. Lee et al. (1999) found that SPAD (Soil and Plant Analyzer Development, Minolta Inc.) readi ngs, which were based on transmittance at 659 nm and 940 nm, were well correlated with N content in corn ear leaves (R2 = 0.962). During in-field reflectance spectra sensi ng, many factors, such as intensity of sunlight, incidence angle of th e sunlight, reflectance angle, cloud cover, distance between sensor and plants, density of canopy and soil background, etc. could significantly affect the reflectance measurement. The major issue is how to normalize or minimize the effect of sensing environment on reflected ra diation. Tumbo et al. (2002) used a back-propagation neural network model to set up a calibration model which could remove the limitation of cloud cover and sola r angles considerabl y. The model showed good correlation between pr edicted and actual chlorophyll me ter readings of the training set (R2 = 0.91). A good relationship was also found in the validation data set (R2 = 0.74). Spectral properties of citrus leaf we re studied by Min and Lee (2005). They reported good relationship between reflectance spectra of citrus leaf and their N contents (R2 = 0.828). Some wavelengths, i.e., 448, 669, 719, 1377, 1773, and 2231 nm, were identified as significant wavelengths fo r N detection. Wavelengths of 448, 669, 719 nm were related to chlorophyll N, 1377 nm was related to water band, and 2231 nm was related to protein N. Kokaly (2001) also re ported that chlorophyll and protein were two main N sources that should be consid ered for N concentration prediction.

PAGE 59

45 Objective The objective of this research is to desi gn an in-field spectral-based N sensing system that could detect N c oncentration in citrus leaf. Th e specific sub-objectives were to: Determine design criteria for an N sensing system Fabricate the sensing system Design algorithm to predict N content of unknown samples Testing the N sensing system in a laboratory environment. Design Criteria Wavelength Selection For successful NIR measurement, a prelimin ary work for N sensor design was to identify significant wavelengths. Based on the study results by Min and Lee (2005), wavelengths near 719 nm played the most im portant role for N prediction (r=0.55). Many important wavelengths were also found in the near infrared region. In figure 5-1, chlorophyll related wavelengths are mostly found in the visibl e range, and protein related wavelengths are concentrated in the near-i nfrared range. For better N prediction, visible (VIS) and near infrared (NIR) detectors w ould be used for covering these wavelength ranges. According to the commercially availa ble detectors and linea r variable filters (LVFs), the wavelength ranges for the sensi ng systems were set to 620-1080 nm and 1400-2500 nm. No Moving Part The requirements for implementing an in-f ield NIR spectrometry in precision agriculture applications are ve ry rigorous. Applications of ten need rapid acquisition and processing of broad spectra. Operating conditions during in-field use of NIR spectrometers may include wide temperatur e fluctuations, dust, humidity, vibration, a

PAGE 60

46 moving target, variable target distances, and stray background light. There are many modern NIR spectrometers available in a wide range of configurations which are used in different principles of operation. These spec trometers generally have broad wavelengths band than required and at high accuracy, such as, FieldSpec (ASD, Inc. Boulder, CO) and Cary 500 (Varian, Inc. Palo Alto, CA). This excess capability results in high cost. A single NIR detector and a rotated grating are used in both spectrometers. A single detector might be less expensive than a de tector array; however, complex mechanical structure causes the whole system stay in a big volume and heavy weight. Also the rotated grating is vulnerable to vibration. Dete ctor array coupled with a fixed grating or a filter provides a better choice for the field NIR spectrometry. Figure 5-2 illustrates an efficient spectrometer using a linear variable filter (LVF) and a de tector array. Incident light of different wavelengths pass the LVF at particular positions and is captured by pixels underneath. 0 0.2 0.4 0.6 0.8 1 1.2 40055070085010001150130014501600175019002050220023502500 Wavelength (nm)Absorbance (Abs) 5N31, N=32.3 g/kg 1N04, N=22.3 g/kg 2054 550 660 450 1450 1940 710 Chloro p h y ll Protein Wate r 2180 Figure 5-1. Selected wavelengths range for N detection. Chlorophyll related wavelengths are concentrated in the visible range, and protein relate d wavelengths are concentrated in the near-infrared range . The wavelength of the light which passes th rough the LVF at any point is a linear function of the position of that point relative to the filter. Once the filter is fixed on the

PAGE 61

47 detector, wavelength can be known by ch ecking pixel number of the detector. Wavelength resolution depends on the length of LVF and the number of pixel in the detector array. (a) (b) Figure 5-2. A spectrometer built with a linear variable filter (LVF) and a detector array to avoid moving parts. (a) Light goes thr ough the LVF and is captured by pixels underneath. Each pixel is according to different wavelength. (b) A 256 pixel detector array (S8377-256Q, Hamamastu, Inc.) and a LVF (JDSU, Inc.) are used for wavelength range 620-1080 nm. Single Leaf Detection During in-field reflectance spectra sensi ng, many factors, such as intensity of sunlight, incidence angle of th e sunlight, reflectance angle, cloud cover, distance between sensor and plants, density of canopy and soil background, etc. could significantly affect the reflectance measurement. Many researches were examining N deficiency by measuring reflectance of canopy/plant l eaf (Lee and Searcy (2000), and Sui and Thomasson (2004)). They continuously coll ected reflectance spectra and predicted N concentration. It was reported that good resu lts would be obtained only when N statuses were separated into few categories (e.g., high and low). It would be difficult to predict more accurate N concentration with canopy size detection. For citrus, spring flush from four to six months old with stable N content were used for leaf sampling. Among citrus groves receiving N applica tion rates of 156-336 kg ha-1, the leaf N range was 24.9-35.3

PAGE 62

48 g·kg-1, separation of reflectance spectra wa s 13.1-19.4% at 550 nm, and 20.9-29.9% at 710 m in laboratory environment (Min, et al. 2004). Measurements on canopy size mixed with old and new leaves that will average variability of reflec tance spectra due to different N distribution. Bac kground noise and other factors would also overwhelm N information contained in spectra. Lee and S earcy (2000) described the difficulties of N prediction on canopy size. For better N predic tion, single leaf and sealed reflectance housing should be used for reflectance spectr a measurement. Single leaf detection will remove mixture effect of different leaves, and reflectance housing will block environmental noises listed above. Measur ement on single leaf will provide more stabilized leaf reflectance spectra. Diffuse Reflectance Measurement Diffuse reflectances are mostly interested in reflection spectroscopy since they are causing fewer measurement problems while maintaining good signa l-to-noise ratio. Diffuse reflectance contains information of ab sorption when light penetrates leaf surface and is partially absorbed. However, specular reflectance mostly is the light bounced back from leaf surface which has no much informa tion. Norris et al. (1976) gave a reflectance measurement model that a monochromatic li ght vertically illuminated a sample and reflected radiation was collected with four lead-sulfide detectors at 45 degrees equal spacing around the incident beam. In figure 5-3, Hatchell (1999) also described an easy way to avoid specular components (solid li nes) and measure only diffuse components (dashed lines) that the spectrometer sensor input was set far off the specular plane. Figures 5-4 and 5-5 show the design of a reflectance housing. The reflectance housing (100×104×62 mm), made by black Delrin sheet (9.5 mm thickness), could well block environmental influence. Light sour ce illuminates sample from back of the

PAGE 63

49 reflectance housing, NIR and VIS de tectors are arranged at bot h sides of the reflectance housing. Figure 5-3. Model for diffuse reflectance measurement. Sensor Structure and Component The basic components of the N sensing syst em (figures 5-4 and 5-5) include a light source, two detectors, two linear variable fi lters, a power supply, a reflectance housing, two data acquisition cards and a laptop. A 6.5 W, 12 V, tungs ten halogen lamp was used as the illumination source (LS-1, Oceanoptics, Inc. Dunedin, FL). It provided wavelength range from 400 nm to 2600 nm (figure 5-6). The light source was mounted at the back of reflectance housing with 45° angle to avoid specular reflectance. A focus lens was mounted in front of the lamp to focus the lamp image onto the sample plane. The size of sample port was 25×45 mm. The light sour ce and the reflectance housing provided a stable and closed measurement environment. A HC235-256 image sensor kit (Hamamatsu, Inc. Shizuoka Pref, Japan) was used for NIR detection. It included an NIR detector, (InGaAs G9208-256W, 900-2550 nm, Hamamatsu, Inc. Shizuoka Pref, Japan), a det ector head which had a two stage TE cooled system, a power supply/interface box and a da ta acquisition card with a 16-bit ADC (NI 6036E, Inc. Austin, TX). A LVF14002500 (1400-2500 nm, JDSU, Inc. San Jose, CA)

PAGE 64

50 and a longpass filter with a cutoff wavelengt h of 900 nm (Edmund, Inc. Barrington, NJ) kept wavelength range between 1400-2500 nm. A CMOS linear array S8377-256Q (200-1100 nm, Hamamatsu, Inc. Shizuoka Pref, Japan) and a LVF6201080 (620-1080 nm, JDSU, Inc. San Jose, CA) were used as VIS sensor with wavelength range of 620-1080 nm . A drive circuit (C9001, Hamamatsu, Inc. Shizuoka Pref, Japan) was used to generate necessary signals (clock and start) and amplify the output. A data acquisition card with a 16-bit ADC (NI 6036E, Inc. Austin, TX) was used to control the detector. Figure 57 illustrates spectral responsivity for both detectors. It was found that sp ectral response curves of bot h detectors were compensated with light source intensity curve which help ed improve the performance of the sensing system. At wavelength range of 900-1000 nm, both detectors had low responsivity, and the light source had the highest light intens ity. Also at wavelengths of 600 nm and 2400 nm, detectors had high responsivity and th e light source had lower light intensity. Both detectors had 256 pixels with ac tive sensing area of 0.5×12.5 mm. The LVFs were attached in front of the detectors w ith filter covers made by Delrin, with a 3×12.5 mm slit on it. Transmittance rate for both LVFs at each position was approximately 60%. The designed field of view (FOV) of each detector was 18 degrees.

PAGE 65

51 Figure 5-4. Design scheme of the N sens ing system. A light source is 45 degree illuminating sample from back of the reflectance housing; reflected radiation is collected by the NIR and the VIS dete ctors arranged at both sides of the reflectance housing. Figure 5-5. A picture of the actual N sensor.

PAGE 66

52 0 500 1000 1500 2000 2500 3000 2004006008001000120014001600180020002200240026002800 Wavelength (nm)Normalized Intensity (uW/cm2/nm) Figure 5-6. Light intensity of the light s ource. It covers wave length range of 400-2600 nm, and has peak intensity at 950 nm. (a) (b) Figure 5-7. Spectral sensitivity for (a) VIS de tector S8377-256Q and (b) NIS detector G9208-256W. Data Acquisition System For the NIR detector, the HC235-256 imag e sensor kit had a well developed software package based on LabView. The kit utilized a multifunction data acquisition card (NI 6036E DAQ) with 16-bits ADC, and 16 analog input channels with up to 200 kHz sampling rate. For the VIS detector, we used the same da ta acquisition board. Data acquisition for detectors requires clock and start input signa ls, and detectors will generate trigger and

PAGE 67

53 output video signals. Figure 5-8 illustrates the timing chart of The VIS detector array. The data acquisition board provided two i ndependent counters which could generate pulse train at certain frequenc y. One counter was used as cloc k signal input to the sensor. Another counter was used as a start signal. Sampling rate could be specified by LabView software, and also needs to be synchronized with start signal to ensu re that we can read the output signal correctly. The NIR detector had very similar timing chart with the VIS detector. Figure 5-8. Timing chart for the VIS detector. The LabView data acquisition block diagra m for the VIS detector was shown in figures 5-9. For the VIS dete ctor, a clock frequency of 4 kHz was used. The system started to acquire data at falling edge of the start pulse . The sampling frequency was 100 kHz. Five measurements of a leaf were acq uired. The interfaces for the VIS and the NIR detectors were shown in figures 5-10 and 5-11. Integration time

PAGE 68

54 Figure 5-9. The LabView data acquisition block diagram for the VIS detector. Figure 5-10. Data acquisition interface for the VIS detector.

PAGE 69

55 Figure 5-11. Data acquisition interface for the NIR detector developed by Hamamatsu, Inc. Sensor Testing Results in Lab Environment Wavelength Calibration Wavelength calibration for the LVFs ha s been accomplished by finding the relationship between a known wavelength and its corresponding pixel number. A spectrophotometer (Cary 500, Varian Inc. Palo Alto, CA) was used to generate single wavelengths which could be used for wave length calibration. In figure 5-12, the VIS detector array with the LVF was placed at th e sample port of the spectrophotometer. Dark current was measured first. Then monochrom ic light generated by the spectrophotometer illuminated the active area, and the pixel which had the highest output after subtracting dark output would correspond to the input wavelength.

PAGE 70

56 Figure 5-12. Wavelength calibration for the VIS detector using Cary 500. Figures 5-13 and 5-14 show th e testing results of the wave length calibration for the VIS detector. The test result showed very good linearity of the LVF in figure 5-15 (r = 0.999). Wavelength prediction error was less than 2 nm (figure 5-15). Wavelengths below 800 nm could not be detected due to very lo w light intensity of the spectrophotometer below 800 nm. The right end of the detector around 1000 nm had lower light intensity than 900 nm due to the very low responsivity of the VIS detector. The full width at half maximum (FWHM) of the LVF for the VIS detector was from 12 nm to 27 nm. Figures 5-16 and 5-17 were th e wavelength calibration for the NIR detector with the LVF. It also showed good linear relationship. The FWHM of the LVF for the NIR detector was between 20 nm and 30 nm . Wavelength error was less than 5 nm (figure 5-18). Using wavelength predicti on equations 5-1 and 5-2, the predicted wavelengths for the VIS and the NIR detect ors were listed in tables 5-1 and 5-2. Wavelength (VIS) = 560.554 + Pixel ×1.995 (5-1) Wavelength (NIR) = 1264.432 + Pixel × 5.086 (5-2) Cary 500 VIS detector A/D convert

PAGE 71

57 1 1.5 2 2.5 3 3.5 1173349658197113129145161177193209225241 PixelVoltage (V) 801 820 840 860 880 900 920 940 960 980 1000 801 840 880 920 960 1000 Figure 5-13. Pixel response for differe nt wavelength for the VIS detector. 700 750 800 850 900 950 1000 1050 100150200250 PixelWavelength (nm)r = 0.999 Figure 5-14. Relationship between monochrome light and pixel number. Linearity of the VIS detector is 0.999. -2 -1 0 1 2 100120140160180200220240 PixelError (nm) Figure 5-15. Errors between pr edicted wavelengths and actual wavelengths. The error is less than 2 nm.

PAGE 72

58 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 1173349658197113129145161177193209225241 PixelVoltage output (V) 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 Figure 5-16. Wavelength calibration for the NI R detector. Pixel res ponses were plotted according to different wavelength. 1400 1600 1800 2000 2200 2400 064128192256PixelWavelength (nm)r =0.999 Figure 5-17. Relationship between monochr ome light and pixel number for the NIR detector. The linearity is 0.999. -6 -4 -2 0 2 4 050100150200250 PixelError (nm) Figure 5-18. Errors between pr edicted wavelengths and actual wavelengths. The error is less than 5 nm.

PAGE 73

59 Table 5-1. Predicted wavelengths for th e VIS detector using pixel number. Pixel Wavelength Pixel Wavelength Pixel Wavelength Pixel Wavelength Pixel Wavelength 1 562.55 52 664.30 103 766. 05 154 867.81 205 969.56 2 564.54 53 666.30 104 768. 05 155 869.80 206 971.55 3 566.54 54 668.29 105 770. 04 156 871.80 207 973.55 4 568.53 55 670.29 106 772. 04 157 873.79 208 975.54 5 570.53 56 672.28 107 774. 03 158 875.79 209 977.54 6 572.52 57 674.28 108 776. 03 159 877.78 210 979.53 7 574.52 58 676.27 109 778. 02 160 879.78 211 981.53 8 576.52 59 678.27 110 780. 02 161 881.77 212 983.52 9 578.51 60 680.26 111 782. 01 162 883.77 213 985.52 10 580.51 61 682.26 112 784. 01 163 885.76 214 987.51 11 582.50 62 684.25 113 786. 01 164 887.76 215 989.51 12 584.50 63 686.25 114 788. 00 165 889.75 216 991.50 13 586.49 64 688.24 115 790. 00 166 891.75 217 993.50 14 588.49 65 690.24 116 791. 99 167 893.74 218 995.49 15 590.48 66 692.23 117 793. 99 168 895.74 219 997.49 16 592.48 67 694.23 118 795. 98 169 897.73 220 999.49 17 594.47 68 696.22 119 797. 98 170 899.73 221 1001.48 18 596.47 69 698.22 120 799. 97 171 901.72 222 1003.48 19 598.46 70 700.21 121 801. 97 172 903.72 223 1005.47 20 600.46 71 702.21 122 803. 96 173 905.71 224 1007.47 21 602.45 72 704.20 123 805. 96 174 907.71 225 1009.46 22 604.45 73 706.20 124 807. 95 175 909.70 226 1011.46 23 606.44 74 708.19 125 809. 95 176 911.70 227 1013.45 24 608.44 75 710.19 126 811. 94 177 913.69 228 1015.45 25 610.43 76 712.18 127 813. 94 178 915.69 229 1017.44 26 612.43 77 714.18 128 815. 93 179 917.68 230 1019.44 27 614.42 78 716.18 129 817. 93 180 919.68 231 1021.43 28 616.42 79 718.17 130 819. 92 181 921.67 232 1023.43 29 618.41 80 720.17 131 821. 92 182 923.67 233 1025.42 30 620.41 81 722.16 132 823. 91 183 925.66 234 1027.42 31 622.40 82 724.16 133 825. 91 184 927.66 235 1029.41 32 624.40 83 726.15 134 827. 90 185 929.66 236 1031.41 33 626.39 84 728.15 135 829. 90 186 931.65 237 1033.40 34 628.39 85 730.14 136 831. 89 187 933.65 238 1035.40 35 630.38 86 732.14 137 833. 89 188 935.64 239 1037.39 36 632.38 87 734.13 138 835. 88 189 937.64 240 1039.39 37 634.37 88 736.13 139 837. 88 190 939.63 241 1041.38 38 636.37 89 738.12 140 839. 87 191 941.63 242 1043.38 39 638.36 90 740.12 141 841. 87 192 943.62 243 1045.37 40 640.36 91 742.11 142 843. 86 193 945.62 244 1047.37 41 642.35 92 744.11 143 845. 86 194 947.61 245 1049.36 42 644.35 93 746.10 144 847. 85 195 949.61 246 1051.36 43 646.35 94 748.10 145 849. 85 196 951.60 247 1053.35 44 648.34 95 750.09 146 851. 84 197 953.60 248 1055.35 45 650.34 96 752.09 147 853. 84 198 955.59 249 1057.34 46 652.33 97 754.08 148 855. 84 199 957.59 250 1059.34 47 654.33 98 756.08 149 857. 83 200 959.58 251 1061.33 48 656.32 99 758.07 150 859. 83 201 961.58 252 1063.33 49 658.32 100 760.07 151 861. 82 202 963.57 253 1065.32 50 660.31 101 762.06 152 863. 82 203 965.57 254 1067.32 51 662.31 102 764.06 153 865. 81 204 967.56 255 1069.32 256 1071.31

PAGE 74

60 Table 5-2. Predicted wavelengths for th e NIR detector using pixel number. Pixel Wavelength Pixel Wavelength Pixel Wavelength Pixel Wavelength Pixel Wavelength 1 1264.43 52 1523.83 103 178 3.23 154 2042.64 205 2302.04 2 1269.52 53 1528.92 104 178 8.32 155 2047.72 206 2307.12 3 1274.60 54 1534.01 105 179 3.41 156 2052.81 207 2312.21 4 1279.69 55 1539.09 106 179 8.49 157 2057.89 208 2317.30 5 1284.78 56 1544.18 107 180 3.58 158 2062.98 209 2322.38 6 1289.86 57 1549.26 108 180 8.67 159 2068.07 210 2327.47 7 1294.95 58 1554.35 109 181 3.75 160 2073.15 211 2332.56 8 1300.04 59 1559.44 110 181 8.84 161 2078.24 212 2337.64 9 1305.12 60 1564.52 111 182 3.93 162 2083.33 213 2342.73 10 1310.21 61 1569.61 112 182 9.01 163 2088.41 214 2347.81 11 1315.30 62 1574.70 113 183 4.10 164 2093.50 215 2352.90 12 1320.38 63 1579.78 114 183 9.18 165 2098.59 216 2357.99 13 1325.47 64 1584.87 115 184 4.27 166 2103.67 217 2363.07 14 1330.55 65 1589.96 116 184 9.36 167 2108.76 218 2368.16 15 1335.64 66 1595.04 117 185 4.44 168 2113.84 219 2373.25 16 1340.73 67 1600.13 118 185 9.53 169 2118.93 220 2378.33 17 1345.81 68 1605.21 119 186 4.62 170 2124.02 221 2383.42 18 1350.90 69 1610.30 120 186 9.70 171 2129.10 222 2388.50 19 1355.99 70 1615.39 121 187 4.79 172 2134.19 223 2393.59 20 1361.07 71 1620.47 122 187 9.87 173 2139.28 224 2398.68 21 1366.16 72 1625.56 123 188 4.96 174 2144.36 225 2403.76 22 1371.24 73 1630.65 124 189 0.05 175 2149.45 226 2408.85 23 1376.33 74 1635.73 125 189 5.13 176 2154.53 227 2413.94 24 1381.42 75 1640.82 126 190 0.22 177 2159.62 228 2419.02 25 1386.50 76 1645.90 127 190 5.31 178 2164.71 229 2424.11 26 1391.59 77 1650.99 128 191 0.39 179 2169.79 230 2429.19 27 1396.68 78 1656.08 129 191 5.48 180 2174.88 231 2434.28 28 1401.76 79 1661.16 130 192 0.56 181 2179.97 232 2439.37 29 1406.85 80 1666.25 131 192 5.65 182 2185.05 233 2444.45 30 1411.93 81 1671.34 132 193 0.74 183 2190.14 234 2449.54 31 1417.02 82 1676.42 133 193 5.82 184 2195.22 235 2454.63 32 1422.11 83 1681.51 134 194 0.91 185 2200.31 236 2459.71 33 1427.19 84 1686.59 135 194 6.00 186 2205.40 237 2464.80 34 1432.28 85 1691.68 136 195 1.08 187 2210.48 238 2469.89 35 1437.37 86 1696.77 137 195 6.17 188 2215.57 239 2474.97 36 1442.45 87 1701.85 138 196 1.26 189 2220.66 240 2480.06 37 1447.54 88 1706.94 139 196 6.34 190 2225.74 241 2485.14 38 1452.63 89 1712.03 140 197 1.43 191 2230.83 242 2490.23 39 1457.71 90 1717.11 141 197 6.51 192 2235.92 243 2495.32 40 1462.80 91 1722.20 142 198 1.60 193 2241.00 244 2500.40 41 1467.88 92 1727.29 143 198 6.69 194 2246.09 245 2505.49 42 1472.97 93 1732.37 144 199 1.77 195 2251.17 246 2510.58 43 1478.06 94 1737.46 145 199 6.86 196 2256.26 247 2515.66 44 1483.14 95 1742.54 146 200 1.95 197 2261.35 248 2520.75 45 1488.23 96 1747.63 147 200 7.03 198 2266.43 249 2525.83 46 1493.32 97 1752.72 148 201 2.12 199 2271.52 250 2530.92 47 1498.40 98 1757.80 149 201 7.20 200 2276.61 251 2536.01 48 1503.49 99 1762.89 150 202 2.29 201 2281.69 252 2541.09 49 1508.57 100 1767.98 151 202 7.38 202 2286.78 253 2546.18 50 1513.66 101 1773.06 152 203 2.46 203 2291.86 254 2551.27 51 1518.75 102 1778.15 153 203 7.55 204 2296.95 255 2556.35 256 2561.44

PAGE 75

61 Dark Current and Signal Output The dark current is a major source of noise for measurement. There exists a small current when a reverse voltage is applied to a device even in a dark state. The dark current was measured at a totally dark envir onment without any light input. Dark current should be subtracted from the measured si gnals of samples to obtain actual signals. Figure 5-19 shows the dark voltage output of the VIS de tector with an integration time of 130 ms. The dark voltage was approximatel y 1.129 V for the VIS detector. For the NIR detector, the integration time was 100 ms, and the output volta ge is shown in figure 5-20, where the pixel 128 has a peak signal output. Figures 5-21 and 5-22 were the signal output for a leaf measurement by subtracting the dark output. For the VIS detector, because of very low leaf reflectance below 680 nm and low spectral responsivity above 950 nm, there was no signal output below pixel 65 and above pixel 195 (figure 5-21). 1.124 1.126 1.128 1.13 1.132 1173349658197113129145161177193209225241 Pixel Output Voltage (V) Figure 5-19. Dark voltage output of the VIS detector using 130 ms integration time.

PAGE 76

62 0.00 1.00 2.00 3.00 4.00 5.00 11835526986103120137154171188205222239256PixelVoltage output (V) Figure 5-20. Dark voltage output of the NIR detector using 100 ms integration time. -0.5 0 0.5 1 1.5 21173349658197113129145161177193209225241PixelSignal output (V) Figure 5-21. Signal output of leaf measuremen t by the VIS detector subtracting the dark. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1173349658197113129145161177193209225241PixelSignal output (V) Figure 5-22. Signal output of leaf measuremen t by the NIR detector subtracting the dark.

PAGE 77

63 Noise and Signal to Noise Ratio Noise for each pixel was evaluated by calcu lating standard deviation of each pixel with ten measurements of one leaf. Figure 523 shows the noise of the VIS detector. The noise level of pixels from 65 to 191 is higher than other pixels because only pixels from 65 to 191 have signal output when measuring le af reflectance. Figure 5-24 shows the leaf reflectance measured by the VIS detector. Du e to low leaf reflect ance at range 500-660 nm and low incident light intensity, there is no signal output from pixels 1 to 64. There is also no signal output for pixels 192 to 256 because of low spectral responsivity beyond 1000 nm for the VIS detector. The mean noise level for the VIS detector is 0.0034 V. Figure 5-25 shows the noise of the NIR det ector. The mean noise level for the NIR detector is 0.0013 V which is less than the one of the VIS detector. Signal to noise ratio was calculated by e quation 5-3. Mean is the average value of ten measurements. Stdev is the standard deviation of the ten measurements. stdev mean SNR (5-3) Figure 5-26 shows the signal to noise ratio fo r the VIS detector. Figure 5-27 shows the signal to noise ratio for the NIR detector. For the VIS detector, the si gnal to noise ratio is 337 for the VIS detector at pixel 70 which is equal to 700 nm. Pixels from 70 to 150 have signal to noise ratio larger than 200. For the NI R detector, most of the pixels have signal to noise ratio larger than 100.

PAGE 78

64 0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 1173349658197113129145161177193209225241 PixelStdev (V) Figure 5-23. Noise for the VIS detector calcu lated using ten measurements of a single leaf. -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 11835526986103120137154171188205222239256PixelVoltage (V) Figure 5-24. A reflectance spectr um of a leaf measurement. 0 0.001 0.002 0.003 0.004 0.005 0163248648096112128144160176192208224240256 PixelStandard deviation (V) Figure 5-25. Noise for the NIR detector calcu lated using ten measurements of a single leaf.

PAGE 79

65 -50 0 50 100 150 200 250 300 11835526986103120137154171188205222239256PixelS/N Figure 5-26. Signal to noise ratio for the VIS detector. 0 100 200 300 0163248648096112128144160176192208224240 PixelS/N Figure 5-27. Signal to noise ra tio for the NIR detector. Linearity Maintaining a linear relationship between the input light intensity and output signal is very important for leaf measurement. A good signal to noise ratio is obtained by increasing the system output as high as po ssible within the system linear range. The linearity of the sensor was measured by studyi ng the linear relati onship between voltage outputs with different integration time. The in tegration time is defined as the time period between two start pulses, which includes pixe l read out time and exposure time. The pixel read out time is related to clock frequenc y. Since the clock frequency was set as a constant, the read out time is fixed. Adjusting the integration time is equal to adjusting the exposure time. Longer exposure time will di rectly result in higher voltage output. For

PAGE 80

66 the VIS detector, the integrat ion times of 100 ms to 170 ms with 10 ms increase were used. Figure 5-28 shows the voltage outputs for on e leaf with different integration times. Figure 5-29 shows the calculated the co rrelation coefficient between voltage outputs and the integration times at each pixe l for the VIS detector. Pixels from 68 to 180 had very good linearity (r >0.99) . For both ends of the detect or, since there is no signal output, there is no linear relati onship. Figure 5-30 shows the linearity of pixel 70, where the r was 0.999, and the residuals were less than 0.01 V (figure 5-31). The linearity of the system was also measured with polytetrafluoroethylene (PTFE), a white reference. Five integrati on times, from 100 ms to 140 ms with 10 ms increment, were used. Since the white refe rence had higher reflectance than leaf, the voltage outputs of some pixels were too high and out of det ector linear range. In figure 533, pixels from 165 to pixel 225 had good linearity with r larger than 0.999. At pixel 165 and 225, the voltage output with 140 ms in tegration time was 3.15 V (figure 5-32), which indicated that the upper level of linear voltage output of the VIS detector was 3.15 V. Voltage output larger than 3.15 V would be out of linear range of the VIS detector. Figure 5-34 shows the voltage outputs for the NIR detector using one leaf with seven different integration times. The NIR det ector used integration times from 100 ms to 160 ms with 10 ms increment. Figure 5-35 shows the calculated linearity at each pixel. The correlation coefficients for most pixels are very close to 1.

PAGE 81

67 1 1.5 2 2.5 3 3.5 1173349658197113129145161177193209225241 PixelVoltage output (V) 100 ms 110 ms 120 ms 130 ms 140 ms 150 ms 160 ms 170 ms Figure 5-28. Voltage outputs for one leaf m easured by the VIS detector with eight different integration times. 0.96 0.97 0.98 0.99 1 1.01 60708090100110120130140150160170180190200PixelCorrelation coefficient (r) Figure 5-29. Linearity for the VIS detector for leaf measurement. 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 80100120140160180Integration time (ms)Voltage output (V)r=0.999 Figure 5-30. Regression line for pixel 70. The r is 0.999.

PAGE 82

68 -0.02 -0.01 0 0.01 0.02 80100120140160180Integration time (ms)Residuals (V) Figure 5-31. Residual for predicted voltage outputs for pixel 70 of the VIS detector. 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 1173349658197113129145161177193209225241Pixel Voltage output (V) 100 ms 110 ms 120 ms 130 ms 140 ms (165, 3.15V) Figure 5-32. Voltage outputs for white refere nce (PTFE) measured by the VIS detector with five different integration times. 0.95 0.96 0.97 0.98 0.99 1 1.01 1173349658197113129145161177193209225241PixelCorrelation coefficient (r) Figure 5-33. Linearity for the VIS detect or for white reference measurement.

PAGE 83

69 1 2 3 4 5 6 7 8 1173349658197113129145161177193209225241 PixelVoltage output (V) 100 ms 110 ms 120 ms 130 ms 140 ms 150 ms 160 ms Figure 5-34. Voltage outputs for one leaf m easured by the NIR detector with seven different integration times. 0.998 0.9985 0.999 0.9995 1 1.0005 0163248648096112128144160176192208224240256PixelCorrelation coefficient (r) Figure 5-35. Linearity for the NIR detector. Stability After two hours system warming up, stabil ity of the system was monitored by 12 repeated measurements of the PTFE disk ove r a 2-hour period with 10 minuets interval. Each measurement was the average of five readings. To compare any differences present, each measurement was normalized by dividing the response of the first measurement. For both detectors, variations were found to be random noise (figur es 5-36 and 5-37). The stability of ratio to the first measurement was within ±0.1% for the VIS detector. While the stability of the NIR detector was w ithin ±0.5% of the first measurement.

PAGE 84

70 0.997 0.998 0.999 1 1.001 1.002 11835526986103120137154171188205222239256 PixelRatio Figure 5-36. Stability test of the VIS de tector during two hours period. Ratios were obtained by dividing each successive measurement to the first scan. 0.99 0.995 1 1.005 1.01 11835526986103120137154171188205222239256PixelRatio Figure 5-37. Stability test of the NIR de tector during two hours period. Ratios were obtained by dividing each successive measurement to the first scan. Leaf Measurement Prior to the diffuse reflectance measuremen t of a sample, a white reference should be obtained to get the reflectance ratio between sample spectra and the white reference. A 50 mm diameter PTFE disk was used to obt ain the optical reference standard. Leaf reflectance was calculated by (leaf-dark)/(ref erence-dark). Figures 5-38 and 5-39 are the reflectance spectrum of a citrus leaf. Since th ere are no data from both ends of the VIS detectors, pixels of 1to 64 and 195 to 256 s hould be removed. The spectra of two leaves obtained by N sensor and the spectrophotometer were plotted togeth er in figure 5-40 by converting the pixel number to wavelength. Compared with actual leaf reflectance spectrum obtained by Cary 500, red edge at pixel 65-78 nm is shown in figure 5-40.

PAGE 85

71 Water bands at pixel 30 and 128 in figure 537 can be recognized. The reflectance of peak at 750 nm obtained by the N sensor is higher than the reflectance at same wavelength measured by spectrophotometer. Since the voltage output of the white reference at 750 nm was out of linear range , the white reflectance at 750 nm was not accurate. The first water band at 1405 nm obtaine d by the N sensor is shifted to the left (red line in figure 5-40) due to blue shift since the angle of the incident light was not perpendicular to the filter plane. -0.2 0 0.2 0.4 0.6 0.8 11835526986103120137154171188205222239256 PixelReflectance (R%)Red edge Figure 5-38. Reflectance spectrum of a citr us leaf measured by the VIS detector. 0 0.1 0.2 0.3 0.4 0.5 1173349658197113129145161177193209225241 PixelReflectance (R%)Water Figure 5-39. Reflectance spectrum of a citr us leaf measured by the NIR detector.

PAGE 86

72 -10 0 10 20 30 40 50 60 70 80 400650900115014001650190021502400Wavelength (nm)Reflectance (R%) Sample 1 Cary 500 Sample 2 Cary 500 Sample 1 N sensor Sample 2 N sensor Figure 5-40. Actual citrus leaf reflectan ce spectrum measured by the spectrophotometer. Development of Calibration Models The calibration model for the system relies on statistical resu lts of a data set collected by the new sensing system. Th e calibration model obtained by Cary 500 spectrophotometer provided wave length selection for the new system. However, different sensing systems have different performan ces. The calibration model developed based on laboratory environment can not be used into the new system dire ctly. A new calibration model should be developed by repeating the cy cle from leaf sampling, data analysis to wavelength selection, etc. In figure 5-41, the design of the sensing system is accomplished in two cycles. The first cycle would be used for sensor wavelength selection, and the second cycle is used for algorithm design. Leaf Sampling and Spectra Measurement For testing the sensing system and de veloping calibration model for N sensing system, a repeated leaf sampling was done in September 2005. A total of 124 samples were collected from the same experiment al grove as year 2003 with the same N

PAGE 87

73 application rates (156.8 (G1), 201.6 (G2), 246.4 (G3), 291.2 (G4), and 336 kg/ha (G5)). The leaves were scanned by Cary 500 spectro photometer and the N sensor. Collected spectra were smoothed using moving average w ith 15 points. Then two or three leaves were combined into one sample. Table 5-3 lists the N concentration an alysis results of the samples collected in 2005. Half of the samples were used fo r calibration, and another half were used for validation. Figure 5-41. Flow chart for N sensing system design which was completed in two flow cycles. The first cycle provides wavele ngth selection for the sensing system, and the second cycle is used for devel oping algorithm for the sensing system. Table 5-3. Results of N concentration analys is of the samples from five different N application rates for samples collected in 2005. N treatment Value G1 G2 G3 G4 G5 Actual N application (kg·ha-1) 156.8 201.6 246.4 291.2 336.0 Number of samples 28 20 25 26 25 Average N (g·kg-1) 25.0 25.9 26.1 26.1 26.9 N Range (g·kg-1) 21.5-27.9 22.0-30.3 20.8-29.1 23.1-29.0 23.1-30.2 Standard deviation (g·kg-1) 1.7 1.7 2.1 1.3 1.9 Statistical Results of Data Se t Measured by Spectrophotometer Since data set 2005 and data set 2003 came from the same experimental citrus grove, results from data set 2005 measured by Cary 500 spectrophotometer had similar

PAGE 88

74 results as data set 2003. In co rrelation coefficient spectrum (figure 5-42), two peaks near 550 nm and 710 nm showed in the visible region. There was no obvious relationship between absorbance and N concentration in wavelength range of 850 to 2500 nm. Peaks in standard deviation spectrum (figure 5-43) almost matched with peaks in figure 3-8 which indicated the similarity of the data sets from different years but the same experimental grove. Table 5-4 lists the SMLR and PLS anal ysis results for data set 2005. The PLS procedure with eight factors generated bette r results than SMLR procedure with seven wavelengths. The SEP and RMSD values were 1.18 g kg-1 and 1.19 g kg-1 for the validation data set using PLS procedure, while 1.22 g kg-1 and 1.21 g kg-1 for the SMLR procedure. Compared with the results fr om data set 2003 which had SEP and RMSD values of 1.19 g kg-1 and 1.20 g kg-1 for the validation data set using PLS procedure with nine factors, the results from data set 2005 wa s very similar to the results from data set 2003. Figure 5-44 shows the predicted N c oncentration using PLS procedure. -0.1 0 0.1 0.2 0.3 0.4 0.5 4005507008501000115013001450160017501900205022002350250 0 Wavelength (nm)Correlation coefficient (r) 710 550 Figure 5-42. Correlation coefficient between absorbance of data set 2005 obtained by spectrophotometer.

PAGE 89

75 0 0.005 0.01 0.015 0.02 0.025 0.03 40055070085010001150130014501600175019002050220023502500Wavelength (nm)Standard deviation 550 708 1450 1891 2020 Figure 5-43. Standard deviation of ab sorbance of data set 2005 obtained by spectrophotometer. Table 5-4. SMLR and PLS analysis result for data sets 2005 measured by Cary 500. 20 22 24 26 28 30 202224262830 Actual N (g/kg)Predicted N (g/kg) Figure 5-44. N prediction for data set 2005 using PLS procedure. Five samples had N prediction error larger than 2 g kg-1. Statistical Results of Data Se t Measured by Nitrogen Sensor For reflectance measurement of each leaf, the averaged data of five scans was obtained. Then the averaged data were sm oothed using moving averaged with seven RMSD (g kg-1 ) R2 Method Selected wavelengths (nm) / factors SEC (g kg-1 ) SEP (g kg-1 )CalibrationValidation Calibration Validation SMLR 692, 1627, 1847, 1838, 1199, 1286, 2247 0.90 1.22 0.84 1.21 0.823 0.390 PLS 8 1.08 1.18 0.99 1.19 0.768 0.475 SEP = 1.18 g kg-1 RMSD = 1.19 g kg-1 R2 = 0.475

PAGE 90

76 points. Two or three smoothed data were combined into one sample. The data set obtained from the N sensor was identified as data set 2005N. In order to remove part of system noise, first derivative was obtained. Ha lf of the samples were used for calibration, and another half were used for validation. Figure 5-45 show s the calculated correlation coefficient (r) spectrum between the 1st deri vative data set and the N concentrations. A peak at 714 nm in the visible range was s hown. Due to system noise, r spectrum in NIR region did not show much useful informati on. Figure 5-46 shows the calculated standard deviation (std) of the data se t and a peak at 718 nm matched with the peak in r spectrum. -0.4 -0.2 0 0.2 0.4 65085010501250145016501850205022502450 Wavelength (nm)Correlation coefficient (r) VIS NIR 714 Figure 5-45. Correlation coefficient spectrum be tween the 1st derivative data set and the N concentrations. 0 0.002 0.004 0.006 0.008 0.01 65085010501250145016501850205022502450 Wavelength (nm)Standard deviation VIS NIR 718 820 Figure 5-46. Standard deviation of the 1st derivative data set.

PAGE 91

77 Since there was no signal output for both ends of the VIS detector, the data used for statistical analysis were restricted in the pixel range from 65 to 100. The converted wavelength ranges covered from 690 nm to 760 nm for the VIS detector and 1264 nm to 2561 nm for the NIR detector. The SMLR and PLS results from 1st derivative of data set 2005N were listed in table 5-5. Figure 5-47 s howed the PLS predicted result of the 1st derivative of data set. Table 5-5. SMLR and PLS analysis results fo r data sets 2005N measured by N sensor. 20 22 24 26 28 30 202224262830 Actual N (g/kg)Predicted N (g/kg) Figure 5-47. N prediction for data set 2005N us ing PLS procedure. Fifteen percent of the samples had prediction N error larger than 2 g kg-1. The SEP and RMSD were 1.87 g kg-1 and 1.86 g kg-1 for the validation data set using PLS procedure, while 2.09 g kg-1 and 2.10 g kg-1 for the SMLR procedure, which were not as good as the results obtained by the spectrophotometer. Some samples had larger errors (>2 g kg-1) which generated R2 value almost close to zero. However, RMSD ( g kg-1 ) R2 Method Selected wavelengths (nm) / factors SEC ( g kg-1 ) SEP ( g kg-1 )CalibrationValidationCalibration Validation SMLR 700, 880, 890, 1676, 1747, 1925, 2205 1.25 2.09 1.17 2.10 0.627 0.030 PLS 6 0.58 1.87 0.55 1.86 0.912 0.083 SEP = 1.87 g kg-1 RMSD = 1.86 g kg-1 R2 = 0.083

PAGE 92

78 considering the N range of the citr us leaf which was about 10 g kg-1, the SEP of 2 g kg-1 took 20% of the whole N range. Actual ly, 75% unknown samples had predicted N concentrations less than 2 g kg-1 error which were in two doted lines in figure 5-47. Discussion Comparing the performance of the N sensor and the spectrophotometer, we can see that the N sensing system still has lot of noise . One possible error source can be the leaf itself. Citrus leaf has wax cuticle (figure 5-48); the uneven surface of citrus leaf can introduce specular reflectance to the syst em. Since reflected light had not been collimated, radiance could come to the LVF through different directions which introduced error resulting from blue shift. Eight repeated measurements of a single leaf at different positions were collected. Standard deviations at each pixel were shown in figures 5-49 and 5-50. The error of repeat ability which was calculated by averaged standard deviation for the VIS detector was 0.022 V which was 6 times of its noise level. And the averaged standard deviation fo r the NIR detector was 0.035 V which was 27 times of its noise level. The actual signal se paration for the VIS detector was 0.2 V, and 0.14 V for the NIR detector. Dividing the volta ge separations of each detector by their error of repeatability, the actual signal to noise ratio should be 9 for the VIS detector and 4 for the NIR detector. That is why the NIR detector was noisi er than the VIS detector. Figure 5-48. Picture of a citrus leaf with uneven cuticle.

PAGE 93

79 0 0.01 0.02 0.03 0.04 0.05 1173349658197113129145161177193209225241PixelStandard deviation (V) Figure 5-49. Standard deviation of repeated measurement of a single leaf at different positions using the VIS detector. 0.02 0.025 0.03 0.035 0.04 0.045 0.05 1163146617691106121136151166181196211226241256PixelStandard deviation (V) Figure 5-50. Standard deviation of repeated measurement of a single leaf at different positions using the NIR detector. 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1173349658197113129145161177193209225241PixelVoltage (V) 0.2 Figure 5-51. Voltage separation fo r the VIS detector is 0.2 V.

PAGE 94

80 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1173349658197113129145161177193209225241PixelVoltage output (V) 0.14 Figure 5-52. Voltage separation fo r the NIR detector is 0.14 V. Summary In this chapter, an in-field N sensing system was developed. The system included visible and near-infrared sens ing ranges which covered wavelengths from 680 to 950 nm and 1400 to 2500 nm. A reflectance housing was designed to block environmental noise and to ensure single leaf measurement. A ha logen light source, two detector arrays, two linear variable filters and two data acquisi tion boards with 16-bit ADC were used to fabricate the sensing system. The designed sy stem has no moving parts, so it is robust and resistant to vibration. The test resu lts showed that the sensing system had good linearity and stability. The VIS detector had linearity (r >0.99), and the NIR detector had linearity (r>0.999). The stability of the VIS de tector was within ±0.1%, and the stability of the NIR detector was within ±0.5%. Us ing first derivative data, the developed calibration model had 75% unknown samples whic h had predicted N concentrations less than 2 g kg-1 error. If the N levels were rated as low (< 25 g kg-1), medium (25~27 g kg-1) and high (>27 g kg-1), the result showed 44.26% samples were misclassified.

PAGE 95

81 CHAPTER 6 FUTURE WORK A spectral-based sensing method has been a promising way for rapid and nondestructive nutrient detection. The results in this research showed that the N sensor has great potential for future use in several directions. Spectroscopic analysis largely relies on having an accurate instrument. Improving the system accuracy could be one thing we n eed to do in the future. However, with the rapid development of MEMS (micro-ele ctro-mechanical systems) technology, a commercially available spectrophotometer will provide many choices of wavelength combination with small volume, high accuracy and low price. For calibration model development, using wi despread N application rate will help build a more robust calibration model. In data sets of 2003 and 2005, two more N application rates, such as 0 kg ha-1and 50 kg ha-1, could be added in the experimental grove for generating some low N concentratio n in citrus leaves. The wider N range in citrus leafs can produce be tter signal to noise ratio and generate better R2 and RMSD values. Combined with a GPS receiver, a nitrogen variation map could be created, and the results will help decide fertilizer application ra te at different locations in citrus groves. If environmental noise could be well normalize d, a more efficient N sensing system could be developed based on canopy size detection. Then sensor-based variable rate application technology could be realized by us ing this real-time N sensor.

PAGE 96

82 LIST OF REFERENCES Alva, A. K. 1997. Best Management Practice for Fertiliz ation of Mature Citrus Trees on the Ridge Area to Minimize Nitr ate Contamination of Groundwater. Lake Alfred, FL: University of Florida, Institute of Food and Agricultural Sciences, Citrus Research and Education Center. Annamalai, P. 2004. Citrus yield mapping syst em using machine vision. Master’s thesis. University of Florida. Gainesville, FL. Association of Analytical Communities ( AOAC). 1995. Protein in animal feed. In Official Methods of Analysis of AOAC International. 16th ed. AOAC 990.03. Arlington, VA: AOAC International. American Society of Agricultural Engineers (ASAE) Standards, 49th ed. 2002. S358.2. Moisture Measurement – Forages. St. Joseph, MI: ASAE. American Society for Testing and Ma terials (ASTM). 1997. E131-95: Standard terminology relating to molecular spectroscopy. In Annual Book of ASTM Standards. West Conshohocken, PA: ASTM. Blackmer, T. M., J. S. Schepers, and G. E. Varvel. 1994. Light reflectance compared with other nitrogen stress measur ements in corn leaves. Agron. J. 86(6): 934-938. Blackmer, T. M., J. S. Schepers, G. E. Varvel, and E. A. Wa lter-Shea. 1996. Nitrogen deficiency detection using reflected s hortwave radiation from irrigated corn canopies. Agron. J. 88(1): 1-5. Bowman, W. D. 1989. The relationship between leaf water status, gas exchange, and spectral reflectance in cotton leaves. Remote Sens. Environ. 30: 249-255. Card, D. H., D. L. Peterson, P. A. Mats on, and J. D. Aber. 1988. Prediction of leaf chemistry by the use of visible and near-infrared reflectance spectroscopy. Remote Sens. Environ. 26(2): 123-147. Carter, G. A. 1991. Primary and secondary eff ects of the water cont ent on the spectral reflectance of leaves. American J. of Botany. 78(7):916-924. Carter, G. A., and A. K. Knapp. 2001. Leaf op tical properties in higher plants: Linking spectral characteristics to stress and chlorophyll concentration. American J. Botany 88(4): 677-684.

PAGE 97

83 Curcio, J. A., and C. C. Petty. 1951. The n ear infrared absorption spectrum of liquid water. J. of the Optical Society of America. 41:302-304. Esbensen, K. H. 2002. Multivariate Data Analysis in Practice. 5th ed. Woodbridge, NJ: CAMO Technologies. Farabee, M. J. 2001. Photosynthesis. Avondale, Ariz.: Estrella Mountain Community College. Available at: www.emc.maricopa.edu/faculty/farab ee/BIOBK/BIOBK/BioBookTOC.html. Accessed 19 December 2003. Florida Agricultural Statistics Serv ice (FASS). 2005. Washington, D.C.: USDA. Available at: www.nass.usda.gov/fl. Accessed 9 May 2003. Frank, I. E., and Friedman, J. H. 1993. A st atistical view of some chemometrics regression tools. Technometrics. 35(2): 109–135. . Gausmann, H. W., W. A. Allen and R. Card enas. 1969. Reflectance of cotton leaves and their structure. Remote Sens. Environ. 1:110-22. Hanlon, E. A., T. A. Obreza, and A. K. Alva. 1995. Tissue and soil analysis. In Nutrition of Florida Citrus Trees. 13-16. Tucker, D. P. H., Alva, A. K., Jackson, L. K., and Wheaton, T.A. eds. Gainesville, FL: University of Florida. Hatchell, D. C. 1999. Technical Guide. 3rd Ed. Analytical Spectral Devices, Inc. Boulder, CO: Analytical Spectral Devices. Helland, I. S. 2001. Some theoretical aspect s of partial least squares regression. Chemometrics and Intelligent Laboratory Systems 58(2): 97-107. Jensen, J. R. 2000. Remote Sensing of the Environment. Upper Saddle River, NJ: Prentice-Hall. Katz, J. J., R. C., Dougherty, and L. J. Boucher. 1966. Infrared and nuclear magnetic resonance spectroscopy of chlorophyll. In The Chlorophylls (L. P. Vernon and G. R. Seely, Eds.). New York: Academic Press. pp. 186–249. Kokaly, R. F. 2001. Investigating a physical basis for spectroscopic estimates of leaf nitrogen concentration. Remote Sens. Environ. 75:153–161. Lee, W. S., and S. W. Searc y. 2000. Multispectral sensor fo r detecting nitrogen in corn plants. ASAE Paper No. 001010. St. Joseph, MI: ASAE. Lee, W. S., S. W. Searcy, and T. Kataoka . 1999. Assessing nitrogen stress in corn varieties of varying co lor. ASAE Paper No. 993034. St. Joseph, MI: ASAE. Lingaerde, O. L., and N. Christophersen. 2000. Shrinkage structure of partial least squares. Scandinavian J. Statistics 27(3): 459-473.

PAGE 98

84 Min, M., and W. S. Lee. 2005. Determination of significant wavele ngths and prediction of nitrogen content for citrus. Trans. ASAE 48(2): 455-461. Min, M., W. S. Lee, and I. Bogrekci. 2004. Th e effect of water and variety on nitrogen sensing of citrus leaf. ASAE Paper No. 041080. St. Joseph, MI: ASAE. Min, M., W. S. Lee, Y. H. Kim, R. A. Bucklin. 2005. Nondestru ctive detection of nitrogen in Chinese cabbage l eaves using VIS-NIR spectroscopy. Hortscience (Acccepted) Morgan, M. and Ess, D. 2003. The Precision-Farming Guide for Agriculturists. 2nd ed. Deere & Company.Moline, IL. Noggle, G. R. and G. J. Fritz. 1976. Water as a plant constituent. In Introductory Plant Physiology, 376-413. Englewood Cliffs, NJ: Prentice-Hall. Norris, K. H., R. F. Barnes, J. E. Moore and J. S. Shenk. 1976. Predicting forage quality by infrared reflectance spectroscopy. J. of Animal Science. 43(4): 889-897. Pollock S. L., B. Lin, and J. Allshouse . 2003. Characteristics of U.S. orange consumption. United States Departme nt of Agriculture, FTS 305-01. SAS. 1990. SAS/STAT User's Guide. Ver. 6, 4th ed. Vol. 1 and 2. Cary, N.C.: SAS Institute, Inc. Sui, R., J. B. Wilkerson, W. E. Hart, a nd D. D. Howard. 1998. Integration of neural networks with a spectral reflectance sensor to detect nitrogen deficiency in cotton. ASAE Paper No. 983104. St. Joseph, MI: ASAE. Sui, R., and J. A. Thomasson. 2004. Plant health sensing system for in situ determination of cotton nitrogen status. ASAE Pa per No. 041081. St. Joseph, MI: ASAE. Sudduth, K. A., and J. W. Hummel. 1993. Port able, near-infrared sp ectrophotometer for rapid soil analysis. Trans. ASAE 36(1): 185-193. Sundberg, R. 1999. Multivariate calibration: Di rect and indirect re gression methodology. Scandinavian J. Statistics 26(2): 161-207. 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 corre cting in-season fertilizer nitrogen deficiencies in winter wheat. Trans. ASAE 39(5): 1623-1631. Thomas, J. R., L. N. Namken, G. F. Oert her, and R. G. Brown. 1971. Estimation leaf water content by reflectance measurements. Agron. J. 63:845-847. Thomas, J. R., and G. F. Oerther. 1972. Es timating nitrogen content of sweet pepper leaves by reflectance measurements. Agron. J. 64(1): 11-13.

PAGE 99

85 Tucker, C. J. 1980. Remote sensing of leaf water content in the near infrared. Remote Sens. Environ. 10:23-32. Tucker, D. P. H., A. K. Alva, L. K. Jackson, and T. A. Wheaton. 1995. Nutrition of Florida Citrus Ttrees. Institute of Food and Agricultu ral Science. Gainesville, FL University of Florida. Tumbo, S. D., D. G. Wagner, and P. H. Heinemann. 2002. Hyperspectral-based neural network for predicting chlorophyll status in corn. Trans. ASAE 45(3): 825-832. Walburg, G., M. E. Bauer, C. S. T. Daughtry, and T. L. Housley. 1982. Effects of nitrogen on the growth, yield, and re flectance characteristics of corn. Agron. J. 74(4): 677-683. Williams, P. and K. Norris. 2001. Near-Infrared Technology in the Agricultural and Food Industries. 2nd ed. St. Paul, Minn.: American Association of Cereal Chemists. Wold, S., H. Martens, and H. Wold. 1983. The multivariate calibration method in chemistry solved by the PLS method. In Proc. Conference in Matrix Pencils, 286293. A. Ruhe and B. Kagstrom, eds. He idelberg, Germany: Springer-Verlag. Woolley, J. T. 1971. Reflectance and tr ansmittance of light by leaves. Plant Physiology. 47: 656-662. Yoder, B. J., and R. E. Pettigrew-Cr osby. 1995. Predicting nitrogen and chlorophyll content and concentrations from reflectan ce spectra (400-2500 nm) at leaf and canopy scales. Remote Sens. Environ. 53(3): 199-211.

PAGE 100

86 BIOGRAPHICAL SKETCH The author was born in 1976 in Anhui, Chin a. She graduated with a Bachelor of Engineering degree in mechatronics in July 1997 from the Anhui Agri cultural University, Anhui, China. She obtained a Master of Engi neering degree in mechanical engineering from Jiangsu University of Science and Technology, Zhenjiang, China. Then she earned her Doctor of Philosophy degree from the Agricultural and Biological Engineering Department at the University of Florida in May 2006.