<%BANNER%>

Record for a UF thesis. Title & abstract won't display until thesis is accessible after 2011-08-31.

Permanent Link: http://ufdc.ufl.edu/UFE0025034/00001

Material Information

Title: Record for a UF thesis. Title & abstract won't display until thesis is accessible after 2011-08-31.
Physical Description: Book
Language: english
Creator: Ruslan, Rashidah
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: Agricultural and Biological Engineering -- Dissertations, Academic -- UF
Genre: Agricultural and Biological Engineering thesis, M.E.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Statement of Responsibility: by Rashidah Ruslan.
Thesis: Thesis (M.E.)--University of Florida, 2009.
Local: Adviser: Ehsani, M Reza.
Local: Co-adviser: Reyes De Corcuera, Jose Ignacio.
Electronic Access: INACCESSIBLE UNTIL 2011-08-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0025034:00001

Permanent Link: http://ufdc.ufl.edu/UFE0025034/00001

Material Information

Title: Record for a UF thesis. Title & abstract won't display until thesis is accessible after 2011-08-31.
Physical Description: Book
Language: english
Creator: Ruslan, Rashidah
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: Agricultural and Biological Engineering -- Dissertations, Academic -- UF
Genre: Agricultural and Biological Engineering thesis, M.E.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Statement of Responsibility: by Rashidah Ruslan.
Thesis: Thesis (M.E.)--University of Florida, 2009.
Local: Adviser: Ehsani, M Reza.
Local: Co-adviser: Reyes De Corcuera, Jose Ignacio.
Electronic Access: INACCESSIBLE UNTIL 2011-08-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0025034:00001


This item has the following downloads:


Full Text

PAGE 1

1 QUANTIFICATION OF TOTAL SOLUBLE SOLIDS CONTENT AND TITRATABLE ACIDITY TO ASSESS CITRUS MATURITY USING A MID IR SPECTROMETER AND A VIS NIR SPECTRORADIOMETER By RASHIDAH RUSLAN A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR TH E DEGREE OF MASTER OF ENGINEERING UNIVERSITY OF FLORIDA 2009

PAGE 2

2 2009 Rashidah Ruslan

PAGE 3

3 To my family who love and support me

PAGE 4

4 ACKNOWLEDGMENTS My deepest appreciation goes to my parents and family for their love, prayers, and constant encouragement, not only during these two years, but also throughout my life. Without them, all this hard work would not be possible.I am also deeply indebted to my advisor, Dr. Reza Ehsani, for accepting me into this program, and guiding me throughout my m aster studies. To my co advisor Dr. Jose Reyes, thank you for the valuable discussion and f or providing chemical equipment needed for the experimen ts. To my committee member, Dr. Won Suk Lee, thank you for sharing valuable knowledge on NIR spectroscopy. I would also like to acknowledge my lab mates, for providing a conducive and fun environment in the lab. To all my friends in the United Sta tes, thank you very much for your encouragement through good and bad times. To my old friend, Nurulhuda Khairu din, thank you for your support ideas, and encour agements in pursuing our m aster as an international student together. Last but not least, a spec ial thank to my fiance, Hafiz Hasri who has always stood by my side, s upporting me to accomplish my m aster degree.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ ............... 4 LIST OF TABLES ................................ ................................ ................................ ........................... 7 LIST OF FIGURES ................................ ................................ ................................ ......................... 9 ABSTRACT ................................ ................................ ................................ ................................ ... 12 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .................. 14 Introduction ................................ ................................ ................................ ............................. 14 Statement of Problem ................................ ................................ ................................ ............. 15 Objectives ................................ ................................ ................................ ............................... 16 2 REVIEW OF LITERATURE ................................ ................................ ................................ 19 Orange Composition ................................ ................................ ................................ ............... 19 Spectroscopy ................................ ................................ ................................ ........................... 20 Electromagnetic Spec trum ................................ ................................ ................................ ...... 21 Near Infrared (NIR) Spectroscopy ................................ ................................ .......................... 21 NIR Spectroscopy for Non Invasive Measurement of Fruits. ................................ ......... 22 NIR Spectroscopy for Aqueous Solution and Dry Extract ................................ .............. 24 Aqueous solutions ................................ ................................ ................................ .... 24 Dry extract ................................ ................................ ................................ ................ 26 Mid Infrared (MIR) Spectroscopy ................................ ................................ .......................... 26 Attenuated Total Reflection (ATR) ................................ ................................ ................. 27 Mid Infrared Spectroscopy for Aqueous Solution and Dry Extract. ............................... 28 Aqueous solutions ................................ ................................ ................................ .... 28 Dry extract ................................ ................................ ................................ ................ 28 Spectral Data Correction ................................ ................................ ................................ ......... 29 Derivati ves ................................ ................................ ................................ ....................... 29 Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV) ............ 30 Mean Filter Smoothing ................................ ................................ ................................ .... 30 Mean Centering ................................ ................................ ................................ ............... 30 Chemometric Techni ques ................................ ................................ ................................ ....... 31 Multiple Linear Regressions (MLR) ................................ ................................ ............... 31 Principal Component Regression (PCR) ................................ ................................ ......... 31 Partial Least Squares Regression (PLS) ................................ ................................ .......... 32 3 MATERIALS AND METHOD S ................................ ................................ ........................... 36 Data Collection and Sample Preparation ................................ ................................ ................ 36

PAGE 6

6 Refractometer Measurement and Chemical Titration ................................ ............................. 37 Near Infrared (NIR) Spectroradiometer ................................ ................................ .................. 37 Mid Infrared (MIR) Spectrometer ................................ ................................ .......................... 39 Chemometric ................................ ................................ ................................ ........................... 40 Model Evaluation ................................ ................................ ................................ .................... 41 4 RESULTS AND DISCUSSI ON FOR THE MIR RANGE ................................ .................... 43 Descriptive Statistics for Destructive Analysis ................................ ................................ ...... 43 Spectral Characteristics ................................ ................................ ................................ .......... 45 Developm ent of Mid Infrared Spectrometer Calibration Model ................................ ............ 47 Soluble Solids Content (SSC) ................................ ................................ ......................... 47 Titratable Acidity (TA) ................................ ................................ ................................ .... 55 Discussion ................................ ................................ ................................ ............................... 62 5 RESULTS AND DISCUSSI ON FOR THE NIR RANGE ................................ ..................... 65 Spectral Characteristics ................................ ................................ ................................ .......... 65 Developm ent of a Calibration Model for the NIR Spectral Range ................................ ........ 68 Soluble Solids Contents (SSC) ................................ ................................ ........................ 68 Titratable Acidity (TA) ................................ ................................ ................................ .... 76 Discussion ................................ ................................ ................................ ............................... 83 6 INSTRUMENT COMPARISO N AND CONCLUSION ................................ ....................... 85 Instrument Comparison ................................ ................................ ................................ .......... 85 Performanc e Comparison ................................ ................................ ................................ 85 Ease of Operational Use ................................ ................................ ................................ .. 8 7 Conclusion ................................ ................................ ................................ .............................. 88 APPENDIX A RESULTS FOR MIR MODE LLING ................................ ................................ ..................... 90 B RESULTS FOR NIR MODE LLING ................................ ................................ ...................... 92 LIST OF REFERENCES ................................ ................................ ................................ ............... 98 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 103

PAGE 7

7 LIST OF TABLES Table page 2 1 Summary of the development of calibration models from non destructive NIR spectroscop y for different citrus varieties by different spectral correction combinations. ................................ ................................ ................................ ..................... 34 2 2 Summary of the development of calibration models from NIR spectroscopy for different aqueous solutions. ................................ ................................ ............................... 35 4 1 Mean and Standard Deviation of SSC and TA for three citrus varieties. .......................... 44 4 2 Summary of the best SSC models developed in the MIR range by PLS and PCR calibration techniques. ................................ ................................ ................................ ....... 49 4 3 Summary of the best TA models developed in the MIR range by PLS and PCR calibration techniques. ................................ ................................ ................................ ....... 56 5 1 Summary of the SSC models d eveloped from PLS and PCR calibration techniques. ...... 70 5 2 Summary of the TA models developed from PLS and PCR calibration technique s. ........ 78 6 1 Summary of the best calibration model for each of the wavelength range. ....................... 86 A 1 Calibration and prediction of SSC and TA with different spectral analyzing pretreatment by PLS technique for MIR spectroscopy. ................................ ..................... 90 A 2 Calibration and prediction of SSC and TA with different spectral analyzing pretreatment by PCR technique for MIR spectroscopy ................................ ..................... 91 B 1 Description of the models developed for NIR analysis. ................................ .................... 92 B 2 Results of PLS calibration and prediction model for Model 1 with various combination of spectral correction in predicting soluble solids content (SSC) and titra table acidity (TA). ................................ ................................ ................................ ....... 92 B 3 Results of PCR calibration and prediction model for Model 1 with various combination of spectral correction in predicting soluble solids content (SSC) and titra table acidity (TA). ................................ ................................ ................................ ....... 93 B 4 Results of PLS calibration and prediction model for Model 2 with various combination of spectral correction in predicting soluble solid content (SSC) and titratable acidity (TA). ................................ ................................ ................................ ....... 93 B 5 Results of PCR calibration and prediction model for Model 2 with various combination of spectral correction in predicting solub le solid content (SSC) and titratable acidity (TA). ................................ ................................ ................................ ....... 94

PAGE 8

8 B 6 Results of PLS calibration and prediction model for Model 3 with various combination of spectral correction in predicting solub le solid content (SSC) and titratable acidity (TA). ................................ ................................ ................................ ....... 94 B 7 Results of PCR calibration and prediction model for Model 3 with various combination of spectral correction in predicting solub le solid content (SSC) and titratable acidity ( TA). ................................ ................................ ................................ ....... 95 B 8 Results of PLS calibration and prediction model for Model 4 with various combination of spectral correction in predicting solub le solid content (SSC) and titratable acidity (TA). ................................ ................................ ................................ ....... 95 B 9 Results of PCR calibration and prediction model for Mo del 4 with various combination of spectral correction in predicting solub le solid content (SSC) and titratable acidity (TA). ................................ ................................ ................................ ....... 96 B 1 0 Results of PLS calibration and prediction model for Model 5 with various combination of spectral correction in predicting solub le solid content (SSC) and titratable ac idity (TA). ................................ ................................ ................................ ....... 96 B 11 Results of PCR calibration and prediction model for Model 5 with various combination of spectral correction in predicting solubl e solid content (SSC) and titratable acidity (TA). ................................ ................................ ................................ ....... 97

PAGE 9

9 LIST OF FIGURES Figure page 1 1 United States of citrus pr oduction for year 1998 2008. ................................ .................... 17 1 2 Florida orange standard requirement of soluble solids content(SSC), total acidity (TA) and ratio of SS C/TA for Flori da oranges. ................................ ................................ 17 1 3 Florida grapefruit standard requirement of soluble solids content(SSC), total acidity (TA) and ratio of SSC/T A ................................ ................................ ................................ 18 2 1 Typical o range composition ................................ ................................ .............................. 19 2 2 Optical configuration for Attenuted Total Reflection (ATR) crystal. .............................. 27 3 1 HR 1024 NIR spectroradiometer ................................ ................................ ..................... 38 4 1 Distribution of the mean SSC and TA versus t he harvesting date ................................ .... 44 4 2 MIR spectra for different concentrations of citric acid, total sugar solution and citrus juices. ................................ ................................ ................................ ................................ 46 4 3 MIR spectra of Hamlin orange, Valencia orange and Thompson Red grapefruit juices. ................................ ................................ ................................ ................................ 46 4 4 SSC model derived from the Hamlin variety developed by PLS ................................ .... 50 4 5 SSC model derived from the Hamlin variety developed by PCR ................................ .... 50 4 6 SSC model derived from the Valencia variety developed b y PLS ................................ .. 51 4 7 SSC model derived from the Valencia variety developed by PCR ................................ 51 4 8 SSC model derived from the Thompson Red grapefruit variety developed by PLS ....... 52 4 9 SSC model derived from the Thompson Red grapefruit variety developed by PCR ...... 52 4 10 SSC model derived from the combination all citrus variety developed by PLS ............. 53 4 11 SSC model derived from the combination all citrus varieties developed by PCR .......... 53 4 12 Plot of RMSEC against number of factors for the SSC model derived from the combination of all citrus varieties. ................................ ................................ ..................... 54 4 13 SSC regression c oefficient derived for combination of all varieties model developed from the PLS and PCR methods. ................................ ................................ ....................... 54 4 14 TA model derived from the Hamlin variety developed by PLS ................................ ...... 57

PAGE 10

10 4 15 TA model derived from the Hamlin variety developed by PCR ................................ ..... 57 4 16 TA model derived from the Valencia variety developed by PLS ................................ .... 58 4 17 TA model derived from the Valencia variety developed by PCR ................................ ... 58 4 18 TA model derived from the Thompson Red grapefruit variety developed by PLS ......... 59 4 19 TA model derived from the Thompson Red grapefruit variety developed by PCR ........ 59 4 20 TA model derived from the combination all citrus varieties developed by PLS ............ 60 4 21 TA m odel derived from combination all citrus varieties developed by PCR .................. 60 4 22 Plot of RMSEC against number of factors for the TA model derived from the combination of all citrus varieties. ................................ ................................ ..................... 61 4 23 TA regression coefficient derived for combination of all varieties model developed from the PLS and PCR methods. ................................ ................................ ....................... 61 5 1 NIR spec trum of citric acid and total sugar solution. ................................ ........................ 66 5 2 Typical relative transmittance spectra for each fruit variety. ................................ ............ 67 5 3 Typical relative absorbance spectra for each fruit variety. ................................ ................ 67 5 4 Model 1 of SSC calibration developed by PLS ................................ ................................ 71 5 5 Model 1 of SSC calibration developed by PCR ................................ ................................ 71 5 6 Model 2 of SSC calibration developed by PLS ................................ ................................ 72 5 7 Model 2 of SSC calibration developed by PCR ................................ ................................ 72 5 8 Model 3 of SSC calibration developed by PLS ................................ ................................ 73 5 9 Model 3 of SSC calibration developed by PCR ................................ ................................ 73 5 10 Model 4 of SSC calibration develop ed by PLS ................................ ................................ 74 5 11 Model 4 of SSC calibrati on developed by PCR ................................ ................................ 74 5 12 Model 5 of SSC calibration developed by PLS ................................ ................................ 75 5 13 Model 5 of SSC calibration dev eloped by PCR ................................ ................................ 75 5 14 Model 1 of TA calibr ation developed by PLS ................................ ................................ .. 78 5 15 Model 1 of TA calibration developed by PCR ................................ ................................ 79

PAGE 11

11 5 16 Model 2 of TA calibration developed by PLS ................................ ................................ .. 79 5 17 Model 2 of TA calibration developed by PCR ................................ ................................ 80 5 18 Model 3 of TA calibration developed by PLS ................................ ................................ 80 5 19 Model 3 of TA calibration developed by PCR ................................ ............................... 81 5 20 Model 4 of TA calibration developed by PLS ................................ ................................ 81 5 21 Model 4 of TA calibration developed by PCR ................................ ............................... 82 5 22 Model 5 of TA calibration developed by PLS ................................ ................................ 82 5 23 Model 5 of TA calibration developed by PCR ................................ ............................... 83

PAGE 12

12 Abstract of Thesi s Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for th e Degree of Master of Engineering QUANTIFICATION OF TOTAL SOLUBLE SOLIDS CONTENTS AND TITRATABLE ACIDITY TO ASSESS CITRUS MATURITY USING A MID IR SPECTROMETER AN D A VIS NIR SPECTRORADIOMETER By Rashidah Ruslan August 2009 Chair: Dr Reza Ehsani Co chair: Dr Jose Reyes Major: Agricultural and Biological Engineering Soluble solids content (SSC) and titratable acidity (TA) are the two major components in the citrus industry that h ave a direct impact on the crop value and the juice quality. It is crucial f or growers to know the stage of maturity i n field before harvesting and for the juice processor to know the fruit quality received at the processing plant. Th is study presents an application of a commercially available VIS NIR s pectroradiometer and a Mid IR (MIR) spectrometer for esti mating SSC and TA of Hamlin and Valencia orange and Thompson Red grapefruit The VIS NIR spectra were acqu ired using a portable spectro meter within the wavelength range of 350 to 2 500 nm in transmission mode. The MIR spectra were obtaine d using a Variable Filter Array ( VFA ) IR spectro meter within the wavelength range of 5.2 to 10.8 m through an attenuated total reflection (ATR) crystal The absorbance spectra were related to the actual value s of SSC and TA by comparison to analyses with bench top refractrometer and by chemical titration, respectively Calibration models relating the NIR and MI R spectra to SSC and TA were developed based on tw o different regression analyses: principal component regression (PCR) and partial least square s (PLS) regression. Selected combinations of spectral preprocessing such as derivatives, mean centering, multiplicative signal corre ction, standard normal variate and

PAGE 13

13 detrending were applied to evaluate the performance of the calibration models P erformance of mo dels in determining the SSC and TA in the NIR and MIR range s were compared. The best calibration model s for the MIR range for the SSC and TA were achieved by PLS regression with coefficient s of determination (R 2 ) of 0.96 and 0.99 and root mean square error of prediction (RMSEP) of 0.20 Brix and 0 .05%, respectively. The best R 2 and RMSEPs from the NIR range were 0.91 and 0.73, and 0.47 Brix and 0.25% for SSC and TA, respectively. Experiment al result s suggest that the MIR spectrometer is a reliable, accurate and fast method for determination of SSC and TA, for future application s

PAGE 14

14 CHAPTER 1 INTRODUCTION Introduction During 2007 08 season, United States produced about 13 million tons of citrus fruit which values for $3.22 billi on (Figure 1 1) Florida produced 75% of total U.S citrus product ion, while California totaled 22 %, and Texas and Arizona produced the remaining three percent. For 2007 08 season, about 1 70 million boxes (National Agricultural Statistics Service, 2008 ) Approximately 95% of the orange crop harvested in Florida is processed into juice or juice espect to sweetness and juice yield This result s from the combination of climatic conditions, tree variety and soil conditions. Because of the climate condition and variety Florida oranges look less appealing for the fresh fruit market. The skin color is not uniform and the peel is fairly hard to remov e which result for consumer r ejection In Florida orange and grapefruit must meet the minimum standard requirement from the U nited S tates D epartment of A griculture (USDA) for percent titratable acidity (TA) and soluble solids content (SSC) before being processed and sold. Figure 1 2 and 1 3 show the standard requirement of the SSC and TA for orange and grapefruit respectively Also, growers are paid based on the weight of the SSC, not by the total weight of the fruit. This is t o assure the grower the va lue paid for their fruits and the processor obtains the desired quality of the finished product For that, a random sample of fruit is taken from each load after weigh in during unloading and taken to the state lab for quality m easurement In the lab, juic e yi eld, SSC, TA and SSC/TA ratio are determined.

PAGE 15

15 Therefor e there is a need for growers to assess the maturity of their fruit before harve sting. By assessing fruit quality before harvesting, growers could better plan their harvesting dat e This is relevant because c itrus is a non climacteric fruits, meaning that it do es not mature after harvesting. H arvesting too early result s in low SSC and a high acidity, and thus, a low er profit for the growers. Statement of Problem Currently, t here is no instr ument that measure s SSC and TA simultaneously in the field SSC is expressed in Brix and is often measured by hydrometer or refractrometer. A h ydrometer measures the density of the juice while a refractrometer measures the refractive index of the juice to estimate its sugar content Typically refractive index is more popular since it can measure over a large scale, 1 70 Brix. Refractrometer measurement is faster, portable and requires a very small sample compare to hydrometer measurement. However, hydromet er is still being used as a standard method in the juice processing plant. For determination of TA a chemical titration is used as a stan dard method in the lab oratory. TA is expressed as a percentage or w/w (weight to weight) C hemical titration is tediou s and time consuming especially when done manually. There are commercial automatic titrators but they are non portable and expensive. Therefore, it would be very useful to develop a new low cost instrument that could measure the SSC and TA simultaneously Simultaneous measurement of SSC and TA will save time and measurement could be done more rapidly on the field. Further, growers could map their SSC a nd TA variability distribution i n the field by using a Global Positioning System ( GPS ) receiver to pinpoint the place of each measurement A v ariability map of SSC and TA will help growers to plan harvesting date to maximize their profits. Recent advances in infrared technology such as miniaturization and improvement in the sensing elements resulted in better a nd cost effective instru ments. Also, infrared technology is

PAGE 16

16 increasingly being used in different applications including agriculture. Applications of infrared spectroscopy in citrus industry have been studied by several researchers (McGlone et al 2003; Gomez et al 2005; Guthrie et al 2005; Cen et al 2007; Cayuela 2008 ; Zude et al., 2008; Lu et al. 2008) Infrared spectroscopy shows a promising future in quantitative method of prediction chemical composition in organic material. In this research, the potential of Near Infrared ( NIR ) and Mid Infrared ( MIR ) wavelength range capabilities in quantify ing total solubl e solids content (SSC) and titratable acidity (TA) of Florida citrus, destructi vely has been studied. Two types of spectro meter were used each covered different wavelength region. A portable NIR spectroradiometer (SpectraVista Corp., New York ) with wavelength of 350 to 2 500 nm and a MIR spectrometer (Wilks Enterprise South Norwalk, CT) covering 5.2 to 10.8 m were employed. Objectives T he long term goal of this study is to develop a low cost portable sensor that could measure th e SSC and TA in real time in a citrus orchard to create variability maps for SSC and TA. The specific objective s of this study is To study the potential of NIR sp ectroscopy (300 2,500 nm) and MIR spectroscopy (5 .2 10 .8 m) for simultaneous ly measuring SSC and TA by develop ing a general calibration model for simultaneous measurement of SSC and TA for different varieties of citrus. To compare the model performance developed by Partial Least Squares (PLS) regression and Principal Component regression (PCR). To compare the NIR an d MIR range calibration performance.

PAGE 17

17 Figure 1 1: United States of citrus production for year 1998 2008. ( Source: National Agricultural Statistics Service, United States Department of Agricultural, 2008) Figure 1 2 : Florida orange standard requirement of so luble solids content(SSC), titratable acidity (TA) and ratio of SSC/TA for Florida oranges. (Source: Florida Department of Agricultu re and Consumer Service, FDACS )

PAGE 18

18 Figure 1 3 : Florida grapefruit standard requirement of so luble solids content(SSC), titratable acidity (TA) and ratio of SSC/TA. (Source: Florida Department of Agriculture and Consumer Service, FDACS)

PAGE 19

19 CHAPTER 2 REVIEW OF LITERATURE Orange Composition Figure 2 1 shows the typical orange composition. The peel of orange consists of flavedo, the waxy, pigmented outer layer and albedo the white fibrous inner layer. The f lavedo contains carotenoid s and oil gla nds. Car atenoid s give the fruit its characteristic skin color and the oil glands provide the aromatic essence of the fruit The a lbe do contains of several substances among others pectin carbohydrate polymers and flavanone glycosides (Braddock, 1999) These substances will influence th e juice quality negatively if they get into the fresh juice. The inner portion contains the juice sacs which are clusters of juice vesicles The juice sacs contain the juice and its soluble components, some cellular organ elles, en zymes and essential oils (Braddock, 1999) T hin segment wall s in the inne r separate the clusters Seeds are also present in the fibrous core region with the oilseed properties of lipids, protein and carbohydrates. Figure 2 1: Typical orange composition ( Reproduced with permission from Reyes De Corcuera, ) About 85% of the citrus fruit is water. The r emaining 15% is composed of soluble suga r s (10%), fiber (2%), organic acid s (1%), amino acids and protein (1%), minerals (0.7%), and oils and lipids (0.3%). The total soluble sugar s in orange juice are composed of sucrose (4%),

PAGE 20

20 glucose (3%) and fructose (3%). The o rganic acid s in citrus juice ar e mainly citric, malic, malonic and oxalic acid. Malic and oxalic acid are mainly present in the peel fraction. The main acid in citrus juice is citric acid. T he total titratable acidity is <1% in orange and mandarin, >1% in grapefruit and >2 3% in lemon and lime (Braddock, 1999). The total soluble solid s (SSC) in citrus juice is determined by the refractive index of sucrose solutions and expressed in terms of degree Brix ( Brix). The titratable acidity (TA) which c onsists of the total acidity in a single strength juice is measured by chemical titration and expressed in percentage. These two main parameters are very important in the determination of citrus juice quality. Spectroscopy Spectro scopy study has been widely used a s a technique in quantit a tive measurement of chemical concentration. Quantitative spectroscopy allows any chemical concentration to be measured based on the absorbance spectrum. The height and peak of the absorbance spectrum is linearly related to the conce ntration of the analyte contained within th e chemical composit ion (Smith, 2002). The r elationship between the absorbance and concentration was explained and de rived by the Beer Lambert law This law states that the amount of light absorbed by a sample, A is equal to the concentra tion of the absorbing analyte in the sample, c the path length of the (2 1) The Beer Lambert Law describes the absorbance in logarithmic form, where I is the intensity of light beamed into t he sample and I 0 is th e intensity of the light coming out of the sample. (2 2)

PAGE 21

21 Absorbance spectra are normally presented i n a graph with the absorbance along Y axis and wavelength along the X axis. Sometimes, spectra are plott ed with Y axis units in transmittance or reflectance. The r elationship of transmittance and absorbance is expressed as follows: A = log (1/T) = log T (2 3) Where T is transmittance, and A is Absorbance. Equation 2 3 shows that transmittance i s not linearly proportional to a bsorbance. T herefore it is necessary to convert the t ransmittance into absorbance for quantitative analysis. T=10 (2 4) Electromagnetic S pectrum In the electromagnetic spectrum, infra r ed (IR) radiation roughly cove rs the r ange from 1m to 100m. With such an extensive range, currently it is impossible to have only one type of instrument (detector) covering the entire region Near infrared (NIR) radiation is normally between 780 to 2 500 nm (Nicolai et al., 2007) whi le mid infrared (MIR) covers the range of 2 ,500 to 25,000 nm. Far i nfrared radiation covers the wavelen gth above 25,000 to 1 000 000 nm. In the section that follows, the focus is exclusively on the application of NIR and MIR regions. Near Infrared (NIR) Spectroscopy NIR spectroscopy has been widely use d for various application s in agriculture The first agricultural application was by Norris (1964) in measuring grain moistur e content (Nicolai et al. 2007). NIR spectroscopy was used for measuring dry matt er, protein content, fat and cellulose of bio products. NIR spectroscopy was also used to quantify soil nitrogen (Ehsani et al., 2001), in pharmaceutical s (Roggo et al., 2007 ), in the beverage s industry (Huang et al., 2008 ) and other application

PAGE 22

22 In this section, the ap plication of NIR spectroscopy to agricultural product s such as fruits and vegetables will be di s cus s ed. The first section describes the non destructive application of NIR sp ectroscopy The second section describes the application of NIR spec troscopy in examining aqueous and dry extract solution s and is particularly focused on sugar and acid determination in fruit juice NIR S pectroscopy for Non I nvasive Measurement of F ruits. For fruits and vegetables, NI R spectros copy applications were i n measuring dry matter, sol uble solids content (SSC), titratable acidity (TA), pH, and firmness of various different fruits such as apple s citrus, melon s peaches and grapes. Most of the work related to fruit quality has been focused on finding the right maturity stages of fruit for consumer benefit s Lammertyn et al. (1998) used a VIS/NIR range of 380 1 560 nm to establish a spectral relationship with the internal quality of the The r elationship s of refle ctance spectra were deriv ed for pH, SSC, stiffness factor and the elastic mo dulus of the apple The best models for pH (R 2 = 0.93), SSC (R 2 = 0.82), stiffness factor (R 2 = 0.90) and elastic modulus (R 2 =0.75) ha ve s tandard error of prediction ( SEP ) of 0.07 0.61, 2.49, and 0.26 respectively. Peirs et al. (2001) predicted the optimum time for harvesting apple s They found that the maturity of apples was dependant on the orchard and data combination s from different orchards for producing a prediction model w ere not acceptable. Temma et al. (2001) predicted apple maturi ty from NIR absorbance spectra by using the SSC of the fruits. A portable low cost sensor named the ound a satisfactory coefficient of determination (R 2 ) at 0.96 and SEP at 0.18 Brix for the relationship between NIR spectra and SSC. The potential of this sensor was further explored on different fruit such as cherries, peaches and tomatoes However, no study was reported on the performance of the sensor with different fruits.

PAGE 23

23 by Partial Least Square s (PLS) regression at spectral range of 300 1 140 nm. The R 2 of the prediction data set f or different fruit positions and models range d between 0.69 and 0.91, and the RMSEP ranged between 7.9 and 15.4% Yan de et al (2007) employed diffuse reflectance FT IR spectroscopy for development of a SSC calibration model. However, the model performa nce was influenced by the probe distance. The highest R 2 and root mean square error of cross validation ( RMSECV ) established for calibrati on model when touching the apple s and the re flectance probe were 0.94 and 0. 74 Brix respectively. Applications of NIR spectroscopy for nondestructive testing of ci trus fruits were summarized in T able 2 1. Among the studies presented, Gomez et al. (2005 ) has the highest R 2 at 0.93 with root mean square error of prediction ( RMSEP ) of 0.32 Brix for an SSC model. The hig hest R 2 ob tained for TA modeling was 0.72 with RMSEP of 0.70 g/L by Lu et al. (2008). D ifferent spectral range s gave a different model performance. McGlone et al. (2003), Guthrie et al. ( 2005a) and Cayuela (2008) agree that no n destructive pr ediction of TA is very difficult because of the low acid concentrations in citrus. Lu et al. (2008) and Gomez et al. (2005) found the PLS calibration technique to be superior to the PCR technique. In non destructive citrus application s the thickness of citrus peel inf luenced the penetrati on of the NIR radiation. Fraser et al. (2003 ) found that most of the energy radiated was absorb ed by the citrus skin. Gut h rie at al (2005b) also found th at the temperature of the fruit affected the bias val ue of the model prediction f or SS C This is beca use of t he effect of H bonding that affects the absorption bands related to OH. Guthrie et al. (2006) found the performance of their model on rockmelon is inferior to its application to citrus. This is due to the irregular skin thickness and heterogeneous distribution of

PAGE 24

24 SSC in rockmelon Abebe (2006) developed a n SSC calibration model with R 2 of 0.82 and RMSEP of 0.42 Brix. Abebe calibration model was s uperior to Guthrie et al. (2006) in terms of the RMSEP value. In a nother approach for predicting SSC of melon, Tsuta et al. (2002) used a cooled charge coupled device (CCD) camera installed with four filter holes with different wavelength s They found that second derivatives absorbance a t 874 and 902 nm gave the highest correlation with SSC. Long and Wa lsh (2005) found that the melon cultivar affect ed the calibration model Th i s proves that NIR spectroscopy is sensitive to the changes of chemical constituent within fruit variety. N on dest ructive measurement of peaches by Ying et al. (2005) result ed in a satisfactory coefficient of determination of 0.92 for SSC and 0.90 for TA, respectively. For grape Larrain et al. (2008) developed NIR sensor and calibration model for sugar, pH and anthoc yianin concentration. Th e instrument was design ed to illuminate the grape on one secto r and collect the light that interacted with the grape from the same side of the fruit. The sensor provide d a good c alibration model with a maximum R 2 of 0.93, 0. 8 0 and 0.68 for sugar, pH and antocyanin, respectively. NIR Spectroscopy for Aqueous Solution and Dry Extract Aqueous solution s Summary of the development of the calibration model for different aqueous solution by NIR spectroscopy was presented in Table 2 2. The m ost popular optical configuration mode used in aqueous spectroscopy is transmission (Lanza and Li, 1984). A path length of 2.2 mm is used for NIR scanning at a spectral range of 1 100 to 2 500 nm for eleven different fruit juices. The best model v a lidatio n result was obtained with orange juice with an R 2 of 0.87 and SEP of 0.25. They conclude d that NIR transmission can potentially be used to predict sugar content

PAGE 25

25 Studies by Giangiacomo and Dull (1986) in the spectral region of 1 550 to 1 850 nm establish ed a satisfactory result on predicting glucose, fructose and sucrose individually in aqueous mixtures of total sugar concentration (10 to 40% concentration). The calibration coefficient of determination ranges from 0.98 to 0.99 and the SEP from 0.35 to 0.8 7 Both studies by Lanza and Li (1984) and Gian giacomo and Dull (1986) employed the Multiple Line ar Regression (MLR) to develop a relationship between selected transmission spectra band with the actual sugar value. Fuj iwara and Honjo (1995) stu died Satsum a mandarin fruit juices and developed a calibration model of sugar content, Brix value and acid content. The model gave a satisfactory SEP of 0.09%, 0.08 Brix and 0.047% (w/v) respectively. The w avelength s 2 272 nm, 1 718 nm and 2 292 nm were thought to be assigned for sugar and acid based on the second derivative spectra of sugar and acid solution. Ou et al. (1997) performed studies similar to Fujiwara and Honjo (1995) to quantify SSC, TA, SSC/TA, sucrose, citric acid and ascorbic acid i n Ponkan mandarin and achieved a good prediction result (R 2 >0.95) by means of transmittance spectroscopy. Chen et al. (2006) studied on raw Japanese apricot fruit juices to determine organic acids using NIR spectroscopy (1 100 1 850 nm). Remarkable peaks at 1 676, 1 720 and 1 784 nm were found for citric acid while malic acid peaks were found at 1 656, 1 694, 1 750 and 1 810 nm Validatio n of the calibration models, le d to a very good result with R 2 and standard error of validation ( SEV ) ranges from 0.96 to 0.98 and 0.22 to 0.27 for malic acid and citric acid, respectively. NIR spectroscopy works v ery well for this fruit juice because the mean concentration of acidity is above 3% and 2% for citric acid and malic acid, respectively.

PAGE 26

26 Cen et al. (2007) used the spectral range of 400 to 1 000 nm to develop a calibration m odel for predicting citric and tartaric acid content of orange juice. Calibr ation models were developed using HPLC as a reference method and VIS/NIR data using a PLS regression. The best R 2 an d RMSEP ac hieved for citric acid was 0.94 and 0.60 and followed by 0.93 and 0.01 for tartaric acid. Dry extract In the analysis of aqueous solution s vibrational water bands often overlap with the analyte of interest. To overcome this obstacle, Meurens et al. (1982, 1990) developed a method of eliminating water in the liquid sample by placing such samples on a fiberglass support until a dry extract is obtained in NIR spectroscopy. The goal wa s to achieve a more sensitive and accurate spectrum. To evaluate this method, Li et al. (1996) used a NIR region of 1 100 to 2 500 nm to quantify separate sugars (glucose, fructose, sucrose) and organic acids (citric and malic acid) in orange ju ice by NIR spectroscopy of the dry extract s The reference value s for individual sugar and organic acids were determined by an enzymatic method. The RMSEP and coefficient of determination between measured and predicted value s obtained from PLS regression w ere 1.57 g/L and 0.92 for glucose, 1.54g/L and 0.93 for fructose, 3.20g/L and 0.96 for sucrose, 0.92 g/L and 0.98 for citric acid and 0.20g/L and 0.90 for malic acid respectively. In this study, PLS and SMLR calibration technique s were compared. A conclu sion was reached that PLS produced a better prediction for individual sugar and acid components in orange juice Mid Infrared (MIR) Spectroscopy The m id infrared (MIR) region is an alternative region for quantitative spectroscopy. The MIR region is very us eful because most of the functional groups absorb infrared radiation regardless of the structure of t he rest of the molecule. In addition, functional groups in the

PAGE 27

27 sample can be detected despite the temperature, pressure, sampling, or changes in molecular structure in other parts of the molecule (Smith, 1999). Attenuated Total Reflection (ATR) The most popular optical conf iguration in MIR spectroscopy i s attenuated total reflection (ATR). ATR is a sampling method that is used for liquid, paste s or powder s t o examine the IR spectrum of the samples. The ATR uses the princip l e of internal total reflection. A beam of infrared light passes through the ATR crystal which is in contact with the sample. The infrared beam is reflected along the crystal until it reach es detect ors at the other end. Figure 2 2 shows the ATR principle in a MIR sensor The reflectance spectrum represents the total summation of the reflectance in contact with the sample. Figure 2 2: Optical configuration for Attenuted Total Reflection (ATR) crystal. (Reproduced with permi ssion from Wilks Enterprise, South Norwalk, CT) In the MIR region, the X axis of the spectrum is often expressed in wavenumber, v 1 ( cm 1 ). This is a unit of frequency rather than wavelength. The relationship between wav enumber, v 1 ( cm 1 ) and wavelength, (nm) is : /v 1 (cm 1 ) (2 5)

PAGE 28

28 Mid Infrared Spectroscopy for Aqueous Solution and Dry Extract. Aqueous solution s The MIR electromag netic spectrum range is where all the fundamental functional groups absorption peak s are found Kems ly et al. (1992) worked o n finding a rapid quantification method for sugar mixtures in the soft drink industry and evaluated the use of an ATR sampling method an Fourier Transform Infrared (FT IR) spectroscopy in the region 4 000 cm 1 to 400 cm 1 (2 500 to 25,000 nm). By means of a P matrix meth od of calibration, they concluded that FT IR spectroscopy provided a potential application for online monitoring and sampling. Mirouze et al. (1993) again used an ATR crystal sampling method and FT IR spectroscopy to determine the glucose syrup to overcome the wet technique of High Performance Liquid Chromatography (HPLC). Satisfactory results were achieved within 3% relative error; which is acceptable by the food industry. Iru dayaraj and Tewari (2003) worked with the infrared region of 6,667 to 10,526 nm (950 to 1500 cm 1 ) through the attenuated total reflection (ATR) sampling method using FT IR spectroscopy. They developed a calibration models to determine sucrose, glucose and fructose in processed apple juice. PLS and PCR techniques produced an R 2 of 0.988 and 0.978 and SEP of 2.2 and 2.5, r espectively during validation. Dry extract Dupuy et al. (1992, 1993) applied the use of an ATR ZeSn crystal for quantitative analys is of powders. P owder impregnation was applied to optimize and improve the optical contact after evaporation. FT Mid IR spectroscopy (2 000 to 700 cm 1 ) of dry extract was employed to overcome the obstacles of high water absorption in the M IR region in determini ng the sug ar and organic acids in orange juice The outcome of their study showed that the dry extract technique are suitable and have a higher sig nal to noise ratio in compared to ATR sampl ing of liquid sample. R esult s obtained from PLS regression gave SEP of 1.63 g/L, 1.19 g/L, 1.27 g/L, 0.21 g/L

PAGE 29

29 and 1.03 g/L for sucrose, glucose, fructose, citric acid and malic acid respectively. SEP of the individual sugar and organic acids from Dupuy et al. (1992) and Li et a l. (19 96) studies show that MIR of dry extracts appear to have a lower SEP than the NIR spectroscopy of dry extract s. B oth of the researchers agree that PLS would pro vide a better calibration model than MLR. Sivakesava and Irudayaraj (2000) improved the model o btained from the FT IR spectroscopy ( 1 530 to 700 cm 1 ) ATR sampling method by experimenting with different techniques of data pretreatment and multivariate calibration. They proposed that spectral selection based on correlation of the absorbance at every w avelength to the concentration of constituents is imp ortant to compensate for deviations from linearity. T his i mprove d the model calibration using a PLS or PCR technique resulting in an R 2 and SEC equal to 0.993 and 3.61% (w/v) with their propose d method for sugar prediction Spectral Data Correction As suggested by many researchers data pretreatment before spectral analysis would help to improve the calibration model performance in terms of R 2 and RMSE P. Some of the pretreatment s that have been appli ed were first and second derivatives, multiplicative scatter correction (MSC), standard normal variate (SNV), mea n filter smoothing, and mean centering Spectral pretreatment selections are not fixed and depend on how the corrections improved the model per formance Derivatives Derivatives are an approach to address problems caused by large baseline variations and overlapping absorption bands in the NIR spectra (Williams and Norris, 1987). The most popular order of derivatives is the second derivatives of a spectrum, since it is able to correct the additive and multiplicative effects such as multiplicative scatter correction ( MSC ) (Nicolai et al, 2007). Derivatives are able to remove overlapping bands because they are related to t he curvature and

PAGE 30

30 have the same sign as the curvature of the spectrum. They are usuall y calculated using the Savitzky Golay algorithm (Naes et al., 2004) or the finite difference method (Williams and Norris, 1987). The Savitzky Golay computation allows the spectra to be smoothed with the order of polynomial and interval width. The latter method is the easiest way to calculate derivatives but it is very sensitive to noise if calculated higher than an order of two. Furthe rmore, it is le ss likely to generate artifacts (Tsai and Philpot, 1998). Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV) In NIR spectra, the reflected energy is a combination of diffuse and specular reflections which are depend ent on the scattering nature and absorption characteristics of the sample (Dhanoa et al. 1994). Additive effects cause vertical shift s whereas multiplicative effects cause a non unity slope s for each spectrum when compared to an ideal or reference spectru m. Dhanoa et al. (1994) explained that MSC and SNV are correction methods to remove the particle size effect. The MSC were used to remove physical effects like particle size and surface blaze, which do not carry any physical or chemical information. Mean Filter S moothing Mean filter smoothing is commonly used to minimize the random noise in spectral data. This method locally smooth es data within a predetermined smoothing window without any polynomial curve fitting or least mean square procedure used in the smoothing computation. This technique uses a local mean computed within the window as the new value of the middle sampling point (Tsai and Philpot, 1998) Mean Centering According to Nicolai et al. (2007), mean centering is a first step for data preproces sing and recommended for all practical applications. Often the operation is to subtract the average from each variable. This method ensures that all the result s are interpretable around the mean.

PAGE 31

31 Chemometric T echniques Chemom etric techniques have played a major role i n spectroscopic sensing since they were developed in the late 1960 s and 1970 s by a number of research groups in chemistry, mainly in analytical and physical organic chemistry (Geladi, 2003). The two main technique groups are classification me thods and regression m ethods. In this review, more focus is given more to regression methods which are capable of correlating the spectrum to quantify the properties of the samples. Multiple Linear Regression s (MLR) Multi ple linear regression s is the olde st method in chemometrics and is less used in application s nowadays because of the improvement in computing power (Roggo et al 2007) and other available chemometric techniques. This technique uses the selected wavelength information that chooses the best correlation coefficient between the spectrum and reference composition value and stop s selecting until some criterion is met. The prediction ability of the search property y j is described as follow s : (2 6) Where b i is the computed coefficient, the absorbance at each chosen wavelength and is the error (Roggo et al, 2007). However, MLR procedures su ffers from two problems; the o verflow of response variables, x by which would be under determin ed and the possibility of co li n earity in response variables, x in which an unstable result of regression model ( Wise et al., PLS Toolbox Manual Ver. 4.0 ). Principal Component Regression (PCR) PCR is a two step procedure created t o address the problems discussed under MLR. The first part of PCR is to treat the data by Principal C omponent Analysis (PCA) and decompose the X variables into numbers of principal components (PC). The second step is to fit a MLR

PAGE 32

32 regression with the optimum numbers of principal components (PC) scores which captur ed the maximum variance of the X variables (Nicolai et al., 2007, Roggo et al., 2007). The pseudoinverse of the response data matrix, X + is estimated as (2 7) Wher e T is the principle components scores and P is the loadings The regression vector, b is determine using a collection of X variables and the property of the interest, y ( Wise et al., PLS Toolbox Manual Ver. 4.0 ) (2 8) Note that in PCR, the PCs rank is uncorrelated and the noise is filtered. The only drawback is i f too many are chosen, the model will over fit. However, the optimum numbe r of PCs is arbitrary and model dependent. For more explanation on PCR calibration technique, read er could refer to Geladi (2003a, 2003b). Partial Least Square s Regression (PLS) PLS was invented to find factors which capture the greatest variance explained in X spectral variables that describe the reference value Y. A l inear relationship is then esta blished between the X and Y. In PLS regression, the variance captured in the X matrices is rank ed accordingly with the Y vector to ensure the latent variables are ordered according to their relevance (Nicolai et al ., 2007; Roggo et al 2007). The pseudoinverse of the response data matrix, X + in PLS is be defined by Equation 2 9: (2 9) Where W is the weight, P is the loadings and T is the scores. The W is required to maintain the orthogonal scores. Giv en a new set of data, X new the scores for these samples are calculated using:

PAGE 33

33 T new = X new W(P T W) 1 (2 10) The prediction of the Y variables for new data is calculated by using Equation 2 11. Y = Xb (2 11) Where b is the regression vector derived from Equation 2 8 and X is the spectral variable ( Wise et al., PLS Toolbox Ver. 4.0 ) For more detail explanati on on PLS regression method, reader could refer to Geladi (2003 a, 2003b ). This PLS approach overcome s both of the problems that occurred with MLR and PCR. This was proven by most of the studies previously described where PLS performance is better than PCR

PAGE 34

34 Table 2 1 : Summary of the development of calibration models from non destructive NIR spectroscopy for different citrus varieties by different spectral correction combination s Note : Correlation coefficient ( R 2 ) and root mean square error of prediction ( RMSEP ) presented are the best R 2 reported by each study with different combination of spectral correction with the present chemometric technique. n/a means not repor ted. Researchers Year Citrus variety Spectral range (nm) Method of optical configuration Spectral correction Calibration technique Parameters R 2 RMSEP Ou et al. 1997 Ponkan mandarin n/a Reflectance n/a n/a SSC 0.76 0.92% Mc Glone et al. 2003 Satsuma mandarin 700 930 Transmission Smoothing Scaling Derivatives PLS SSC TA 0.93 0.65 0.32% 0.15% Miller and Zude 2004 Indian River grapefruit n/a Online monitoring Autoscale PLS SSC 0.67 n/a Guthrie et al. 2005 a Imperial mandarin 720 950 Interactance Derivatives SNV Detrend PLS SSC TA Dry matter 0.75 n/a 0.90 0.14% n/a 0.25% Gomez et al. 2005 Satsuma mandarin 400 2 350 Reflectance Reduced measurement Moving average MSC PLS SSC pH Firmness 0.94 0.80 0.83 0.33 Bx 0.18 8.53 N Lu et al. 2005 Gannan navel orange 800 2,500 Reflectance Derivatives PLS SSC 0.87 0.45 Lu et al. 2008 Gannan navel orange 500 1000 Transmission Derivatives MSC SNV PLS SSC TA pH 0.85 0.72 0.65 0.462 B x 0.704 g/l 0.12 3 Cayuela 2008 Valencia Late orange 570 1 048 and 1 100 1 850 Reflectance Derivatives MSC SNV PLS SSC TA 0.91 0.67 0.51 0.33

PAGE 35

35 Table 2 2: Summary of the development of calibration models f r om NIR spectroscopy for different aqueous solutions Note : C oefficient of determination (R 2 ) correlation coefficient (r), root mean square error of prediction (RMSEP) and standard error of prediction (SEP) presented are the best reported by each study with the present chemometric technique. n/a means not reported. Researchers Year Aqueous solution Spectral range (nm) Method of optical configuration Spectral correction Calibration technique Parameters R 2 or r RMSEP or SEP Lanza & Li 1984 Orange juice 1,100 2,500 Transmission Second derivatives Multiple linear regression SSC 0.87 0.25% Giangiacomo & Dull 1986 Individual sucrose, fructose and glucose 1,550 1,850 Transmission Second derivatives Multiple linear regression Sucrose Fructose Glucose 0.98 0.99 0.35 0.69% (w/w) Fujiwara & Honjo 1995 Satsuma mandarin juice n/a Transmission Second derivatives Multiple linear regression Sugar content Brix value Acidity n/a 0.09% 0.08 Brix 0.05% (w/v) Ou et al. 1997 Ponkan mandarin juice n/a Transmission n/a n/a SSC TA SSC/TA Sucrose Citric acid Ascorbic acid >0.95 n/a Cen et al. 2006 Orange juice 400 1,000 Reflectance MSC Normalization PLS Citric acid Tartaric acid 0.94 0.93 0.60 0.01 Chen et al. 2006 Japanese apricot 1,100 1,850 Transmittance Second derivatives PLS Citric acid Malic acid 0.98 0.96 0.27% 0.21%

PAGE 36

36 CHAPTER 3 MATERIALS AND METHOD S Data Collection and Sample Preparation Valencia oranges, Hamlin oranges and Thompson Red grapefruit were chosen for this study Valencia and Hamlin are the two most commonly planted varieties of sweet oranges in Florida and constitute about half of the oranges grown in the United States. Six t rees per variety were randomly selected and tagged from the Citrus Research and Education Center groves in Lake Alfred, Florida. Data collection was performed o n fi ve different occasions since October 22, 20 08 until January 13, 2009. Fruit maturity stage f or ea ch variety were observed over the entire data collection to ensure the data obtained would cover a wide ran ge of total soluble solids content (SSC) and the titratable acidity (TA) of the fruit s. Four fruits were harvested f r om each tree randomly and a total of 120 fruits were obtained throughout the entire data collection. Fruits were labeled and stored in a packing house cold room at 40 F before processing. Citrus is a non climacteric fruits, therefore the chemical con tent such as sugar and acid do not change rapidly after harvesting However, young citrus such as gr apefruits are susceptible to chilling injuries Citrus fruits can be stor ed up to three months before any processing take s place. The samples were brought to room temperature prior th e m easurement. Each fruit was cleaned with water and cut into halves before being squeezed with a manual juicer. All the juice sampl es were filtered to remove pulp and stored in an air tight container before being analyzed. A total number of 24 fruits from ea ch harvesting period were prepared and used to collect NIR, MIR spectra and simultaneously reference value s of SSC and TA

PAGE 37

37 Refractometer Measurement and Chemical Titration Samples of filtered juice were taken to the laborato ry refractrometer for the refere nce measurement of the SSC A digital bench top Abbe refractrometer with temperature compensation was used for SSC measurement. Each sample measurements were averaged from three readings. S ingle st rength juice titratable acidity (TA) was determined based o n Florida standard law. TA was determined by chemical titration. A sample of 25 ml of juice was titrated with 0.3125N Sodium Hydroxide (N aOH) the end point of pH 8.2. Several drops of phenolphthalein w ere used as a color indicator for the reaction Percent age of TA was calculated based on the amount of Na OH used over the specified sample volume (Kimball, 1999). (3 1) Near Infrared ( NIR ) S pectroradiometer A field portable NIR spectroradiometer named HR 1024 spectroradiometer (SVC Corp., New York) which cover the Visible (VIS) and NIR wavelength s fr om 350 to 2 500 nm was brought in for this particular study. This pho todiode spectroradiometer uses three diffract ion grating spectrometers with one silicon (350 to 1 100 nm) and two InGaAs (1 100 2 500 nm) diode arrays. The silicon array has 512 discrete detectors (350 to 1,000 nm) and the InGaAs arrays (1,000 to 2,500 nm) each ha ve 256 discrete detectors that provide the capability to read 1024 spectral bands The spectral resolution of the sensor are based on the detectors spectral range where < 3.5nm for a range of 350 to1,000 nm, < 9.5 nm for a range of 1, 000 to 1,850 nm and < 6.5 nm for a range of 1,850 to 2,500 nm. The sensor was setup for an illumination module with cuvette holder. This illumination setting allows for transmission mode of the sensor. The setup allows energy to couple wi th fiber

PAGE 38

38 optic cable. Figure 3 1 shows the sensor setup. A fiber optic probe was used to transmit light from a 10 watt tungsten halogen fiber optic il luminator to cuvette holder. A path length of 1.0cm was used. A h igh intensity illumination setting was used to operate the halogen tungsten bul b. Another fiber optic probe was connected to the spectroradiometer in transmitting residual light coming out from the cuvette. The s ample was h eld in a cuvette holder which has two transparent sides and two translucent sides. The cuvette should be placed in the holder with the transparent sides facing the fiber optic cables. Prior to acquiring data, the illumination module was warmed up to let it settle down Spectroradiometer needs to be warm ed up for at least 15 20 minutes to ensure quality of the data collection. Juice sample s were placed in the cuvette and four to six spectral reading s were taken for each sample. Prior to the sample measurement, a reference reading of an empty cell was taken and stored in the memory. The integration time was set to 15 seconds for each measurement to allow for the best signal to noise ratio for each of the spectrum. Later the spectra were averaged for each of the juice sample. The c uvet te was cleaned and dried aft er each of the sample measurement s to ensure that no resi due was left that would interfere with a new sample. A B Figure 3 1: HR 1024 NIR spectroradiometer A) System configuration for Transmisson mode with illumination setting. B) A cuvette for sample placement with the cuvette holder.

PAGE 39

39 Mid Infrared (MIR) S pectro meter A commercial low cost rugged and mobile MIR spectrometer (VFA IR spectrometer, Wilks Corp.) was used for this study. The MIR spectrometer was specifically designed with no optical path exposed to the air and no moving parts like other spectrom eter such as FTIR. ATR crystal sampling was formerly used as an accessory for FTIR spectrometer where it is an alternative method in obtaining difficult sample that cannot be examined by normal transmission method. ATR crystal sampling unit is suitable for sampling of thick or highly absorbing material such as powder, paste and aqueous sample. The sensor was design ed in light weight and is able to be use d either in a laboratory or receiving dock of any processing facility. Variable Filter Array ( VFA ) IR Sp ectrometer has a standard spectral analyzer range cover ing from 5.2 to 10.8 m. The source and detector array are optically connected through an attenuated total reflectance (ATR) ZnSe crystal. The detector array has 128 pixels with 25 cm 1 spectral resol ution. A multi pixel linear pyroelectric array is coupled with a linear variable filter (LVF) and illuminated by an electronically modulated source. Light emitted from the source reflects down the crystal towards the detec tor array. Sample substances were placed on the surface of the ATR crystal and covered with a clear plate. The s ample on top of the crystal absorbs the light emitted and the det ector array measures the amount of infrared light absorbed by the sample. The LVF separates the absorbed infrared light into different wavelength s in order to provide an infrared spectrum. Figure 2 3 shows the ATR configuration. Figure 3 2 shows the Mid Infrared spectrometer. Deionized water was used as a refere nce spectrum Three to four readings were taken for each sample. The spectra collections for each sample were average prior to calibration. Each reading of spectrum c ollection involved 60 sca ns of the sample equivalent to 60 times averaging of the sample. The s pectra were saved in an absorbance format for f urther analysis.

PAGE 40

40 Figure 3 2 : Infraspec VFA IR Spectrometer Chemometric The software package Matlab PLS Toolbox Ver 4.0 was used for chemometric analysis. The r elationship of SSC and TA with the spectral data was developed using both PLS and PCR techniqu es. Calibration analyses were conducted for each citrus fruit variety and a combination of all the spectra with respect to their spectral range. Each data set was divided into a calibration set ( 75 % of the data) and the remaining 25 % of samples to predicti on data set. A c ontiguous block for cross validation was employed to avoid over f itting by retention of too great number of factors (PC or LV). The m aximum number of factors was set to 20 with different cross validation samples depending on the number of t he training sample. Different spectral pretreatments were applied to the calibration dataset. The effect of pretreatment was tested on the performance of the models developed. For the MIR data set mean centering the data followed by first or second deriv ati ves of Savitzky Golay (15 points filter width with polynomial order of two ) approach es and multiplicative scatter correct ion (MSC) are among the combination of spectral correction s applied for both of the calibration techniques The combination of prep rocessing was tabulated in A ppendix A for MIR modelling Improvement of the calibration model with mean centered and none mean centered data can be clearly seen, although it was very minimal. The use of

PAGE 41

41 derivatives also helped to improve some of the calibr ation models however, MSC worsened the results especially in Valencia and Grapefruit models. Therefore, for the development of combination of the varieties model, MSC pretreatment was dropped and focused are given more on the use of mean centering and deri vatives. For the NIR data set, all the spectra were smoothed with filter width of 15 points by Savitzky Golay method Baseline correction of order one was performed followed by MSC, SNV or detrending the spectra. Later, spectra were transformed to either first or second derivatives of Savitzky Golay (15 points filter width with polynomial order of two ) prior to the calibration process. The combination of the spectra preprocessing modeling was presented in Appendix B. Overall, the use of MSC and SNV for NIR spectra did not differ much in most of the model calibration Dhanoa et al. (1994) found that these two methods can be used interchangeably. By removing the surface blaze and particle size effect in the spectra, MSC help to improve the calibration performance. To enhance the sensitivity of the calibration, the effect of derivatives were observed in both PLS and PCR techniques. Model Evaluation To assess the effective ness of each of the model developed by PLS and PCR calibration The prediction ability of a model is based on the root mean square of calibration (RMSEC), root mean squa re of prediction (RMSEP), root mean square of cross validation (RMSECV) and the coefficient of determination (R 2 ) of calibration and prediction data set A good model is considered to have a low RMSEC, RMSECV and RMSEP with a high R 2 RMSEC V and RMSEP were deter mined by the following equation (Nicolai et al., 2007). (3 2)

PAGE 42

42 Where, is the number of the validated samples, and is the predicted and measured value at the i th observation i n the prediction set, respectively. A small difference between RMSECV and RMSEP is preferable. A large difference indicates too many number of factor s are retain ed in the model which will result in noise modeling. The optimal number of factor s was found to be somewhat arbitrary but the plot of RMSEC/RMSECV against the number of factors provided a guideline. Additional factors were included in the model if the reduction of RMSEC/ RMSECV is more than 2%. Too many factors will result in a greater complexity of the model In the following analysis, the performance of the R 2 is grouped in three different categories. A model with an R 2 lower than 0.5 0 considered worse or bad model, an R 2 greater than 0.5 0 but lower than 0.80 is consider as satisfacto ry and an R 2 greater than 0.80 is very satisfactory or very good.

PAGE 43

43 CHAPTER 4 RESULT S AND DISCUSSION FOR THE MIR RANGE Descriptive Statistic s for Destructive Analysis In general, the distribution of SSC increased and TA decreased ov er the harvest season. T he mean and standard deviation of the fruit varieties at each harvesting date is summarized in Table 4 1. F ruit was harvest from October 22, 2008 until January 13, 2009. A long harvesting period ensure d a wi de range of variability covere d for the calibration model Figure 4 1 a visualizes the distribution of the mean SSC The SSC distribution in Valencia orange s had increased from 9.36 to 12.11 Brix throughout the entire picking date. SSC of Hamlin orange s did not increase rapidly since the fruit has reached their maturity level during the entire harvesting date. T hompson Red grapefruit also showed increments in t he SSC distribution which ranged from 9.64 to 10.48 Brix This result is in agreement with Braddock (1999) where Valenc ia oranges dramatically increasing the sweetness over tartness from October to March. The increment of SSC in grapefruit is slower compared to Valencia oranges as seen in Braddock (1999) report, which is true with the results presented here. Figure 4 1b represents the distrib ution of the mean TA over the entire harvesting date Hamlin variety had a lower acidity range (0.59 to 0.73 %) while Va lencia had a higher acidity range (1.19 to 1.76 %) and Thompson Red grapefruit was in the middle (1 28 to 1.33 %). Valencia orange showed a rapid reduction in TA as the maturity increased over the time. The acid reduction results in high Brix/acid ratio, therefore increase the sweetness sensation in the juice (Braddock, 1999). The standard deviation for SSC within a fruit batch wa s below 0.90 Brix with exception of the last batch of Hamlin orange s (1.0 9 Brix) and the fourth batch of Thompson Red grapefruit (1.13 Brix) The variability of TA wa s low with Hamlin oranges having the lowest

PAGE 44

44 standard deviation of 0.09 % Therefore, it wa s concluded that the fruit maturity is almost consistent throughout the tree canopy a t different stage s of maturity. Table 4 1: Mean and Standard Deviation of SSC and TA for three citrus varieties. Citrus variety Date of harvest No of fruit sample Total Soluble Solids ( Brix) Titratable acidity (%) Mean Std. deviation Mean Std d eviation Hamlin Orange 22 Oct 2008 24 8.817 0.440 0.711 0.125 7 Nov 2008 24 9.365 0.798 0.727 0.090 21 Nov 2008 24 9.452 0.758 0.681 0.096 15 Dec 2008 24 9.743 0.178 0.671 0.094 13 Jan 2009 24 10.020 1.089 0.590 0.110 Valencia Orange 22 Oct 2008 24 9.361 0.345 1.756 0.243 7 Nov 2008 24 10.179 0.322 1.537 0.246 21 Nov 2008 24 10.558 0.511 1.452 0.142 15 Dec 2008 24 11.368 0.280 1.237 0.195 13 Jan 2009 24 12.113 0.529 1.186 0.142 Thompson Red Grapefruit 22 Oct 2008 23 9.638 0.804 1.283 0.178 7 Nov 2008 24 9.715 0.904 1.326 0.167 21 Nov 2008 23 9.800 0.705 1.298 0.165 15 Dec 2008 23 10.475 1.131 1.300 0.174 13 Jan 2009 24 10.209 0.715 1.313 0.124 Figure 4 1 : Distribution of the mean SSC and TA versus the harvesting date a) Mean SSC ( Brix ) b) Mean ( %TA ) 8 9 10 11 12 13 14 22 Oct 08 22 Nov 08 22 Dec 08 a) Picking date Hamlin orange Valencia orange Thompson Red grapefruit 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 22 Oct 08 22 Nov 08 22 Dec 08 % TA b) Picking date

PAGE 45

45 Spectral Characteristic s Each MIR spectrum had 128 data points Spectra were acquired in % refl ectance and later converted to absorbance (log 1/R) to obtain a linear relationship between the const ituent s to be measured and MIR spectra. The spectral range used for this analysis was 5.2 m to 10.8 m ( 933 1942 cm 1 ). Figure 4 2 re present s the MIR s pectra of pure citric acid and total sugar solution overlapped with citrus juice The absorbance of the spectra increase d with increasing total concentrat i on of citric acid and sugar content In citric acid spectrum, the range of 5.7 to 6.0 m ( 1 75 0 1 6 8 0 cm 1 ) were correspond to the carbon oxygen double bond, C=O stretch with a narrow peak at 5.81 m ( 1 722 cm 1 ) The carbon oxygen single b ond, C O was found from 7.7 to 10.0 m (1, 3 00 to 1, 0 00 cm 1 ). A broad peak was assigned for C O stretch ba nd at 8.1 to 8.2 m (1,229 1 223 cm 1 ) Oxygen hydrogen, O H in plane bendin g was observed between 7.2 to 6.9 m ( 1, 440 1 ,395 cm 1 ) (Smith, 1999). For the total sugar spectrum, a broad peak in the fingerprint region of 6.7 to 20 m ( 1 500 500 cm 1 ) was observ ed around 9.53 m ( 1 049 cm 1 ) which correspond s to C O stretch. Figure 4 3 sh ows the M IR spectra of citrus juice for each of the selected variety. C=O stretch peak was found at 5.8 m ( 1722 cm 1 ) which correspond to the acid distribution in the juice. Since the total acidity in citrus juice is ve ry small (around 1%), the peak was less intense which match with the peak found in citric acid solution. Another broad peak was observed around 9.5 to 9.6 m ( 1 043 1 053 cm 1 ) to which corresponds to C O bond in total sugar solution. Therefore it is concluded that the entire MIR spectral region is important for measuring SSC and TA.

PAGE 46

46 Figure 4 2 : MIR spectra for different concentrations of citric acid, total sugar solution an d citrus juices. Figure 4 3 : MIR spectra of Hamlin orange, Valencia orange and Thompson Red grapefruit juices. 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0 Absorbance Wavelength (m) 20% 18% 16% 10 Brix 12 Brix 13 Brix Valencia Hamlin Grapefruit 9.53 8.1 8.2 5.81 0.0 0.1 0.2 0.3 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0 Absorbance Wavelength (m) Valencia Hamlin Grapefruit 5.8 9.5 9.6

PAGE 47

47 Development of Mid Infrared Spectrometer Calibration Model Partial Least Square s (PLS) Regression and Principal Component Regression (PCR) were employed to find the relationship between MIR spectra with SSC and TA, respectively. Several choices of different preprocessing techniques were evaluated for the effect of the calibrat ion performance such as mean centering and Savitzky Golay derivatives. Results for various combinations of pretreatments effect were tabulated in Appendix B. Most researchers (Gomez et al. (2005) ; Guthrie et al., (2006) ; Lu et al., (2008) ) a re in favor of PLS compared to PCR i n quantif i cation modeling method. This also proved to be true in this analysis where PLS provides better model performance in terms of coefficient of determination (R 2 ) and root mean square error of prediction (RMSEP) than PCR PLS modeling did not i nclud e any latent variables which are not important in the models for describing the variance captured in the X and Y blocks (Gomez et al 2006). As a result, the PLS model will always result in lower number of factors compared to the PCR technique. Soluble Solids C ontent (SSC) Overall, obtained SSC model was very satisfactory in terms of the R 2 and the RMSEP of the prediction set for different pretreatments except for the multiplicative scatter correction (MSC) as tabulated in Appendix A It was observed t hat MSC pre processing yielded lower R 2 and raised error in calibration and prediction for both PLS and PCR techniques. Table 4 2 summarized the best calib ration s and prediction s developed from the PLS and PCR techniques. For Hamlin oranges, the best models obtained from PLS and PCR have a coefficient of determination of 0.92 and 0.94 and RMSEP s of 0.25 Brix and 0.25 Brix respectively. The n umber of factors retained in the models were three for PLS and four for PCR. Figure 4 4 and 4 5 show the graph of calibration and prediction data set.

PAGE 48

48 For Valencia oranges, the best correlation coefficient and RMS EP developed from the PLS were 0.98 and 0.18 Brix respectively. The best model d eveloped by the PCR technique has an R 2 of 0.97 and RMS EP of 0.19 Brix PLS uses five latent variables while PCR uses seven principal components to develop the models. Figur e 4 6 and 4 7 show the graph of calibration and prediction data set. For Thompson Red grapefruit, the best models from the PLS and PCR te chnique s each have an R 2 of 0.94 and with RMSEP s of 0.22 Brix and 0.23 Brix, respectively. T he number of facto rs modeled in PLS was three compared to five in PCR. Figure 4 8 and 4 9 visualize the predicted and measured SSC relationship for the calibratio n and prediction data set. The effect of different preprocessing was also observed where there are changes in RMSEC and RMSEP amo ng all the models. Mean centering the dataset improved the m odel performance mostly for the PCR technique. Nevertheless, the b est models obtained in PLS for Valencia and Thompson Red grapefruit are when the data had been mean centered To further improve the calibration model, all the data were combined to provide a single calibration model of SSC for citrus. In multivariate anal ysis, the more samples added to the calibration set the better resulting model became. In PLS, the best model for calibration and prediction of the data combination was obtained by mean centering the dataset prior to calibration as shown in Table 4 2 The coefficient of determination was 0.96 with an RMSEP of 0.20 Brix RMSEC of 0.22 Brix and RMSECV of 0.2 4 Brix The PLS techniq ue used five latent variables for this model. In PCR analysis, the best model for calibration and prediction was obtained w ithout any pretreatment to the spectra Six principal components were included in the models which result ed in an R 2 of 0.949, RMSEP of 0.24 Brix, RMSEC of 0.2 5 Brix and RMSECV of 0.25 Brix In

PAGE 49

49 conclusion, SSC model derived from PLS analysis provide d a better performance compared to PCR technique. For visua lization of the model, Figure s 4 10 and 4 11 show a graph of calibration and validation data set of the SSC developed by PLS and PCR. The x axis repre sents the actual SSC while the y axis represents the predicted value from MIR spectroscopy. Figure 4 12 shows the plot of RMSEC against the number of fact ors derived from the PLS and PCR calibrations for SSC The RMSEC values for the PLS model decreased significantly compared to the PCR model. Lower numb er of factors is needed to provide the lowest RMSEC in PLS model. Fi gure 4 13 sh ows the regression coefficient, b deri ved for the best SSC model from PLS and PCR calibration techniques. The regression coefficient value for both technique was relatively sta ble across the spectral window, with PCR model follow ing closely with PLS model. Table 4 2: Summary of the best SSC models developed in the MIR range by PLS and PCR calibration techniques. Variety Calibration technique Spectral Pre treatment No of factors Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP All PLS Mc 5 0.96 0.217 0.236 0.96 0.203 PCR None 6 0.95 0.245 0.251 0.95 0.235 Hamlin PLS 1 st der 3 0.92 0.255 0.275 0.91 0.249 PCR Mc 4 0.93 0.257 0.262 0.89 0.253 Valencia PLS Mc 5 0.98 0.127 0.168 0.98 0.177 PCR Mc, 2 nd der 7 0.97 0.184 0.202 0.98 0.185 Grapefruit PLS Mc 3 0.95 0.216 0.240 0.94 0.215 PCR Mc, 1 st der 8 0.95 0.212 0.237 0.93 0.225 Note: Mc = mean centering. None = no pretreatment was applied prior to calibration. 1 st der, 2 nd der = first and second derivative of Savitzky Golay method.

PAGE 50

50 Figure 4 4 : SSC model derived from the Hamlin variety developed by PLS a)Calibration data set b)Prediction data set Figure 4 5 : SSC model derived from the Hamlin variety developed by PCR a)Calibration data set b)Prediction data set y = 0.9477x + 0.4978 R = 0.92 8 9 10 11 12 13 8 9 10 11 12 13 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.9489x + 0.5609 R = 0.92 8 9 10 11 12 8 9 10 11 12 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.9259x + 0.7085 R = 0.93 8 9 10 11 12 13 8 9 10 11 12 13 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.8919x + 1.1243 R = 0.89 8 9 10 11 12 13 8 9 10 11 12 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 51

51 Figure 4 6 : SSC model derived from the Valencia variety developed by PLS a)Calibration data set b)Prediction data set Figure 4 7 : SSC model derived from the Valencia variety developed by PCR a)Calibration data set b)Prediction data set y = 0.9833x + 0.1813 R = 0.98 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.9001x + 1.0694 R = 0.98 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.9723x + 0.2995 R = 0.97 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.931x + 0.7785 R = 0.98 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 52

52 Figure 4 8 : SSC model derived from the Thompson Red grapefruit variety developed by PLS a)Calibration data set b)Prediction data set Figure 4 9 : SSC model derived from the Thompson Red grapefruit va riety developed by PCR a)Calibration data set b)Prediction data set y = 0.9205x + 0.7904 R = 0.95 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.9352x + 0.5639 R = 0.94 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.9462x + 0.5326 R = 0.95 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.9881x + 0.0436 R = 0.94 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 53

53 Figure 4 10 : SSC model derived from the combination all citrus variety developed by PLS a)Calibration data set b)Prediction data set Figure 4 11 : SSC model derived from the combination all citrus varieties developed by PCR a)Calibration data set b)Prediction data set y = 0.9594x + 0.4094 R = 0.96 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.9466x + 0.5712 R = 0.96 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.9648x + 0.3528 R = 0.95 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.9559x + 0.4823 R = 0.95 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 54

54 Figu re 4 12 : Plot of RMSEC against number of factors for the SSC model derived from the combination of all citrus varieties. Figure 4 13 : SSC regression coefficient derived for combination of all varieties model developed from the PLS and PCR methods. 0.15 0.17 0.19 0.21 0.23 0.25 0.27 0.29 0 5 10 15 20 Brix Number of factors RMSEC PLS RMSEC PCR 15.0 10.0 5.0 0.0 5.0 10.0 15.0 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 Regression coefficient, b Wavelength (m) PCR PLS

PAGE 55

55 Titratable A cidity (TA) For TA calibration and prediction mo dels, the results were very satisfactory for both techniques, except for Hamlin orange s variety as shown in Appendix A. This was perhaps because of the lower variability within the Hamlin samples. However, the Haml in orange calibration model was improved significantly with mean centering prior to the analysis. A s umma ry of the best models for TA is shown in Table 4 3. The highest R 2 achieved by Hamlin orange s for bo th PLS and PCR methods are 0.90 and 0. 87 with RMSEP of 0.004% and 0.04 %, respectively. The improvement of t he models by mean centering was also observed in Valencia orange s and Thompson Red grapefruit for PLS modeling. The best model for Valencia orange and Thompson Red grapefruit has a n R 2 of 0.98 and 0.94 with RMSEP of 0. 04% and 0.05 %, respectively developed by PLS modeling Valencia orange uses four latent variables while the Thompson Red grapefruit uses three latent variables. For PCR modeling the best models for Valencia orange s were obtained by mean centering followed by the second derivatives of the Savitzky Golay method to the dataset. An R 2 of 0.98 was obtained with an RMSEP of 0. 19 %. Five principal components were used for the development of the model. For Thompson Red grapefruit, the best model was developed with six principal components which led to R 2 of 0.93 and an RMSEP of 0.05 % Figure 4 14 to 4 19 presented the estimat ed ver sus measured value of the TA model. To further improve the calibration model, all the data from each variety were combined to provide a single calibration model of TA for citrus The best model for TA was performed by PLS regression with an R 2 of 0.99 and an RMSEP of 0.05 %. The model was achieved by mean centering the data followed by second derivatives from the Savitzky Golay method with four latent variables. The PCR model provide d a slightly lower R 2 of 0.98 with an RMSEP of 0.05 %

PAGE 56

56 for the TA calibr ation and pred iction dataset. PCR uses five principal components to developed the model. The results show that the PLS model is slightly superior to the PCR in terms of the correlation coefficie nt and the number of factor s retained in the model Figure 4 20 and 4 21 show the calibration and validation data set s for the TA developed by the PLS and PCR methods Figure 4 22 shows of the RMSEC against number o f factors derived by PLS and PCR methods for TA calibration. The RMSEC values for the PLS and PCR model s did not differ much. PLS calibration techniques provided a slightly lower RMSEC as compared to the PCR techniques. Figure 4 23 shows the regression coefficient, b derived for the best TA calibration developed by PLS and PCR techniques. The PLS regression coefficient has a higher b value because of the transformation of the spectra into second derivative, however the PCR model was developed without any pretreatment prior to the calibration. Table 4 3: Summary of the best TA models developed in the MIR range by PLS and PCR calibration techniques. Variety Calibration technique Spectral Pre treatment No of factors Calibration Prediction R 2 RMSEC RMSECV R 2 RMSE P All PLS Mc, 2 nd der 4 0.99 0.044 0.045 0.99 0.050 PCR None 5 0.98 0.048 0.049 0.98 0.055 Hamlin PLS Mc 4 0.90 0.034 0.041 0.91 0.004 PCR Mc, 1 st der 5 0.86 0.041 0.045 0.90 0.038 Valencia PLS Mc 4 0.98 0.040 0.050 0.98 0.039 PCR Mc, 1 st der 5 0.98 0.044 0.048 0.98 0.042 Grapefruit PLS Mc 3 0.94 0.038 0.043 0.94 0.045 PCR Mc 6 0.93 0.040 0.043 0.94 0.049 Note: Mc = mean centering. None = no pretreatment was applied prior to calibration. 1 st der, 2 nd der = first and second derivative of the Savitzky Golay method.

PAGE 57

57 Figure 4 14 : TA model derived from the Hamlin variety developed by PLS a)Calibration data set b)Prediction data set Figure 4 15 : TA model derived from the Hamlin variety developed by PCR a) Calibration data set b) Prediction data set y = 0.901x + 0.0665 R = 0.90 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 Predicted TA (%) a) Measured TA (%) y = 0.9x + 0.0573 R = 0.91 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 Predicted TA (%) b) Measured TA (%) y = 0.8618x + 0.0929 R = 0.86 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 Predicted TA (%) a) Measured TA (%) y = 0.9014x + 0.0622 R = 0.90 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 Predicted TA (%) b) Measured TA (%)

PAGE 58

58 Figure 4 16 : TA model derived from the Valencia variety developed by PLS a) Calibration data set b) Prediction data set Figure 4 17 : TA model derived from the Valencia variety developed by PCR a ) Calibration data set b) Prediction data set y = 0.9795x + 0.0289 R = 0.98 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Predicted TA (%) a) Measured TA (%) y = 0.95x + 0.077 R = 0.98 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Predicted TA (%) b) Measured TA (%) y = 0.9752x + 0.035 R = 0.98 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 2.0 Predicted TA (%) a) Measured TA (%) y = 0.9634x + 0.0604 R = 0.98 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 2.0 Predicted TA (%) b) Measured TA (%)

PAGE 59

59 Figure 4 18 : TA model derived from the Thompson Red grapefruit variety developed by PLS a) Calibration data set b)Prediction data set Figure 4 19 : TA model derived from the Thompson Red grapefruit vari ety developed by PCR a) Calibration data set b)Prediction data set y = 0.938x + 0.0817 R = 0.94 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) a) Measured TA (%) y = 0.9563x + 0.0749 R = 0.94 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) b) Measured TA (%) y = 0.9327x + 0.0887 R = 0.93 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) a) Measured TA (%) y = 0.9241x + 0.1205 R = 0.94 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) b) Measured TA (%)

PAGE 60

60 Figure 4 20 : TA model derived from the combination all citrus varieties developed by PLS a)Calibration data set b)Prediction data set. Figure 4 21 : TA model derived from combination all citrus varieties developed by PCR a)Calibration data set b)Prediction data set y = 0.9867x + 0.0153 R = 0.99 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Predicted TA (%) a) Measured TA (%) y = 0.9904x + 0.0142 R = 0.99 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Predicted TA (%) b) Measured TA (%) y = 0.9831x + 0.0194 R = 0.98 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Predicted TA (%) a) Measured TA (%) y = 0.9971x + 0.0012 R = 0.98 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Predicted TA (%) b) Measured TA (%)

PAGE 61

61 Figure 4 22: Plot of RMSEC against number of factors for the TA model derived from the combination of all citrus varieties. Figure 4 23: TA regression coefficient derived for combination of all varieties model developed from the PLS and PCR methods (Note: PLS regression coefficients were scaled down by a factor of 10 from original value) 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0 5 10 15 20 % TA Number of factors RMSEC PLS RMSEC PCR 15.0 10.0 5.0 0.0 5.0 10.0 15.0 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 Regression coefficient, b Wavelength (m) PCR PLS

PAGE 62

62 Discussion One universal algorithm was developed for e ach of the attributes ; SSC and TA of citrus juice In MIR region, the absorption band can be cle arly seen related to the molecules in orange juice as shown in Figure 4 2 Overlapping spectra of orange juice with citric acid and total sugar concentration in Figure 4 2 explain the chemical bonds which exist in the juice. The development of SSC model was very satisfactory. T he highest R 2 was 0.96 with an RMSEP of 0.20 Bri x. The number of factors used in the model was five. This model was developed using the PLS method by combining all the three citrus varieties. However, the R 2 value obtained in this study was lower to Chang (1998) work in predicting SSC from fresh fruit juices. They obtained a R 2 of 0.98 with a SEP of 0.62 with a single wavelength universal calibration through NIR region. In comparison to previous work by Lu et al. (2008), Cayuela (2008), McGlone et al. (2003), the R 2 achieved from ATR sampling with MIR spectroscopy from this study was higher than non destr uctive NIR spectroscopy. Lu et al. (2008) shown an overall R 2 was 0.80 with an RMSEP 0.46 Brix for predicting SSC of Gannan Navel orange. Cayuela (2008) recorded an R 2 of calibration of 0.91 and an RMSEP of 0.51 Brix for quantifying the SSC of Late Valen cia ora nge. McGlone et al. (2003) obtained an R 2 of 0.93 with a RMSEP 0.32 Brix for predicting the SSC of Satsuma mandarin nondestructively. Again, the MIR spectroscopy provided the lowest error of prediction at 0.2 Brix in comparison to the previous work listed above. I t was observe d that mean centering do es not affect much on the model calibration and prediction. T he diff erence between the RMSEP and RMSEC was very small between mean centered and no n mean centered data for the SSC model. For the combi nation model, the RMSEP and RMSEC difference for both spectral preprocessing approaches was almost

PAGE 63

63 negligible These results prove that, user could always use the model derived without any mean centering prior to the calibration for the ease of use and les s complicated programming. The best R 2 and RMSEP ac hieved for modeling TA were 0.99 and 0.05 %, respectively with four latent variables This model was achieved by PLS modeling of the combination al l the three citrus varieties. Many researchers have tried to model the TA attribute in citrus with non destructive NIR spectrocopy led to failure (Mc Glone et al., 2003 (R 2 = 0.65, RMSEP = 0.15%) ; Cayuela, 2008 (R 2 = 0.67, RMSEP = 0.33%) ). Chen et al. (2007) modeled citric and tartaric acid for orange juice with in NIR range with transmission method. They obtained a good R 2 of 0.94 for prediction of citric acid with MSC pretreatment. However, the ir R 2 w as less than the one achieved in this study Gomez et al (2006) also modeled the acidity of mandarin through pH. The best R 2 was 0.87 which is also less than t he result s achieved by this study. From the data in Table 4 2 and 4 3, it is noticeable that the RMSEP for TA is much smaller than for SSC. This is an interesting result because the TA concentratio n present was smaller than the SSC, therefore provide a smaller absor bance peak. According to Smith (2002), this phenomenon was probably because of the linearity. By his observation in predicting ethanol (EtOH) and isopropanol (IPA) concentration with MIR spectroscopy he found that a smaller concentration range (0 to 9%) and absorbance peak of EtoH, PLS could predict with lower SEP at 0.03 volume percent, while for a wider concentration range of IPA (0 to 54%) the SEP was 0.5 volume percent. He concluded t hat there is a possibility in Beer Law that is not well followed in a high er concentration samples. Smith (2002) fin ding is in agreement with the result from this study. TA concentration in citrus juice is very small (mean from 0.68 to 1.8%) while SSC has a higher concentration (mean from 8.82 to 12 Brix).

PAGE 64

64 Results show that this new concept of M IR spectroscopy which employed an ATR crystal as the medium of measurement has a grea t potential in citrus industry with the development of o ne universal calib ration model for both SSC and TA

PAGE 65

65 CHAPTER 5 RESULT S AND DISCUSSION FOR THE NIR RANGE Spectral Characteristic s Figure 5 1 shows the standard spectra of citric acid and total sugar solution obtained within the 820 to 2 000 nm spectral range. A s observed, both of the spectra exhibit similar peaks and valleys. The assignment of ov ertones and other bands occurring in the NIR spectrum is a challenge. The s econd overtone of O H stretch was observed at 960 nm and 955 nm for citric acid and su gar solution respectively. Another second overtone of O H stretch, internally bonded occurs within 1,030 to 1,330 nm with clear peak s at 1,160 nm and 1,220 nm for citric acid and sugar solution, respectively. Both citric acid and sugar solution have a pea k at 1,360 nm which corresponds to O H stretch intramolecular O H bonds wit h single bridge. A broad valley can be seen at ~ 1,620 to 1,700 nm for both spectra which correspond to the first overtone of O H stretch. C=O molecule differentiate between citric a cid and total sugar solution spec tra. Two small valley s can be spotted at 1,460 and 1,510 nm for the third overtone of C =O stretch (William and Norris, 1987). Figure 5 2 shows typical NIR spectrophotometer relative transmittance spectra of three di fferen t citrus varieties for wavelengths from 350 to 2 500 nm. The reference (blank cuvette) spectra were obtained with a low illumination setting while the target spectra were acquired with high illumination In addition to that, a 0.1 mm thick Kodak 50% neutral density filter was used to attenuate the intensity of the light which passed through the blank sample. Relative transmittance spectra were calculated by a ratio of the target to reference value. In Figure 5 2 a ver y low signal to noise ratio was observed in the t ransmittance spectrum above the 1 350 nm wavelength Due to the low signal conditions, the spectral range included in the analysis was reduced from 35 0 to 1 340 nm. The s elected range of the spectra was later

PAGE 66

66 transformed in to absorban ce unit s to provide a linear relationship between analyte concentrations and NIR spectra. Figure 5 3 shows the absorbance spectra for the selected spectral range. The absorption curves of the citrus juice had a peak at ~390 nm which corresponds to the vio let color in the visible regi on. Y ellow and oran ge color wavelength bands occur around 565 to 625 nm. Strong water absorption bands exist from 960 to 990 nm and 1,000 to 1,200 nm. This is true since water is the predominate component to citrus juice ( 80 to 90% ). Figure 5 1: NIR spectrum of citric acid and total sugar solution. 0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.05 1.10 1.15 1.20 800 1,000 1,200 1,400 1,600 1,800 2,000 Transmittance Wavelength (nm) citric acid total sugar 955 960 1,160 1,220 1,360 1,430 1,510 1, 680 1,700

PAGE 67

67 Figure 5 2: Typical relative transmittance spectra for each fruit variety. Figure 5 3: Typical relative absorbance spectra for each fruit variety. 0.00 0.05 0.10 0.15 0.20 0.25 350 850 1350 1850 2350 Transmittance Wavelength (nm) Thompson Red Grapefruit Hamlin Orange Valencia Orange 1.5 2 2.5 3 3.5 4 4.5 5 5.5 350 550 750 950 1150 1350 Absorbance Wavelength (nm) Thompson Red Grapefruit Hamlin Orange Valencia Orange 960 to 990 1,000 to 1,200 390

PAGE 68

68 Deve lopment of a Calibration Model for the N I R Spectral Range In this analysis, several different models and data set s were used to develop a calibration model for both SSC an d TA The VIS NIR spectra measurements were used to determine the corresponding value of the SSC and TA. As stated earlier, PLS and PCR multivariate analysis were used to find the relationship between the NIR spectra and SSC and TA, respectively. Followings are the list of the models derived and the description of each model. This list will be referred to later in the disc ussion. Model 1: Combination of all (Hamlin, Valencia, and Grapefruit) absorbance spectra. Model 2: Combination of Hamlin and Valencia absorbance spectra. Model 3: Thompson Red Grapefruit absorbance spectra. Model 4: Hamlin orange absorbance spectra. Model 5: Valencia orange absorbance spectra. Soluble Solids Contents (SSC) Results of PLS and PCR calibration methods for modeling SSC were tabulated in A ppendix B. A s ummary of the best model derived from both t echniques is shown in Table 5 1 The best PLS result for SSC modeling of Model 1 has an R 2 of 0.63 and an RMSEP of 0.73 Brix with six PLS factors. The optimal pretreatment for Model 1 was MSC spectral correction and followed by first derivative. The bes t PCR technique for Model 1 h as an R 2 of 0.6 0 and an RMSEP of 0.74 Brix w ith MSC pretreatment. This model used seven factors for calibration development Figure 5 4 and 5 5 shows the relationship between the measured and estimated SSC for the calibration and prediction data set. For Model 2, the best models obtained from PLS and PC R have a same R 2 of 0.72 and an RMSEP of 0.65 Brix and 0.63 Brix respectively Although bot h of the models have a close agreement in terms of the RMSEC and RMSECV, PLS method was chosen as the be st. PLS uses only four l atent variables while PCR need seven principal components to achieve the same

PAGE 69

69 model performance. MSC was the optimal pretreatmen t for the developme nt both of the model. Figure 5 6 and 5 7 shows the measured versus estimated value for calibration and prediction data set s. F or Model 3, the best coefficient of determination and RMSEP devel oped from PLS was 0.75 and 0.57 Brix, respect ively. The best model developed by PCR has an R 2 of 0.55 and an RMSEP of 0.79 Brix. Both PLS and PCR uses 11 number of factors to develop the model. The optimal pretreatment for PLS model was MSC and followed by first derivatives, while for PCR, the best preprocessing was MSC and followed by second derivatives. Figure s 5 8 to 5 9 show the calibra tion and validation data set s for the SSC developed by t he PLS and PCR. Model 4 which was developed from the Hamlin orange absorbance spectra achieved the highest R 2 at 0.81 and an RMSEP of 0.59 Brix by PLS calibration technique. The best model developed by PCR has an R 2 of 0.63 and an RMSEP of 0.77 Brix. Bo th of the techniques uses 14 number of factors, with the best spectra preprocessing was MSC and followed by First derivatives. H owever, the PLS calibration technique performed better than the PCR technique For visual ization of the model, Figure 5 10 and 5 11 shows the measured and estimated value from NIR spectra relationship. For Model 5, the best PLS result has an R 2 of 0.91 and an RMSEP of 0.47. PLS uses seven latent variables in predicting the SSC in Model 5. The best pretreatment was detrend and foll owed by first derivatives transformation. The best result from PCR has an R 2 of 0.86 and RMSEP of 0.43 Brix which uses eight principal components. This model was developed by MSC spectral correction. The relationship s of the measured and predicted va lues are presented in Figure 5 12 and 5 13

PAGE 70

70 Overall, the model performance improved from Model 1 to Model 2. This was perhaps because of the different characteristics exhibit ed by T hompson Red grapefruit were removed in Model 2. Thompson Red grapefruit have a slightly less thick juice compared to orange juice. The difference in juice color perhaps contributed to the spectral characteristic s However Model 3 did not perform as expected. The result was satisfactory (R 2 within the range of 0.5 to 0.8) because a small number of calibration samples were used for the model development. For Model 4, the performance was satisfactory however, there is a possibility of over fitting in the calibration data set with the indication in lower R 2 achieved by the prediction da ta set for both of the PLS and PCR technique. The performance of Model 5 was very satisfactory with the highest R 2 achieved was 0.91 and modeled with seven latent variables. Table 5 1: Summary of the SSC models developed from PLS and PCR calibration techn iques. Model Calibration t echnique Spectral P re treatment No of factors Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP 1 PLS 1 st der, MSC 6 0.65 0.645 0.677 0.56 0.728 PCR None, MSC 7 0.62 0.668 0.687 0.54 0.737 2 PLS None, MSC 4 0.74 0.593 0.626 0.65 0.653 PCR None, MSC 7 0.74 0.593 0.617 0.67 0.631 3 PLS 1 st der, MSC 11 0.76 0.420 0.633 0.69 0.572 PCR 2 nd der, MSC 11 0.58 0.561 0.702 0.44 0.791 4 PLS 1 st der, MSC 14 0.81 0.385 0.593 0.60 0.588 PCR 1 st der, MSC 14 0.63 0.539 0.691 0.31 0.772 5 PLS 1 st der, Detrend 7 0.91 0.319 0.385 0.84 0.466 PCR None, MSC 8 0.86 0.396 0.436 0.82 0.432 Note: Mc = mean centering. None = no pretreatment except smoothing and baseline correction. 1 st der, 2 nd der = first and second derivative of Savitzky Golay method

PAGE 71

71 Figure 5 4: Model 1 of SSC calibration developed by PLS a) Calibration data set b) Prediction data set Figure 5 5: Model 1 of SSC calibration developed by PCR a) Calibration data set b) Prediction data set y = 0.6946x + 3.0707 R = 0.65 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.658x + 3.4285 R = 0.56 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.6171x + 3.8576 R = 0.62 8 9 10 11 12 13 14 8 9 10 11 12 13 14 predicted SSC (Brix) a) Measured SSC (Brix) y = 0.5287x + 4.7638 R = 0.54 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 72

72 Figure 5 6: Model 2 of SSC calibration developed by PLS a) Calibration data set B) Prediction data set Figure 5 7: Model 2 of SSC calibration developed by PCR a) Calibration data set b) Prediction data set y = 0.7359x + 2.6843 R = 0.74 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.6295x + 3.7022 R = 0.65 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.7357x + 2.6865 R = 0.74 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.6718x + 3.2532 R = 0.67 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 73

73 Figure 5 8: Model 3 of SSC calibration developed by PLS a) Calibration data set b) Prediction data set Figure 5 9 : Model 3 of SSC calibration developed by PCR a) Calibration data set b) Prediction data set y = 0.7537x + 2.4245 R = 0.76 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.7205x + 2.7402 R = 0.69 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.6535x + 3.4041 R = 0.58 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.5603x + 4.3552 R = 0.44 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 74

74 Figure 5 10 : Model 4 of SSC calibration developed by PLS a ) Calibration data set b) Prediction data set Figure 5 11 : Model 4 of SSC calibration developed by PCR a) Calibration data set b) Prediction data set y = 0.8146x + 1.7599 R = 0.82 8 9 10 11 12 13 8 9 10 11 12 13 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.5799x + 3.9259 R = 0.60 8 9 10 11 12 8 9 10 11 12 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.6277x + 3.5352 R = 0.63 8 9 10 11 12 13 8 9 10 11 12 13 Predicted SSC (Brix) a) Measured SSC (Brix) y = 0.3184x + 6.3831 R = 0.31 8 9 10 11 12 8 9 10 11 12 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 75

75 Figure 5 12: Model 5 of SSC calibration developed by PLS a) Calibration data set b) Prediction da ta set Figure 5 13: Model 5 of SSC calibration developed by PCR a) Calibration data set b) Prediction data set y = 0.9098x + 0.9654 R = 0.91 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 1.1146x 1.162 R = 0.84 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix) y = 0.8607x + 1.4905 R = 0.86 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) a) Measured SSC (Brix) y = 1.0061x 0.0545 R = 0.82 8 9 10 11 12 13 14 8 9 10 11 12 13 14 Predicted SSC (Brix) b) Measured SSC (Brix)

PAGE 76

76 Titratab l e Acidity (TA) Results of PLS and PCR calibration methods for modeling TA were tabulated in A ppendix B. In general, the TA calibration performance was within the range of poor (R 2 < 0.5) to satisfactory (0.5 < R 2 < 0.8) in terms of the R 2 S ummary of the best model derived from both t echniques is shown in Table 5 2 For Model 1, the best result was achieved with the PLS method with 12 factor s retained in the model that le d to an R 2 of 0.72 and an RMSEP of 0.20 %. The optimum data pretreatment was SNV spectral correction which yield ed a low RMSEC (0.21 %) and RM SECV (0.23 %). The best PCR techniques resulted in R 2 of 0.54 with RMSEP of 0.26 %. Fifteen number s o f factors were used for this model with the optimum data pretr eatment was detrending of order one followed by the first derivative of the Savitzky Golay method. The graphs of estimated and predicted value by NIR spec troscopy are shown in Figure 5 14 and 5 15 For Model 2, the best result was achieved by PLS which yield ed an R 2 of 0.74 and an RMSEP of 0.25 %. MSC was the optimum spectral preprocessing step for this data set which dev eloped the lowest RMSEC of 0.22% and RMSECV of 0.24 % compared to the PCR method. The PCR technique achieved an R 2 of 0.68 with RMSEP of 0.26 %, which is slightly higher than that of the PLS Model 2. The optimal preprocessing for PCR technique was SNV. PCR used 17 factors to develop t he model while PLS only need ed seven factors. The relationship s between estimated and predicte d values are shown in Figure 5 16 and 5 17 Model 3 which solely use d the Thompson Red grapefruit data set result ed in a poor TA calibration model. The best R 2 that PLS method could achieve was 0.45 with an RMSEP of 0.11 % with four latent variables retained in the model The optimum spectral preprocessing method was MSC. The PCR method yielded a lower R 2 at 0.43 with an RMSEP of 0.11% and

PAGE 77

77 uses six principal components for developing the model. Figure s 5 18 to 5 19 show the calibr ation and validation data set s for the TA developed by the PLS and PCR technique. For Model 4, which was developed only from Hamlin oranges absorbance spectra yield ed an R 2 of 0.66 and an RMSEP of 0.07% by PLS calibration technique. However this model seems to be over fitted with the prediction data set achieved an R 2 of 0.03 Ten latent variables were used for the development of the model. The best model by PCR cali bration technique achieved an R 2 of 0.52 with the RMSEP of 0.28%. Eight principal components were used which resulted an RMSEC of 0.19% and RMSECV of 0.21%. Figures 5 20 to 5 21 show the calibration and validation data sets for the TA developed by the PLS and PCR technique. For Model 5, the highest R 2 achieved with PLS technique was 0.56 and an RMSEP of 0.18%. PLS calibration technique retained five latent variables for the development of the model. The best R 2 achieved with PCR calibration technique was 0. 54 and an RMSEP of 0.17%. Eight principal components were used for the PCR model. The optimal spectral preprocessing for both of the techniques was MSC, however PLS performance outperformed the PCR in terms of the number of factors used and the R 2 of the c alibration and prediction data set. The graphs of estimated and predicted value by NIR spec troscopy are shown in Figure 5 22 and 5 23 In ge neral, Model 1, 2 4 and 5 were able to satisfactory determine TA which has an R 2 ranges from 0.52 to 0.72 However, the performance of Model 3 was very p oor (R 2 below 0.5). Perhaps low er number of calibration sampl es contributed to the poor performances in Model 3. Model 4 was found to over fit the calibration data set with R 2 obtained for the prediction data set was 0.03. Comparison on the model development for each of the variety between M odel 3, 4 and 5 shows that Model 5 modeled the TA satisfactory with an R 2 of 0.56 and RMSEP of 0.18 %.

PAGE 78

78 Table 5 2: Summary of the TA model s developed from PLS and PCR calibration techniques. Model Calibration technique Spectral P re treatment No of factors Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP 1 PLS None, SNV 12 0.72 0.207 0.234 0.72 0.201 PCR 1 st der, Detrend 15 0.54 0.269 0.284 0.54 0.257 2 PLS 1 st der, SNV 7 0.74 0.218 0.238 0.71 0.249 PCR None, SNV 14 0.68 0.224 0.245 0.69 0.255 3 PLS None, MSC 4 0.45 0.124 0.136 0.50 0.110 PCR None, MSC 6 0.43 0.126 0.136 0.49 0.112 4 PLS 2 nd der, SNV 10 0.66 0.105 0.097 0.03 0.066 PCR 2 nd der, MSC 8 0.52 0.297 0.307 0.65 0.277 5 PLS None, MSC 5 0.56 0.190 0.212 0.63 0.176 PCR None, MSC 8 0.54 0.195 0.212 0.65 0.172 Note: Mc = mean centering. None = no pretreatment except for smoothing and baseline correction. 1 st der, 2 nd der = first and second derivative of Savitzky Golay method Figure 5 14 : Model 1 of TA calibration developed by PLS a) Calibration data set b) Prediction data set y = 0.7247x + 0.3129 R = 0.72 0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) a) Measured TA (%) y = 0.7867x + 0.2289 R = 0.72 0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) b) Measured TA (%)

PAGE 79

7 9 Figure 5 15 : Model 1 of TA calibration developed by PCR a) Calibration data set b) Prediction data set Figure 5 16 : Model 2 of TA calibration developed by PLS a) Calibration data set b) Prediction data set y = 0.5417x + 0.5204 R = 0.54 0.4 0.8 1.2 1.6 2.0 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) a) Measured TA (%) y = 0.6049x + 0.4466 R = 0.54 0.4 0.8 1.2 1.6 2.0 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) b) Measured TA (%) y = 0.7418x + 0.2772 R = 0.74 0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) a) Measured TA (%) y = 0.6211x + 0.424 R = 0.71 0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) b) Measured TA (%)

PAGE 80

80 Figure 5 17 : Model 2 of TA calibration developed by PCR a) Calibration data set b) Prediction data set Figure 5 18 : Model 3 of TA calibration developed by PLS a) Calibration data set b) Prediction data set. y = 0.727x + 0.2932 R = 0.73 0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) a) Measured TA (%) y = 0.622x + 0.4292 R = 0.69 0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4 Predicted TA (%) b) Measured TA (%) y = 0.4463x + 0.7256 R = 0.45 1.0 1.2 1.4 1.6 1.8 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) a) Measured TA (%) y = 0.4975x + 0.6412 R = 0.50 1.0 1.2 1.4 1.6 1.8 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) b) Measured TA (%)

PAGE 81

81 Figure 5 19 : Model 3 of TA calibration developed by PCR a) Calibration data set b) Prediction data set. Figure 5 20 : Model 4 of TA calibration developed by PLS a) Calibration data set b) Prediction data set. y = 0.4308x + 0.7459 R = 0.43 1.0 1.2 1.4 1.6 1.8 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) a) Measured TA (%) y = 0.5017x + 0.6323 R = 0.49 1.0 1.2 1.4 1.6 1.8 1.0 1.2 1.4 1.6 1.8 Predicted TA (%) b) Measured TA (%) y = 0.664x + 0.2296 R = 0.66 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 Predicted TA (%) a) Measured TA (%) y = 0.4451x + 0.3982 R = 0.31 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 Predicted TA (%) b) Measured TA (%)

PAGE 82

82 Figure 5 21 : Model 4 of TA calibration developed by PCR a) Calibration data set b) Prediction data set. Figure 5 22 : Model 5 of TA calibration developed by PLS a) Calibration data s et b) Prediction data set. y = 0.3463x + 0.4462 R = 0.31 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 Predicted TA (%) a) Measured TA (%) y = 0.1492x + 0.8045 R = 0.04 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 Predicted TA (%) b) Measured TA (%) y = 0.5585x + 0.6353 R = 0.56 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 Predicted TA (%) a) Measured TA (%) y = 0.5353x + 0.6156 R = 0.63 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 Predicted TA (%) b) Measured TA (%)

PAGE 83

83 Figure 5 23 : Model 5 of TA calibration developed by PCR a) Calibration data set b) Prediction data set. Discussion PLS calibration technique proved to be superior to the PCR in q uantitative study of SSC and TA. The PLS factors are chosen based on how the factors score correlated to the predicted variable. However in PCR, the factors chosen for the models are solely selected on the amo unt of variation explained in x (spectra wavelength) onl y without the consideration to y relationship. The factor s chosen were not rank ed according to their importance, therefore PCR always ends with higher number of factors included in the models. Overall, the PLS calibrat ion technique is preferable to the PCR technique in developing calibra tion model for SSC and TA. In general, SSC predic tions for all the models were within the range satisfactory (0.5 < R 2 < 0.8) and very satisfactory (R 2 > 0.8 ) The best coefficient of determination a mong the five models ranged from 0.58 to 0.91 whil e the RMSEP ranged from 0.43 Brix to 0.79 Brix. Model 5 which was developed for the Valencia orange was chosen as the best NIR model for SSC calibration with an R 2 of 0.91 and an RMSEP of 0.47 Brix However, the v alue of coefficient of y = 0.5372x + 0.666 R = 0.54 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 Predicted TA (%) a) Measured TA (%) y = 0.5476x + 0.6011 R = 0.65 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 Predicted TA (%) b) Measured TA (%)

PAGE 84

84 determination obtained from this study was low er than those of Kawano et al. (1993) and McGlone et al (2003). Kawano et al. (1993) achieved an R 2 up to 0.98 and a lower RMSEC of 0.28% while McGlone et al. (2003) obtained an R 2 of 0.93 with an RMSEP of 0.32 Brix in developing SSC calibration model of Satsuma mandarin. In comparison to Lu et al. (2008) R 2 value from this study was higher for Valencia orange in predicting the SSC Lu et al. (2008) obtained an R 2 of 0.80 with an RMSEP of 0.46 Brix Cayuela (2008) recorded a sam e R 2 of 0.91 for quantifying SSC of the Late Valencia orange with slightly higher RMSEP of 0.51 For TA prediction, the models developed were within the range of poor (R 2 <0.5) to satisfactory (0.5 < R 2 < 0.8) A maximum of an R 2 of 0.73 and an RMSEP of 0.25% was obtained by Model 2 (combination of Hamlin and Valencia orange absorbance spectra) The best value from this study was superior to the stu dies from Lu et al. (2008), whose highest R 2 for TA prediction was 0.64 for Gannan Navel orange. McGlone et al. (2003) obtained an R 2 of 0.60 with an RMSEP of 0.15% in predicting TA for Satsuma m andarin. Cayuela (2008) achieved an R 2 of 0.56 and an RMSEP of 0.33% in modeling TA for the Late Valencia orange s. Therefore, this study shows that TA calibration is po ssible by mean s of destructive NIR analysis. Spectral range chosen for this study was corresponds with the p revious studies by Lu et al. (2008) They suggested that wavelength range of 500 to 1 000 nm contains a number of carbohydrate and water absorption bands and appropriate for SSC and TA d etermination Chang et al. (1998) found that wavelength 2 270 nm to be a dominant factor in estimatin g Brix values in fruit juices, h owever, due t o the low signal to noise ratio above the wavelength 1 ,350 nm in this studies, wavelength above 1,350 nm was discarded.

PAGE 85

85 CHAPTER 6 INSTRUMENT COMPARISO N AND CONCLUSION Instrument C omparison Traditionally, fruits internal quality i s determined quantitatively by destructive chemical analysis in a laborato ry. This approach is both laborious and time consuming. Recent de velopments in spectroscopy, have allowed researcher to explore new methodology to replace the destructive chemical analysis. In this chapter, comparison o f the NIR spectroradiometer and the M IR spectrometer will be discuss ed based on the calibration performance result s achieved by the two instruments. Performance Comparison According to the analyses in Chapter s 4 and 5, PLS was chosen as the best calibration technique in deriving the SSC and T A model for both of the MIR and NIR wavele ngth range. Table 6 1 summarizes the best calibration model achieved by the two wavelength range s The best calibration and prediction result estimates for SSC and TA by MIR spectrometer have a coefficient s of dete rmination of 0.96 and 0.99, and RMSEP of 0.20 Brix and 0.05%, respecti vely. The best R 2 and RMSEPs from the NIR spectroradiometer were 0.91 and 0.73, and 0.47 Brix and 0.25% for SSC and TA, respectively. NIR range us ed seven latent variables in the development of the SSC and TA model while MIR range used five and four latent variables for SSC and TA, respectively. This indicated that the NIR model was more complex and needed a higher number of factors to describe the va riance captured in the x variables (wavelength). These results indicate that M IR wav elength region outperformed the NIR wavelength region in providing the best calibr ation model for SSC and TA. M IR spectrometer produced models that were low in RMSEP and high in correlation coefficient for both SSC and TA parameters. Additionally, the limit of detection (LOD) and the sensitivity of the calibration were

PAGE 86

86 evaluated The LOD was determined from multiple measurements (n = 10) of the spectral response of deioniz ed water (no concentration) at a specific wavelength using E quation 6 1 : LOD = 3 x Standard Dev. (6 1) The LOD for the MIR range (SSC = 0.004 Brix, TA = 0.003%) appears to be lower than for the NIR range (SSC = 0.04 Brix, TA = 0.32%) for both of t he SSC and TA attributes. These indicate that the MIR range co uld quantify the SSC and TA at very low concentration s The slope of the peak height versus the concentration of the SSC and TA represent the analytical sensitivity reported in Table 6 1. Sensit ivities of 0.023 Brix and 0.027% was achieved by the MIR range for SSC and TA, respectively. NIR range achieved a sensitivity of 0.7 Brix and 1.9% for both of the SSC and TA, respectively. The sensitivities result s indicate that the TA model is more sens itive than the SSC model in both of the wavelength range s The error of prediction for the M IR range was very low as c ompared to NIR range result (Table 6 1) The results proved that using the M IR range, it is possible to c orrelate the absorbance with an unknown concentration of the SSC and TA sample at a very low error. Therefore, this study suggests using Mid IR spectrometer for analyzing citrus juice quality components. Table 6 1: Summary of the best calibration model for each of the wavelength range. Attributes Parameter s NIR range MIR range SSC R 2 0.91 0.96 RMSEP 0.47 Brix 0.20 Brix Number of PLS components 7 5 Limit of detection (LOD) 0.315 Brix 0.00 4 Brix Sensitivity 0.7 00 0.02 3 TA R 2 0.73 0.99 RMSEP 0.25% 0.05% Number of PLS components 7 4 Limit of detection (LOD) 0.043 % 0.003 % Sensitivity 1.9 00 0.027

PAGE 87

87 There are several factors which contributed to the lower performance of the N IR range. Listed are a few sources of error that were believed to contribute to the NIR range performance Sample temperature: Temperature affects the degree of the hydrogen bonding and the hydration status of the constituents (William s and Norris, 1987). The samples were taken from laboratory col d room prior to calibration and al low to cool to room temperat ure prior the measurement with the NIR sensor; h owever it was not iced that the temperature varied from sample to sample based on the observation of the condensation outside the cuvette. Cuvette cell pathlength: The illumination mode provided with the NIR spectroradiometer has a pathlength of 1.0 cm that allows more energy to be absorbed rather than transmitted. According to Settle (1997) a normal t ransmission mode has a pathlength range of 0.1 to 1.0 mm thick. Saturation: The NIR spectroradiometer has a tendency to saturate at a certain wavelength range where the transmission will go flat The saturation was observed during the measuremen t of the actual solution of citric acid and total sugar. However the saturation could be avoid ed with an application of a normal density filter film in front of the light source. The filter could range from 50 to 70 % Ease of Operational Use Ease of operation is very important in the development of a new sensor. A sensor must be both users friendly and robust in order to attract consumers. In order to develop a low cost sen sor with the consumer target being citrus growers, a few factors need to be taken care of. These factors were identified from the design of the two instruments used in thi s project. The lists of the factors are as follows. Easy to use and clean: The e nd user prefer s an instrument that is less complicated with easy operational sampling. ATR sampling is very suitable for a low cost spec trometer where sampling is easy and vers atile for different types of substances. Simpler mechanical design: With no moving parts and no optical p ath exposed to the air less wear and more reliability will be resulted Portable: Portability is important to user if the sensor is to be used outdoor Fast measurement: A slow spectral respons e is time consuming to the user. However, the scanning time of the spectrum is user identified. A longer scanning time, allows having a good spectrum with lower signal to noise ratio Small and l ight weight: This will contribute for the ease of handling for one operator.

PAGE 88

88 Affordable: A s ensor which is less than $5000 would attract buyers. User interface and programming: User friendly s oftware design for the spectrometer allows the user to s witch from the reflectanc e to the absorbance wind ow and vice versa, and also provide the calibration model for the quantification of the attributes to be measured. Spectral region: M IR region was known to identify the key com ponents of a substance and h as long been used by chemist s to identify chemical substances. MI R region is less sensitive than NIR region in chemical changes of the substances to be analyzed. Conclusion The first objective of this study was to evaluate the potential of the NIR spectroradiometer and the MIR spectrometer in measuring citrus total soluble solids content and titratable acidity This objective was achieved by developing a calibration model for each of the citrus varieties. From the analysis, spectral relationships between SSC and TA for each cit rus variety were established according to their respective wavelength range. The second objective of the study was to compare the models developed from the PLS and PCR calibration technique s Results indicated that t he PLS ca libration technique is more fav orable than PCR because of the stability of the models developed. The third objective of this study was to compare the calibration performance from the two spectral ranges, NIR and MIR. This study proved that the MIR range is more suitable in predicting th e SSC and TA by means of the ATR crystal sampling method. Predictive models based on destructive ATR sampling appear to be optimal for estimating citrus fruits internal quality. According to the results obtained from the analysis, the MIR wavelength range outperformed the NIR range in terms of the calibration model p erformance Therefore, MIR spectrome ter is the preferred method for simultaneous measurement of SSC and TA of citrus juices.

PAGE 89

89 With the use of the MI R spectrometer and the models developed for c itrus quali ty, one could ensure a rapid and simultaneous measurem ent of SSC and TA The VFA M IR spectrometer from Wilks Corp. was design to do quantitative analysis provided one knows the calibration model for the substances in interest. However, the model s developed from this study could be improved in the f uture by adding more samples to the calibration data set. To evaluate the model performance independent test ing should be carried out and reported The unknown samples need to be prepared in the same manner as the calibration samples. The spectrum needs to be treated with the same spectral treatment as the calibration model samples. Ideally, the same spectrometer should be used for future sampling and application.

PAGE 90

90 APPENDIX A R ESULTS FOR MIR MODELLING Table A 1 : Calibration and prediction of SSC and TA with different spectral analyzing pretreatment by PLS technique for MIR spectroscopy. Variety Parameter Pretreatment L V Calibration Prediction R 2 RMSEC RMSECV RMSEP R 2 Combination SSC None 5 0.96 0.219 0.237 0.203 0.96 Mc 5 0.96 0.217 0.236 0.203 0.97 Mc, 1 st der 4 0.95 0.230 0.243 0.229 0.96 Mc 2 nd der 4 0.95 0.231 0.244 0.246 0.95 TA None 4 0.99 0.045 0.047 0.053 0.98 Mc 4 0.99 0.044 0.046 0.052 0.96 Mc 1 st der 4 0.99 0.045 0.047 0.049 0.99 Mc 2 nd der 4 0.99 0.044 0.045 0.050 0.99 Hamlin SSC None 3 0.93 0.250 0.277 0.283 0.89 Mc 2 0.92 0.253 0.260 0.287 0.89 1 st der 3 0.92 0.255 0.275 0.249 0.91 2 nd der 2 0.91 0.272 0.287 0.244 0.91 MSC 6 0.84 0.365 0.581 0.616 0.49 TA None 4 0.87 0.039 0.045 0.051 0.84 Mc 4 0.90 0.034 0.041 0.004 0.91 1 st der 4 0.88 0.039 0.042 0.041 0.89 2 nd der 4 0.87 0.040 0.046 0.040 0.91 MSC 5 0.91 0.034 0.043 0.041 0.89 Valencia SSC None 5 0.98 0.139 0.173 0.185 0.98 Mc 5 0.98 0.127 0.168 0.177 0.98 1 st der 4 0.97 0.187 0.213 0.230 0.97 2 nd der 5 0.98 0.157 0.190 0.158 0.98 MSC 3 0.76 0.483 0.543 0.465 0.84 TA None 2 0.94 0.070 0.075 0.072 0.94 Mc 4 0.98 0.040 0.050 0.039 0.98 1 st der 3 0.97 0.047 0.050 0.046 0.98 2 nd der 3 0.97 0.048 0.052 0.043 0.98 MSC 3 0.93 0.073 0.080 0.068 0.95 Grapefruit SSC None 3 0.94 0.218 0.240 0.219 0.95 Mc 3 0.94 0.216 0.240 0.215 0.95 1 st der 2 0.94 0.233 0.251 0.235 0.93 2 nd der 3 0.95 0.207 0.242 0.222 0.94 MSC 5 0.74 0.464 0.663 0.547 0.62 TA None 4 0.94 0.039 0.043 0.045 0.94 Mc 3 0.94 0.038 0.043 0.045 0.94 1 st der 3 0.92 0.044 0.048 0.059 0.90 2 nd der 3 0.93 0.042 0.005 0.056 0.92 MSC 5 0.86 0.064 0.084 0.099 0.77

PAGE 91

91 Table A 2 : Calibration and prediction of SSC and TA with different spectral analyzing pretreatment by PCR technique for MIR spectroscopy Variety Parameter Pretreatment P C Calibration Validation R 2 RMSEC RMSECV RMSEP R 2 Combination SSC None 6 0.95 0.245 0.251 0.235 0.95 Mc 4 0.94 0.257 0.262 0.253 0.94 Mc, 1 st der 5 0.95 0.252 0.258 0.270 0.94 Mc 2 nd der 6 0.95 0.240 0.249 0.255 0.94 TA None 5 0.98 0.048 0.049 0.055 0.98 Mc 4 0.98 0.049 0.050 0.054 0.98 Mc 1 st der 4 0.98 0.053 0.054 0.056 0.98 Mc 2 nd der 4 0.98 0.049 0.050 0.053 0.98 Hamlin SSC None 6 0.92 0.256 0.270 0.280 0.89 Mc 5 0.93 0.249 0.262 0.284 0.89 Mc, 1 st der 5 0.92 0.261 0.272 0.275 0.89 Mc, 2 nd der 7 0.92 0.263 0.285 0.249 0.91 MSC 1 3 0.68 0.517 0.668 0.773 0.23 TA None 4 0.76 0.054 0.058 0.073 0.69 Mc 6 0.82 0.047 0.051 0.052 0.85 Mc, 1 st der 7 0.86 0.041 0.045 0.038 0.90 Mc, 2 nd der 8 0.86 0.041 0.047 0.040 0.91 MSC 6 0.73 0.057 0.061 0.070 0.65 Valencia SSC None 6 0.95 0.219 0.232 0.252 0.96 Mc 6 0.96 0.192 0.211 0.249 0.97 Mc, 1 st der 4 0.96 0.191 0.204 0.199 0.98 Mc, 2 nd der 9 0.97 0.1 63 0. 179 0.18 4 0.98 MSC 5 0.71 0.528 0.564 0.499 0.81 TA None 3 0.94 0.067 0.072 0.073 0.94 Mc 5 0.97 0.048 0.056 0.043 0.98 Mc, 1 st der 5 0.98 0.044 0.048 0.042 0.98 Mc, 2 nd der 6 0.98 0.042 0.048 0.043 0.98 MSC 4 0.90 0.090 0.094 0.080 0.93 Grapefruit SSC None 4 0.94 0.230 0.247 0.225 0.94 Mc 5 0.94 0.226 0.251 0.224 0.94 Mc, 1 st der 8 0.95 0.212 0.237 0.225 0.94 Mc, 2 nd der 6 0.94 0.219 0.243 0.229 0.94 MSC 5 0.09 0.868 0.922 0.782 0.20 TA None 5 0.92 0.043 0.047 0.047 0.94 Mc 6 0.93 0.040 0.043 0. 049 0.94 Mc, 1 st der 6 0.93 0.041 0.045 0.056 0.91 Mc, 2 nd der 6 0.93 0.042 0.047 0.054 0.93 MSC 5 0.55 0.103 0.110 0.139 0.38

PAGE 92

92 APPENDIX B RESULTS FOR NIR MODE LLING Table B 1 : Description of the models developed for NIR analysis. Model Description Model 1 Combination of all (Hamlin, Valencia, and Grapefruit) absorbance spectra. Model 2 Combination of Hamlin and Valencia, absorbance spectra. Model 3 Thompson Red grapefruit absorbance spectra. Model 4 Hamlin orange absorbance spectra Model 5 Valencia orange absorbance spectra Table B 2 : Results of PLS calibration and prediction model for Model 1 with various combination of spectral correction in predicting solubl e solids content (SSC) and titratable acidity (TA). Param eter Spectrum Spectral pretreatment LV Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 5 0.63 0.660 0.679 0.54 0.730 SNV 5 0.54 0.736 0.756 0.39 0.852 Detrend 4 0.54 0.743 0.769 0.47 0.789 1 st der MSC 6 0.65 0.645 0.677 0.56 0.728 SNV 4 0.55 0.747 0.773 0.43 0.852 Detrend 5 0.61 0.677 0.712 0.55 0.731 2 nd der MSC 8 0.62 0.679 0.743 0.46 0.870 SNV 8 0.60 0.700 0.769 0.41 0.896 Detrend 5 0.63 0.659 0.710 0.55 0.734 TA none MSC 12 0.72 0.210 0.238 0.72 0.201 SNV 12 0.72 0.207 0.234 0.72 0.201 Detrend 10 0.65 0.235 0.261 0.65 0.224 1 st der MSC 9 0.68 0.226 0.247 0.68 0.217 SNV 9 0.69 0.221 0.241 0.67 0.218 Detrend 8 0.63 0.240 0.258 0.60 0.242 2 nd der MSC 6 0.57 0.260 0.278 0.63 0.227 SNV 9 0.71 0.213 0.240 0.67 0.214 Detrend 6 0.59 0.253 0.273 0.62 0.232

PAGE 93

93 Table B 3 : Results of PCR calibration and prediction model for Model 1 with various combination of spectral correction in predicting soluble solids content (SSC) and titratable acidity (TA). Parameter Spectrum Spectral pretreatment PC Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC N one MSC 7 0.62 0.668 0.687 0.54 0.737 SNV 8 0.58 0.701 0.725 0.48 0.782 Detrend 6 0.55 0.742 0.769 0.43 0.836 1 st der MSC 6 0.51 0.782 0.798 0.42 0.873 SNV 6 0.46 0.826 0.844 0.35 0.927 Detrend 5 0.59 0.696 0.712 0.56 0.723 2 nd der MSC 4 0.32 0.969 0.981 0.21 1.129 SNV 8 0.48 0.823 0.865 0.33 0.997 Detrend 8 0.58 0.705 0.742 0.50 0.775 TA N one MSC 8 0.39 0.308 0.315 0.41 0.290 SNV 8 0.41 0.303 0.311 0.42 0.286 Detrend 8 0.40 0.305 0.313 0.46 0.275 1 st der MSC 13 0.49 0.281 0.292 0.53 0.259 SNV 13 0.53 0.272 0.284 0.55 0.255 Detrend 15 0.54 0.269 0.284 0.54 0.257 2 nd der MSC 15 0.50 0.279 0.301 0.59 0.241 SNV 15 0.52 0.274 0.296 0.61 0.236 Detrend 13 0.45 0.292 0.310 0.52 0.258 Table B 4 : Results of PLS calibrat ion and prediction model for Model 2 with various combination of spectral correction in predicting soluble solid content (SSC) and titratable acidity (TA). Parameter Spectrum Spectral pretreatment LV Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 4 0.74 0.593 0.626 0.65 0.653 SNV 3 0.67 0.669 0.695 0.55 0.642 Detrend 3 0.70 0.631 0.651 0.58 0.729 1 st der MSC 4 0.69 0.645 0.678 0.59 0.705 SNV 4 0.68 0.656 0.688 0.63 0.675 Detrend 4 0.72 0.606 0.643 0.62 0.692 2 nd der MSC 4 0.62 0.724 0.665 0.54 0.768 SNV 4 0.58 0.760 0.832 0.54 0.777 Detrend 4 0.72 0.724 0.797 0.62 0.677 TA none MSC 5 0.57 0.284 0.296 0.63 0.280 SNV 5 0.57 0.282 0.294 0.63 0.281 Detrend 5 0.56 0.287 0.298 0.60 0.288 1 st der MSC 7 0.74 0.220 0.240 0.71 0.248 SNV 7 0.74 0.218 0.238 0.71 0.249 Detrend 4 0.58 0.280 0.293 0.58 0.299 2 nd der MSC 5 0.66 0.249 0.265 0.70 0.254 SNV 5 0.67 0.247 0.263 0.69 0.255 Detrend 6 0.69 0.240 0.259 0.65 0.271

PAGE 94

94 Table B 5 : Results of PCR calibration and prediction model for Model 2 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA) Param eter Spectrum Spectral pretreatment PC Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 7 0.74 0.593 0.617 0.67 0.631 SNV 2 0.57 0.764 0.774 0.54 0.792 Detrend 4 0.69 0.651 0.665 0.57 0.741 1 st der MSC 2 0.54 0.820 0.828 0.44 0.950 SNV 6 0.66 0.681 0.704 0.60 0.694 Detrend 11 0.73 0.600 0.639 0.66 0.642 2 nd der MSC 3 0.53 0.813 0.830 0.42 0.934 SNV 3 0.45 0.894 0.912 0.40 0.994 Detrend 3 0.74 0.673 0.684 0.50 0.817 TA none MSC 7 0.48 0.312 0.318 0.57 0.303 SNV 7 0.49 0.309 0.314 0.57 0.302 Detrend 7 0.45 0.319 0.326 0.51 0.322 1 st der MSC 6 0.49 0.308 0.314 0.59 0.296 SNV 10 0.65 0.256 0.269 0.60 0.286 Detrend 14 0.73 0.224 0.245 0.69 0.255 2 nd der MSC 8 0.52 0.297 0.307 0.65 0.277 SNV 8 0.51 0.293 0.361 0.65 0.358 Detrend 3 0.31 0.357 0.303 0.40 0.276 Table B 6 : Results of PLS calibration and prediction model for Model 3 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA). Param eter Spectrum Spectral pretreatment LV Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 3 0.44 0.638 0.678 0.53 0.749 SNV 8 0.57 0.563 0.671 0.55 0.712 Detrend 4 0.34 0.714 0.782 0.28 0.887 1 st der MSC 11 0.76 0.420 0.633 0.69 0.572 SNV 5 0.45 0.637 0.725 0.54 0.727 Detrend 4 0.35 0.691 0.727 0.33 0.902 2 nd der MSC 8 0.67 0.494 0.633 0.49 0.733 SNV 6 0.55 0.596 0.725 0.39 0.878 Detrend 3 0.36 0.687 0.727 0.50 0.840 TA none MSC 4 0.45 0.124 0.136 0.50 0.110 SNV 4 0.43 0.127 0.139 0.45 0.116 Detrend 4 0.35 0.135 0.145 0.33 0.133 1 st der MSC 4 0.45 0.124 0.137 0.48 0.113 SNV 4 0.44 0.126 0.140 0.44 0.119 Detrend 3 0.32 0.138 0.150 0.20 0.164 2 nd der MSC 4 0.38 0.137 0.150 0.40 0.130 SNV 7 0.53 0.115 0.145 0.32 0.138 Detrend 5 0.42 0.127 0.144 0.21 0.157

PAGE 95

95 Table B 7 : Results of PCR calibration and prediction model for Model 3 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA). Param eter Spectrum Spectral pretreatment PC Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 3 0.42 0.653 0.679 0.49 0.757 SNV 3 0.35 0.694 0.722 0.44 0.771 Detrend 5 0.34 0.711 0.779 0.18 0.935 1 st der MSC 6 0.47 0.624 0.679 0.55 0.712 SNV 6 0.41 0.658 0.720 0.47 0.742 Detrend 5 0.31 0.716 0.760 0.08 1.068 2 nd der MSC 11 0.47 0.561 0.701 0.44 0.795 SNV 11 0.41 0.584 0.728 0.38 0.874 Detrend 4 0.31 0.704 0.733 0.49 0.844 TA none MSC 6 0.43 0.126 0.136 0.49 0.112 SNV 6 0.42 0.127 0.138 0.43 0.117 Detrend 5 0.32 0.138 0.151 0.31 0.136 1 st der MSC 5 0.41 0.130 0.138 0.46 0.118 SNV 5 0.39 0.133 0.142 0.42 0.125 Detrend 4 0.29 0.142 0.149 0.17 0.172 2 nd der MSC 5 0.30 0.143 0.153 0.38 0.137 SNV 3 0.26 0.153 0.158 0.32 0.149 Detrend 4 0.30 0.140 0.147 0.13 0.181 Table B 8 : Results of PLS calibration and prediction model for Model 4 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA). Param eter Spectrum Spectral pretreatment LV Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 12 0.76 0.431 0.606 0.59 0.601 SNV 12 0.76 0.429 0.594 0.55 0.629 Detrend 13 0.78 0.418 0.646 0.61 0.588 1 st der MSC 14 0.81 0.385 0.593 0.60 0.588 SNV 11 0.77 0.421 0.582 0.53 0.644 Detrend 10 0.72 0.465 0.633 0.52 0.656 2 nd der MSC 9 0.68 0.502 0.785 0.36 0.743 SNV 10 0.72 0.469 0.731 0.30 0.775 Detrend 9 0.68 0.501 0.739 0.46 0.684 TA none MSC 14 0.72 0.059 0.098 0.13 0.106 SNV 14 0.72 0.060 0.101 0.15 0.106 Detrend 14 0.72 0.060 0.100 0.09 0.111 1 st der MSC 13 0.76 0.059 0.097 0.16 0.107 SNV 13 0.71 0.061 0.101 0.17 0.106 Detrend 12 0.72 0.060 0.102 0.13 0.112 2 nd der MSC 10 0.68 0.064 0.102 0.32 0.097 SNV 10 0.66 0.105 0.097 0. 03 0.066 Detrend 11 0.71 0.060 0.109 0.31 0.097

PAGE 96

96 Table B 9 : Results of PCR calibration and prediction model for Model 4 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA). Param eter Spectrum Spectral pretreatment PC Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 13 0.61 0.553 0.689 0.30 0.777 SNV 14 0.61 0.551 0.703 0.29 0.780 Detrend 14 0.60 0.561 0.710 0.31 0.775 1 st der MSC 14 0.63 0.539 0.691 0.31 0.772 SNV 14 0.62 0.542 0.700 0.26 0.800 Detrend 14 0.60 0.558 0.712 0.26 0.801 2 nd der MSC 10 0.39 0.702 0.796 0.01 1.044 SNV 10 0.43 0.677 0.760 0.00 1.056 Detrend 10 0.51 0.617 0.718 0.09 0.916 TA none MSC 18 0.45 0.083 0.120 0.04 0.150 SNV 12 0.36 0.090 0.120 0.13 0.155 Detrend 18 0.46 0.083 0.122 0.05 0.154 1 st der MSC 11 0.37 0.088 0.111 0.09 0.153 SNV 11 0.36 0.090 0.113 0.07 0.150 Detrend 11 0.37 0.089 0.117 0.13 0.163 2 nd der MSC 11 0.31 0.094 0.107 0.04 0.156 SNV 11 0.30 0.094 0.109 0.03 0.152 Detrend 11 0.34 0.092 0.108 0.11 0.162 Table B 10 : Results of PLS calibration and prediction model for Model 5 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA). Param eter Spectrum Spectral pretreatment LV Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 2 0.81 0.459 0.468 0.82 0.427 SNV 6 0.86 0.403 0.444 0.78 0.498 Detrend 6 0.89 0.359 0.417 0.83 0.451 1 st der MSC 3 0.78 0.498 0.528 0.84 0.445 SNV 3 0.72 0.574 0.609 0.80 0.487 Detrend 7 0.91 0.319 0.385 0.84 0.466 2 nd der MSC 6 0.88 0.370 0.433 0.89 0.427 SNV 4 0.73 0.572 0.688 0.84 0.502 Detrend 6 0.88 0.370 0.433 0.89 0.427 TA none MSC 5 0.56 0.190 0.212 0.63 0.176 SNV 7 0.58 0.186 0.214 0.57 0.190 Detrend 6 0.55 0.193 0.216 0.61 0.179 1 st der MSC 5 0.56 0.199 0.216 0.52 0.200 SNV 5 0.56 0.190 0.221 0.49 0.203 Detrend 8 0.62 0.176 0.220 0.56 0.194 2 nd der MSC 6 0.61 0.179 0.220 0.52 0.204 SNV 6 0.58 0.187 0.241 0.46 0.211 Detrend 5 0.58 0.185 0.242 0.58 0.189

PAGE 97

97 Table B 11 : Results of PCR calibration and prediction model for Model 5 with various combination of spectral correction in predicting soluble solid content (SSC) and total acidity (TA). Param eter Spectrum Spectral pretreatment LV Calibration Prediction R 2 RMSEC RMSECV R 2 RMSEP SSC none MSC 8 0.86 0.396 0.436 0.82 0.432 SNV 8 0.82 0.454 0.500 0.79 0.474 Detrend 8 0.83 0.433 0.499 0.84 0.423 1 st der MSC 3 0.76 0.522 0.545 0.83 0.462 SNV 9 0.83 0.443 0.500 0.79 0.468 Detrend 8 0. 83 0. 437 0.4 97 0.8 0 0.4 82 2 nd der MSC 6 0.72 0.584 0.722 0.86 0.485 SNV 6 0.63 0.681 0.800 0.82 0.517 Detrend 9 0.84 0.426 0.483 0.84 0.488 TA none MSC 8 0.54 0.195 0.212 0.65 0.172 SNV 7 0.51 0.200 0.216 0.63 0.172 Detrend 8 0.51 0.201 0.218 0.61 0.177 1 st der MSC 15 0.60 0.182 0.217 0.55 0.200 SNV 15 0.59 0.184 0.219 0.52 0.202 Detrend 9 0.49 0.205 0.229 0.49 0.199 2 nd der MSC 14 0.59 0.185 0.218 0.47 0.213 SNV 14 0.56 0.189 0.224 0.44 0.218 Detrend 19 0.60 0.181 0.229 0.48 0.210

PAGE 98

98 LIST OF REFERENCES Abebe, A. T. 2006. Total Sugar and Maturity Evaluation of Intact Watermelon Using Near Infrared Spectroscopy. Journal Near Infrared Spectroscopy 14 : 67 70. Agricultural Statistic Board. 2008. Citrus Fruit 2008 Summary. 1 49. Braddock, R. J. 1999. Handbook of citrus by products and p r ocessing t echnolog y. New York, NY : John Wiley & Sons Inc. Cayuela, J. A. 2008. Vis/NIR Soluble Solids Prediction in Intact Oranges (Citrus sinesi L.) c v. Va lencia Late by reflectance. Postharvest Biology and Technology 47 : 75 80. Cen, H., Y. Bao, Y. He and D. Sun. 2007. Visible and Near Infrared Spectroscopy for Rapid Detection of Citric and Tartaric Acids in Orange Juice. Journal of Food Engineering 82 : 253 260. Chang, W. H., S. Chen and C. C. Tsai. 1998. Development of Universal Algorithm for use of NIR in Estimation of Soluble Solids in Fruit Juices Transaction of the ASABE 41(6): 1793 1745. Chen, J. Y., H. Zhang and R. Matsunaga. 2006. Rapid Determinat ion of the Main Organic Acid Composition of Raw Japanese Apricot Fruit Juices Using Near Infrared Spectroscopy. Journal of Agricultural and Food Chemistry 54 ( 9 ): 652 9657. Clark, C. J., A. V. McGlone and R. B. Jordan. 2003. Detection of Brownheart in 'Bra eburn' Apple by Transmission NIR spectroscopy Postharvest Biology and Technology 28 : 87 96. Dhanoa, M. S., S. J. Lister, R. Sanderson and R. J. Barnes. 1994. The Link between Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV) trans formations of NIR spectra Journal Near Infrared Spectroscopy 2 : 43 47. Dupuy, N., B. Meurens, P. Legrand and J. P. Huvenne. 1992. Determination of Sugars and Organic Acids in Fruit Juices by FT Mid IR Investigation of Dry Extract Applied Spectroscopy 4 6(5): 860 863. Dupuy, N., M. Meurens, B. Sombert, P. Legrand and J. P. Huvenne. 1993. Multivariate Determination of Sugar Powders by Attenuated Total Reflectance Infrared Spectroscopy. Applied Spectroscopy 47(4): 452 457. Ehsani, M. R., S. K. Upadhyaya, W. R. Fawcett, L. V. Protsailo and D. Slaughter. 2001. Feasibility of Detecting Soil Nitrate Content Using a Mid Infrared Technique. Transaction of ASAE 44(6): 1931 1940. Fujiwara, T. and T. Honjo. 1995. Determination of Sugar and Acid Contents in Fruit J uice of Satsuma Mandarin by Near Infrared Spectroscopy Journal of the Japanese Society For Food Science and Technology 42(2): 109 117.

PAGE 99

99 Geladi, P. 2003. Chemometrics in Spectroscopy Part 1. Classical Chemometrics. Spectrochimica Acta Part B 58 : 767 782. Geladi, P., B. Sethson, J. Nystrom, T. Lillhonga, T. Lestander and J. Burger. 2004. Chemometrics in Spectroscopy Part 2. Examples. Spectrochimica Acta Part B 59 : 1347 1357. Giangiacomo, R. and G. G. Dull. 1986. Near Infrared Spectrophotometric Determination of Individual Sugars in Aqueous Mixtures. Journal of Food Science 51(3): 679 683. Gomez, A. H., Y. He and A. G. Pereira. 2006. Non destructive Measurement of Acidity, Sol uble Solids and Firmness of Satsuma Mandarin Using Vis/NIR Spectroscopy Techniques Journal of Food Engineering 77 : 313 319. Guthrie, J. A., C. J. Lienbenberg and K. B. Walsh. 2006. NIR Model Development and Robustness in Prediction of Melon Fruit Total Soluble Solids. A ustralian Journal of Agricultural Research 5 : 71 8. Guthrie, J. A., D. J. Reid and K. B. Walsh. 2005. Assessment of Internal Quality of Mandarin Fruit. 1. NIR Calibration Model Development. Australian Journal of Agricultural Research 56: 405 416. Guthrie, J. A., D. J. Reid and K. B. Walsh. 2005. Assessment of Internal Quality Attributes of Mandarin Fruit. 2. NIR Calibration Model Robustness Australian Journal of Agricultural Research 56 : 417 426. Huang, H., H. Yu, H. Xu and Y. Ying. 2 008. Near Infrared Spectroscopy for On/In Line Monitoring of Quality in Foods and Beverages. Journal of Food Engineering 87 : 303 313. Irudayaraj, J. and J. Tewari. 2003. Simultaneous Monitoring of Organic Acids and Sugars in Fresh and Processed Apple Juic e by Fourier Transform Infrared Attenuated Total Reflection Spectroscopy. Applied Spectroscopy 57(12): 1599 1604. Kemsley, E. K., L. Zhuo, M. K. Hammouri and R. H. Wilson. 1992. Quantitative Analysis of Sugar Solutions Using Infrared Spectroscopy. Journal of Food Chemistry 44 : 299 304. Kimball, D. A. 1999. Citrus processing: A complete g uide Second ed Gaithersburg,Maryland : Aspen Publisher. L ammertyn, J., B. Nicolai, K. Ooms, V. D. Smedt and D. J. Baerdemaeker. 199 8. Non destructive measurement o f acid ity, soluble solids and firmness of Jonagold apples using NIR spectroscopy Transaction of the ASABE 41(4): 1089 1094. Lanza, E. and B. W. Li. 1984. Application for Near Infrared Spectroscopy for Predicting the Sugar Content of Fruit Juices Journal of Fo od Science 49 : 995 998.

PAGE 100

100 Larrain, M., A. R. Guesalaga and E. Agosin. A multipurpose Portable Instrument for Determining Ripeness in Wine Grapes Using NIR Spectroscopy. IEEE Transaction on Instrumentation and Measurement 57(2): Li, W., P. Goovaerts and M. Meurens. 1996. Quantitative Analysis of Individual Sugars and Acids in Orange Juices by Near Infrared Spectroscopy of Dry Extract. Journal Agricultural Food Chemical 44 : 2252 2259. Long, R. L. and K. B. Walsh. 2006. Limitation to the measurement of Intac t Melon Total Soluble Solids Using Near Infrared Spectroscopy. A ustralian Journal of Agricultural Research 57 : 403 410. Lu, H., H. Jiang, X. Fu, H. Yu, H. Xu and Y. Ying. 2008. Non Invasive Measurement if The Internal Quality of Intact 'Gannan' Navel Oranges by VIS/NIR Spectroscopy. Transaction of the ASABE 51(3): 1009 1014. McGlone, A. V., D. G. Fraser, R. B. Jordan and R Kunnemeyer. 2003. Internal Quality Assesment of Mandarin Fruit by Vis/NIR Spectroscopy. Journal Near Infrared Spectroscopy 11 : 323 332. Meurens, M. 1982. Analysis of Aqueous Liquids by Near Infrared Reflectance Spectroscopy. In International Meeting of Chemical Engineering (82): Miller, W. M. and M. Zude Sasse. 2004. NIR Based Sensing to Measure Soluble Solids Content of Florida Citrus. Applied Engineering in Agriculture 20(3): 321 327. Mirouze, D. L. F., J. C. Boulou, N. Dupuy, M. Meurens, J. P. Huven ne and P. Legrand. 1993. Quantitative Analysis of Glucose Syrups by ATR/FT IR Spectroscopy. Applied Spectroscopy 47(8): 1187 1191. Naes, T., T. Isaksson, T. Fearn and T. Davies. 2004. A User Friendly Guide to Multivariate Calibration and Classification Charlton, Chichester, UK: NIR publications. Nicolai, B. M., K. Beullens, E. Bobelyn, A. Peirs, W. Saeys, I. K. Theron and J. Lammertyn. 2007. Nondestructive measurement of fruit and vegetable quality by means of NIR spectroscopy: A review Postharvest Bio logy and Technology 46 : 99 118. Ou, A. S., S. Lin, T. Lin, S. Wu and M. Tiarn. 1997. Studies on the Determination of Quality Related Constituents in Ponkan mandarin by Near Infrared Spectroscopy Journal of the Chinese Agricultural Chemical Society 35(4): 462 474. Peirs, A., J. Lammertyn, K. Ooms and B. M. Nicolai. 2000. Prediction of the Optimal Picking Date of Different Apple Cultivars by Means of VIS/NIR spectroscopy. Postharvest Biology and Technology 21189 199. Peng, Y. and R. Lu. 2002. New Approach es of Analyzing Multispectral Scattering Profiles for Predicting Apple Fruit Firmness and Soluble Solids Content. ASABE Meeting Paper No 066234 St. Joseph, Mich.

PAGE 101

101 Polesello, A., R. Giangiacomo, E. Forni and F. Braga. 1990. The use f NIRS Spectrophotometry to Estimate the Pectin Substances in Fruit and Fruit Products. Carbohydrate Polymers 12 : 27 38. Roggo, Y., P. Chalus, L. Maurer, C. Lema Martinez, A. Edmond and N. Jent. 2007. A Review of Near Infrared Spectroscopy and Chemometrics in Pharmaceutical Techn ologies. Journal of Pharmaceutical and Biomedical Analysis 44 : 683 700. Saranwong, S., J. Sornsrivichai and S. Kawano. 2004. Prediction of Ripe Stage Eating Quality of Mango Fruit From Its Harvest Quality Measured Nondestructively by Near Infrared Spectro scopy. Postharvest Biology and Technology 31 : 137 145. Settle, F.A. 1997. Han dbook of Instrumental Techniques for Analytical Chemistry, Prentice Hall. Sivakesava, S. and J. Irudayaraj. 2000. Determination of sugars in Aqueous Mictures Using Mid Infrared Spectroscopy. Applied Engineering in Agriculture 16(5): 543 550. Smith, B. C. 2002. Quantitative spectroscopy: Theory and p ractice San Diego, California : Academic Press. Smith, B. C. 1999. Infrared spectral i nterpretation: A systematic approach Bota R acon, Florida: CRC Press LLC. Temma, T., K. Hanamatsu and F. Shinoki. 2002. Development of a Portable Near Infrared Sugar Measuring Instrument. Journal Near Infrared Spectroscopy 10 : 77 83. Tetra, P. 1998. The orange b ook Lund, Sweeden: Tetra Pak Proce ssing Systems AB, Ruter Press. Tsai, F. and W. Philpot. 1998. Derivative Analysis of Hyperspectral Data. Remote Sensing Environment 66 : 41 51. Tsuta, M., J. Sugiyama and Y. Sagara. 2002. Near Infrared Imaging Spectroscopy Based on Sugar Absorption Bands for Melons. Journal of Agricultural and Food Chemistry 50(1): 48 52. Williams, P. and K. Norris. 1987. Ne ar i nfrared technology in the agricultu ral and food i ndustries Minnesota, USA: American Association of Cereal Chemist, Inc. Wise, B. M., N. B. Gallagher, R. Bro, J. M. Shaver, W. Windig and S. R. Koch. PLS Toolbox ver4.0, to be Used with Matlab Yan de, L., Y. Yi bin, F. Xiaping and L. Huis han. 2007. Experiments on Predicting Sugar Content in Apples by FT IR Technique. Journal of Food Engineering 80 : 986 989. Ying, Y. B., Y. D. Liu, J. P. Wang, X. P. Fu and Y. B. Li. 2005. Fourier Transform Near Infrared Determination of Total Soluble Solid s and Available Acid in Intact Peaches Transaction of the ASABE 48(1): 229 234.

PAGE 102

102 Zude Sasse, M., M. Pflanz, C. Kaprielian and B. L. :. Aivazian. 2008. NIRS as a Tool for Precision Horticulture in the Citrus Industry. Biosystem Engineering 99 : 455 459. Zud e Sasse, M., W. B. Herppich, H. Dou and W. M. Miller. 2002. Non Destructive Prediction of Grapefruit Sensitivity to Chilling Injury. ASAE Meeting Paper No. 026187 St Joseph, Mich.

PAGE 103

BIOGRAPHICAL SKETCH Rashidah R u s lan was born in 1983 to Mr Ruslan A bdul Shukor and Mrs. Padzilah Othman. She was sent to Mara Junior Science College boarding school since she was 13 years old. She received her b achelor degree in biological & agricultural e ngineering from University Putra Malaysia, Malays ia in 2006 and later was appointed as an Assistant Researcher for a year in the same university. Knowing her passion in the academia world, she pursued her dream to be an academician. With high hopes and financial support from the Ministry of Higher Educa tion Malaysia she flew to the United States to do her Master of Engineering at the University of F lorida in August 2007. In future, she hopes to finish her higher education studies with flying colors and return home with a good amount of knowledge and experiences to be shared with her future stu dent