Dublin Institute of Technology
ARROW@DIT
Doctoral
Science
2010-01-01
Visible and Hyperspectral Imaging Systems for the
Detection and Discrimination of Mechanical and
Microbiological Damage of Mushrooms
Edurne Gaston (Thesis)
Dublin Institute of Technology
Follow this and additional works at: http://arrow.dit.ie/sciendoc
Part of the Environmental Microbiology and Microbial Ecology Commons
Recommended Citation
Gaston, E.: Visible and Hyperspectral Imaging Systems for the Detection and Discrimination of Mechanical and Microbiological
Damage of Mushrooms. Doctoral Thesis. Dublin Institute of Technology, 2010.
This Theses, Ph.D is brought to you for free and open access by the Science
at ARROW@DIT. It has been accepted for inclusion in Doctoral by an
authorized administrator of ARROW@DIT. For more information, please
contact yvonne.desmond@dit.ie, arrow.admin@dit.ie.
This work is licensed under a Creative Commons AttributionNoncommercial-Share Alike 3.0 License
Visible and hyperspectral imaging systems for the
detection and discrimination of mechanical and
microbiological damage of mushrooms
Edurne Gaston Estanga, B.Sc.
A thesis submitted to Dublin Institute of Technology
in fulfilment of the requirements for the degree of
DOCTOR OF PHILOSOPHY
School of Food Science and Environmental Health
College of Science
Dublin Institute of Technology
Supervisors:
Dr. Jesús M. Frías Celayeta
Dr. Patrick J. Cullen
2010
Abstract
Horticultural products such as mushrooms are exposed to environmental conditions
during their postharvest life, which may affect product quality. Loss of whiteness during
storage is particularly important in the mushroom industry. Rough handling and
distribution, fruiting body senescence and bacterial infections are among the main
causes of mushroom discolouration. The aim of this work was to study the use of visible
and hyperspectral imaging (HSI) systems for the detection and discrimination of
mechanical and microbiological damage of mushrooms.
This piece of research involved a) monitoring the browning of mushroom with visible
computer imaging systems, b) investigating the effect of mechanical damage on the
kinetics of enzymes responsible for mushroom browning, c) exploring the potential use
of Vis-NIR HSI to predict PPO activity in mushroom caps and d) studying the potential
application of Vis-NIR HSI for microbial and viral detection on mushroom caps and for
their discrimination from mechanical damage.
Results presented in this thesis show that the efficacy of commercial webcams was
limited in the detection of mechanical damage on mushroom caps. Damage increased
the activity of PPOs on mushroom pileipellis, but the effect of the extent of damage was
not significant at the levels of study. Vis-NIR HSI showed some potential as a tool to
estimate the activity of PPO enzymes on mushroom caps. The combination of HSI with
chemometric tools allowed for the differentiation of mechanically and microbiologically
damaged mushroom classes.
Results from this study could be used for developing non-destructive monitoring systems
for mechanical and microbiological damage detection and discrimination. The potential
application of such systems as on-line process analytical tools would facilitate rapid
assessment of mushroom quality.
i
Declaration
I certify that this thesis which I now submit for examination for the award of Doctor of
Philosophy (PhD) is entirely my own work and has not been taken from the work of
others, save and to the extent that such work has been cited and acknowledged within
the text of my work.
This thesis was prepared according to the regulations for postgraduate study by
research of the Dublin Institute of Technology and has not been submitted in whole or in
part for another award in any institute.
The work reported on in this thesis conforms to the principles and requirements of the
Institute’s guidelines for ethics in research.
The Institute has permission to keep, lend or copy this thesis in whole or in part, on
condition that any such use of the material of the thesis be duly acknowledged.
Signature
Date
Edurne Gaston Estanga (candidate)
ii
09.08.2010
We shall not cease from exploration
And the end of all our exploring
Will be to arrive where we started
And know the place for the first time.
T.S. Eliot, 1942.
iii
Acknowledgements
This thesis belongs not only to me, but also to the many people who have contributed to
make this achievement possible.
It is a pleasure to thank all the people who have generously shared their time and
knowledge with me:
Dr. Jesús M. Frías, for the many things I learnt while working by his side.
Dr. PJ Cullen, for playing strategically and showing support when it was most needed.
Dr. Carl Sullivan, for all his selfless help in experimental design.
Prof. Colm O’Donnell, for a warm welcome in UCD.
Dr. Aoife Gowen, for the supervision that supervisors don’t have time for.
Carlos Esquerre, for being open to discuss hyperspectral and chemometric issues.
Dr. Paula Bourke and Dr. Patricia Nobmann, for providing the tips a chemist needs to
deal with microbes.
Ted Cormican and Dr. Helen Grogan, for the provision of mushrooms and kind advice.
Noel Grace and Tony Fitzpatrick, for helping out with the things a student in a lab
needs helping out with.
Seán Dowling, for proofreading the thesis thoroughly.
The academic, administrative, technical and research staff in both institutions.
On a personal level, some people have been invaluable:
Aitor, thanks to whom following this path has felt so natural.
Aoife O’Gorman, for holding out her hand to me.
Arturo, for his ability to adapt and the very special person he is.
Carl, for innumerable expressions of affection.
Haizea, for the enormous amount of patience she has had with me.
Héctor, for being one phone call away, day or night!
Josu, for sharing a home with me.
Laura Álvarez, Leixuri and Patricia, for being there on good and bad days.
Laura Massini, for sharing her taste for music and art.
iv
Nagore, for being the mate my soul needs.
Seán, for loving me tenderly.
Txopin, for being there when she was and was not among us.
Vasilis, for some special moments when the going was tough.
There are lots of people who have made Dublin a very special place to be:
Adriana, Alka, Anne, Aoife Kelly, Aoife Smith, Beatriz, Bérengère, Celine, Colette,
Dharmendra, Diarmuid, Esther, Flórence, Jacek, Jane, Jesús, John, Khadija, Larry,
Laura Thornton, Lorena, Lubna, Maeve, Manuel, Mary, Mato, Michael, Mohammed,
O’Toole family, Prabhash, Sapna, Santos, Sergi, Sharon, Sofia, Sonal, Sylvain,
Tommek, Ulla, Urtzi, Wojtek and Zied.
The following, being “so far, so close”, have contributed enormously to my well-being:
Ainara, Aintzane, Andrea, Ane, Antton, Bengoa family, Estanga-Mujika family,
Gaston-San Juan family, Iñaki Etaio, Iñaki Rikarte, Joseba, Kintana, Laura, Lope,
Maider, Mederi, Mery, Mikel, Mónica, Natxo, Nuria, Oihane, Patxi, Pily, Unai and
Urko.
My final words go to my family, my treasure. Their support and faith in me, together with
my persistence, have been the key to success. A big Thank you goes to:
My parents, for an outstanding example of modesty and hard work.
Natalia, for being generous and protective.
Lidia, for her devotion, clarity and candour.
Maria, for moments of creativity and mischief.
Rakel, for being open to share worries and thoughts, and for her wise words.
Andoni, Iker, Gari and Pablo, for caring for me, each in different ways.
My nephews and nieces - Eñaut, Ander, Beñat, Aiert, Naroa, Laida and Jone - for
welcoming me with their biggest smiles every time we meet.
v
Abbreviations list
a*
CIELAB greenness/ redness value
b*
CIELAB blueness/yellowness value
CCD
Charge-coupled device
CIELAB
Colour space defined by the International Commission on Illumination.
CV
Computer vision
ΔE
Total colour difference
EAU
Enzyme activity units
GSV
Grayscale value
HSI
Hyperspectral imaging
λ
Wavelength
LOD
Limit of detection
LV
Latent variable
# LV
Number of latent variables
L*
CIELAB lightness/darkness value
MD
Mushrooms subjected to mechanical damage
MLR
Multiple linear regression
MSC
Multiple scatter correction
MVX
Mushroom virus X
PC
Principal component
PCA
Principal component analysis
PCR
Principal component regression
PLS-DA
Partial least squares-discriminant analysis
PLSR
Partial least squares regression
PPO
Polyphenol oxidase
PT
Mushrooms inoculated with P. tolaasii
RGB
Red, green, blue
vi
RMSECV
Root mean squared error of cross-validation
RMSEP
Root mean squared error of prediction
ROI
Region of interest
RPD
Ratio of percentage deviation
SNV
Standard normal variate
U
Undamaged mushrooms
Vis-NIR
Visible near-infrared
vii
Table of contents
1. INTRODUCTION ....................................................................................................... 1 2. LITERATURE REVIEW ............................................................................................ 4 2.1 An insight into mushrooms ......................................................................................... 5 2.1.1 History of mushroom cultivation ........................................................................... 5 2.1.2 The Fungi kingdom .............................................................................................. 6 2.1.3 Morphology of mushrooms .................................................................................. 6 2.1.4 Mushroom growing .............................................................................................. 8 2.1.5 Irish mushroom industry..................................................................................... 10 2.2 Key issues for the mushroom industry ..................................................................... 12 2.2.1 Mushroom shelf-life............................................................................................ 12 2.2.2 Variability ........................................................................................................... 13 2.3 Mushroom quality ..................................................................................................... 16 2.3.1 Mushroom colour as a quality parameter........................................................... 17 2.4 Mushroom discolouration ......................................................................................... 19 2.4.1 Why brown? ....................................................................................................... 20 2.4.2 Causes of discolouration in healthy mushrooms ............................................... 24 2.4.3 Causes of discolouration in diseased mushrooms............................................. 27 2.5 Computer imaging for product quality measurement................................................ 31 2.5.1 The necessity for automating quality assessment ............................................. 31 2.5.2 Computer vision ................................................................................................. 32 2.5.3 Hyperspectral Imaging ....................................................................................... 38 2.6 Spectral data pre-processing techniques ................................................................. 52 2.6.1 Standard normal variate..................................................................................... 53 2.6.2 Multiplicative scatter correction .......................................................................... 54 2.6.3 Differentiation: Savitzky-Golay ........................................................................... 55 viii
2.7 Chemometrics .......................................................................................................... 56 2.7.1 Qualitative data analysis .................................................................................... 56 2.7.2 Quantitative data analysis .................................................................................. 58 3. AIM AND OBJECTIVES ......................................................................................... 61 4. EFFICACY OF A VISIBLE IMAGING SYSTEM TO DIFFERENTIATE AND
MONITOR THE BROWNING OF MUSHROOMS .......................................................... 63 4.1 Introduction ............................................................................................................... 64 4.2 Materials and methods ............................................................................................. 65 4.2.1 Mushroom supply............................................................................................... 65 4.2.2 Simulated process and transport damage protocol ........................................... 66 4.2.3 Image acquisition system................................................................................... 68 4.2.4 Experimental design .......................................................................................... 69 4.2.5 Image acquisition ............................................................................................... 69 4.2.6 Image analysis ................................................................................................... 70 4.3 Results and discussion ............................................................................................. 72 4.3.1 Standardisation of damage protocol .................................................................. 72 4.3.2 Browning kinetics ............................................................................................... 75 4.3.3. Limit of detection ............................................................................................... 87 4.3.4 Time to discrimination ........................................................................................ 93 4.4 Conclusions .............................................................................................................. 95 5. THE EFFECT OF MECHANICAL DAMAGE ON THE ENZYME ACTIVITY OF
MUSHROOMS ............................................................................................................... 97 5.1 Introduction ............................................................................................................... 98 5.2 Materials and methods ........................................................................................... 100 5.2.1 Sample preparation.......................................................................................... 100 5.2.2 Experimental design ........................................................................................ 101 ix
5.2.3 Extract preparation........................................................................................... 101 5.2.4 Spectrophotometric assay ............................................................................... 101 5.2.5 Data analysis ................................................................................................... 102 5.3 Results and discussion ........................................................................................... 103 5.3.1 RGB images of undamaged and damaged mushrooms .................................. 103 5.3.2 PPO kinetics .................................................................................................... 104 5.3.3 The effect of Damage, Temperature and Day on enzyme activity ................... 107 5.4 Conclusions ............................................................................................................ 111 6. PREDICTION OF PPO ACTIVITY USING VIS-NIR HYPERSPECTRAL IMAGING
ON MUSHROOM CAPS .............................................................................................. 113 6.1 Introduction ............................................................................................................. 114 6.2 Material and methods ............................................................................................. 116 6.2.1 Sample preparation.......................................................................................... 116 6.2.2 Experimental design ........................................................................................ 116 6.2.3 Hyperspectral imaging ..................................................................................... 116 6.2.4 Enzyme extraction ........................................................................................... 118 6.2.5 Image processing and data analysis................................................................ 118 6.3 Results and discussion ........................................................................................... 124 6.3.1 Spectral variation arising from mushroom shape............................................. 124 6.3.2 Spectra............................................................................................................. 126 6.3.2 Enzyme activity ................................................................................................ 126 6.3.3 Modelling.......................................................................................................... 127 6.4 Conclusions ............................................................................................................ 137 7. VIS-NIR
HYPERSPECTRAL
IMAGING
FOR
THE
DETECTION
AND
DISCRIMINATION OF BROWN BLOTCH ON MUSHROOM CAPS .......................... 139 7.1 Introduction ............................................................................................................. 140 x
7.2 Materials and methods ........................................................................................... 141 7.2.1 Sample preparation.......................................................................................... 141 7.2.2 Preparation of pathogenic P. tolaasii solution .................................................. 142 7.2.3 Hyperspectral imaging ..................................................................................... 144 7.2.4 Confirmation of P. tolaasii ................................................................................ 144 7.2.5 Image processing............................................................................................. 144 7.2.6 Partial least squares- discriminant analysis (PLS-DA) ..................................... 145 7.2.7 Prediction maps ............................................................................................... 146 7.2.8 Mushroom classification................................................................................... 148 7.3 Results and discussion ........................................................................................... 151 7.3.1 RGB images..................................................................................................... 151 7.3.2 Confirmation of P. tolaasii ................................................................................ 152 7.3.4 Spectra............................................................................................................. 153 7.3.5 PLS-DA analysis .............................................................................................. 155 7.3.6 Prediction maps ............................................................................................... 157 7.3.7 Mushroom classification................................................................................... 161 7.4 Conclusions ............................................................................................................ 163 8. HYPERSPECTRAL IMAGING FOR THE DISCRIMINATION OF MUSHROOMS
INFECTED WITH MUSHROOM VIRUS X ................................................................... 164 8.1 Introduction ............................................................................................................. 165 8.2 Materials and methods ........................................................................................... 166 8.2.1 Sample preparation.......................................................................................... 166 8.2.2 Experimental design ........................................................................................ 166 8.2.3 Image acquisition ............................................................................................. 167 8.2.4 Data analysis ................................................................................................... 167 8.3 Results and discussion ........................................................................................... 168 8.4 Conclusions ............................................................................................................ 171 xi
9. OVERALL CONCLUSIONS AND RECOMMENDATIONS .................................. 172 REFERENCES ............................................................................................................. 176 LIST OF PUBLICATIONS ............................................................................................ 203 Journal publications ...................................................................................................... 204 Conference abstracts ................................................................................................... 204 xii
List of figures
Figure 2.1 Diagram of mushroom tissues. ................................................................... 7 Figure 2.2 Mushroom growing systems. .................................................................... 10 Figure 2.3 CIELAB diagram. ...................................................................................... 18 Figure 2.4 Structure of phenol. .................................................................................. 20 Figure 2.5 Natural melanogenous phenols present in A. bisporus. ........................... 21 Figure 2.6 Hydroxylation of a monophenol (Cresolase activity)................................. 21 Figure 2.7 Oxidation of an o-diphenol (Catecholase activity). ................................... 21 Figure 2.8 Structure of a melanin. ............................................................................. 22 Figure 2.9 A. bisporus caps showing typical symptoms of bacterial blotch. .............. 28 Figure 2.10 A. bisporus caps showing symptoms of MVX disease. .......................... 30 Figure 2.11 Diagram of a typical computer vision system. ........................................ 34 Figure 2.12 Hypercube structure. .............................................................................. 40 Figure 2.13 Unfolding of a hypercube. ....................................................................... 41 Figure 2.14 Methods for acquiring three dimensional hypercubes. ........................... 42 Figure 2.15 Diagram and image of a pushbroom hyperspectral imaging system. ..... 43 Figure 2.16 Schematic diagram of hyperspectral data analysis process. .................. 45 Figure 2.17 Effects of pre-processing techniques on mushroom spectra. ................. 55 Figure 4.1 Stages in mushroom harvesting and transportation. ................................ 64 Figure 4.2 Devices tested for inducing mechanical damage in mushrooms. ............. 67 Figure 4.3 Average CIELAB colour coordinates as a function of damage time. ........ 73 Figure 4.4 Average CIELAB colour coordinates and their calibration curves as a
function of damage time. ........................................................................................... 74 Figure 4.5 Grayscale Value (GSV) of mushrooms as a function of storage time. ..... 76 Figure 4.6 Delta Grayscale Value (ΔGSV) of mushrooms as a function of storage
time. ........................................................................................................................... 77 xiii
Figure 4.7 Delta Grayscale Value (ΔGSV) of the MD 20 / C 3 experiment as a
function of storage time. ............................................................................................ 79 Figure 4.8 White-referenced Corrected ∆GSV (∆GSVref) of mushrooms as a function
of storage time. .......................................................................................................... 80 Figure 4.9 White-referenced Corrected ∆GSV (∆GSVref) of the MD 20 / C3 experiment
as a function of storage time. ..................................................................................... 81 Figure 4.10 Weighed ∆GSV (∆GSVref.w) of mushrooms as a function of storage time.
................................................................................................................................... 83 Figure 4.11 GSVref.w time derivative (D∆GSVref.w) of mushrooms as a function of
storage time. .............................................................................................................. 84 Figure 4.12 Weighed GSV (GSVref.w) of mushrooms as a function of storage time. .. 86 Figure 4.13 Weighed ∆GSV (∆GSVref.w) of MD mushrooms as a function of storage
time. ........................................................................................................................... 89 Figure 4.14 D∆GSVref.w derivative (D∆GSVref.w) of MD mushrooms as a function of
storage time. .............................................................................................................. 91 Figure 4.15 Weighed GSV (GSVref.w) of MD mushrooms as a function of storage time.
................................................................................................................................... 92 Figure 5.1 Cresolase and catecholase activities of catechol oxidase. ....................... 98 Figure 5.2 Representative RGB images of various damage levels at day 0 of storage.
................................................................................................................................. 103 Figure 5.3 Representative RGB images of various damage levels at different days of
storage. .................................................................................................................... 104 Figure 5.4 PPO enzyme activity as a function of storage time. ............................... 105 Figure 5.5 Effect of first-order factors on the enzyme activity of mushrooms. ......... 108 Figure 5.6 Effect of second-order Damage*Time interaction on the enzyme activity of
Rep.1 mushrooms. .................................................................................................. 109 Figure 5.7 Effect of the third-order interaction (i.e. Damage*Temperature*Time) on
the enzyme activity of Rep. 1 mushrooms. .............................................................. 111 xiv
Figure 6.1 Schematic representation of a hyperspectral imaging hypercube. ......... 115 Figure 6.2 Pushbroom hyperspectral imaging system ............................................. 117 Figure 6.3 Diagram showing modelling strategies 1 and 2. ..................................... 119 Figure 6.4 Typical hyperspectral image of the surface of a mushrooms. ................ 125 Figure 6.5 Summary of spectral and enzymatic characteristics of the data set. ...... 127 Figure 6.6 Predicted PPO vs measured PPO activity. ............................................. 131 Figure 6.7 Undamaged mushroom caps. ................................................................ 133 Figure 6.8 Damaged mushroom caps. .................................................................... 133 Figure 6.9 Prediction of an imaginary line drawn through the centre of mushroom
caps. ........................................................................................................................ 136 Figure 7.1 Decision tree for mushroom hypercube classification. ........................... 149 Figure 7.2 Histograms showing number of mushrooms as a function of the
percentage of pixels in the BINPT binary images. ..................................................... 150 Figure 7.3 Representative RGB images of the mushrooms under investigation. .... 151 Figure 7.4 Colonies in PAB plates. .......................................................................... 152 Figure 7.5 Mean spectra of the mushroom classes under investigation. ................. 153 Figure 7.6 RMSECV and performance statistics of PLS-DA models as a function of
the number of LV. .................................................................................................... 155 Figure 7.7 Prediction images of mushrooms (after no erosion in Step 2). ............... 157 Figure 7.8 Effect of varying the radius of the SE in Step 2. ..................................... 159 Figure 7.9 Effect of the omitting/incorporating Step 4. ............................................. 161 Figure 8.1 PCA scores plot for Vis MSC-corrected spectra representing control
mushrooms and mushrooms infected with different strains. .................................... 169 Figure 8.2 PCA scores plot for Vis raw reflectance spectra representing control
mushrooms and mushrooms infected at different timing/concentrations. ................ 170 Figure 8.3 PCA scores plot for NIR MSC-corrected reflectance spectra representing
different storage days. ............................................................................................. 171 xv
List of tables
Table 4.1 Experimental design. ................................................................................. 69 Table 4.2 GSV-derived parameters calculated in order to cancel the different sources
of noise in the signal. ................................................................................................. 71 Table 4.3 Limits of detection (LODs) set using the different GSV-derived parameters.
................................................................................................................................... 71 Table 4.4 Total colour difference (ΔE) as a function of damage time. ....................... 75 Table 4.5 Estimation of the weight of the reference tile for each camera. ................. 82 Table 4.6 LOD values and minimum and maximum values of MD mushrooms for
each camera and parameter. ..................................................................................... 87 Table 4.7 Time to discrimination of all cameras and damage levels, based on
LODΔGSVref.w and (ΔGSVref.w )MD . ................................................................................ 93 Table 4.8 Maximum time for MD/U discrimination of each camera and damage level,
based on
LODDΔGSVref.w and D∆GSVref.w. ...................................................................... 94 Table 4.9 Time to discrimination of each camera and damage level, based on
LODGSVref.w and (GSVref.w )MD ...................................................................................... 95 Table 5.1 ANOVA of Damage, Temperature and Time factors and their interactions
on Rep. 1 data. ........................................................................................................ 107 Table 6.1 Ratio percentage deviation (RPD) for the different model strategies,
spectral pre-processing techniques and chemometric methods under investigation.
................................................................................................................................. 128 Table 7.1 Performance statistics at spectra level of PLS-DA models built on
reflectance spectra. ................................................................................................. 156 Table 7.2 Performance statistics at pixel level of PLS-DA models built on raw
reflectance spectra. ................................................................................................. 162 Table 8.1 Experimental design of the first study. ..................................................... 166 xvi
Table 8.2 Experimental design of the second study. ............................................... 167 Table 8.3 Chemometric tools employed and main purpose of utilisation................. 168 Table 8.4 Summary of results obtained. .................................................................. 169 xvii
1. INTRODUCTION
Chapter 1. Introduction
Button mushroom (Agaricus bisporus) production is a fermentation industry that is able
to produce quality protein from cellulose-based agricultural by-products. White button
mushrooms are one of the most important horticultural crops grown in Ireland, with an
export value exceeding €100 million in 2008 (Bord Bía, 2009). The Irish mushroom
sector is competing with other producing countries such as Poland and the Netherlands
in the British import market, which puts it under pressure to develop market led value
added products.
Fresh mushrooms are highly perishable and have a short shelf-life compared to most
vegetables. They are very sensitive to (post-)harvest practices, including unhygienic
growing conditions, rough handling and distribution, which can cause irreversible injuries
and enhance mushroom cap discolouration. Loss of whiteness has an important
economical impact on the Irish mushroom industry, since the colour of the cap is the first
characteristic consumers notice when purchasing mushrooms. There is a definite need
for research in the fields of i) physical and biological factors involved in the loss of
mushroom whiteness and ii) techniques for its detection/discrimination.
Computerised
image
analysis
offers
advantages
over
human-assisted
visual
assessment of produce, such as objectivity, accuracy, consistency and reduced labour
costs and time of analysis. In recent years, several techniques combining machine
vision and artificial intelligence have proven effective for mushroom inspection.
Hyperspectral Imaging (HSI) is an emerging technique that combines conventional
imaging and spectroscopy. This integrated feature expands the capability for detecting
chemical constituents and intrinsic characteristics of a product. Simple computer and
hyperspectral imaging systems offer promising scope for research in the field of rapid
detection, identification and discrimination of blemishes on mushrooms caps.
A summary of this thesis and its division into relevant chapters is outlined below:
2
Chapter 1. Introduction
Chapter 2: Literature review, which presents published information in the field and
results relevant to the topic of this thesis.
Chapter 3: Aim and objectives of the thesis.
Chapter 4: Efficacy of a visible imaging system to differentiate and monitor the browning
of mushrooms, where inexpensive webcams were used to follow the browning kinetics
of sound and bruised mushrooms.
Chapter 5: The effect of mechanical damage on the enzyme activity of mushrooms,
which investigates how the kinetics of mushroom polyphenol oxidases (PPOs) are
affected by damage and storage.
Chapter 6: Prediction of polyphenol oxidase activity using Vis-NIR hyperspectral imaging
on mushroom caps, which explores the correlation between the spectral response of
mushroom caps and the activity of the PPO enzymes responsible for mushroom
browning.
Chapter 7: Vis-NIR hyperspectral imaging for the detection and discrimination of brown
blotch disease on mushrooms caps, which studies the ability of this technique to identify
and differentiate microbial lesions on mushroom caps.
Chapter 8: Hyperspectral imaging for the discrimination of mushrooms infected with
Mushroom Virus X, which investigates the ability of this technique to discriminate
mushrooms subjected to viral infection fro control ones.
Chapter 9: Overall conclusions and recommendations, where the findings of this piece of
research are summarised and some guidelines for future work are provided.
3
2. LITERATURE REVIEW
Chapter 2. Literature review
2.1 An insight into mushrooms
2.1.1 History of mushroom cultivation
Mushrooms have been part of fungal diversity for around 300 million years. Prehistoric
humans probably used mushrooms collected in the wild as food and possibly for
medicinal purposes. After the intentional cultivation of plants for food by early
civilisations, it was inevitable that mushrooms would eventually be deliberately grown
too.
Mushroom cultivation did not come into existence until the 7th century, when the Chinese
cultivated Auricularia auricula on wooden logs. These mushrooms were grown outdoors
without using any specially prepared spawns. There are historical records showing that
other species were cultivated in a similar manner during the following centuries.
The biggest technological advance in the field happened around AD 1600, when
Agaricus bisporus (A. bisporus), commonly known as champignon or button mushroom,
was first cultivated in France. The first cultivation techniques used horse manure as a
substrate. After mushroom bed harvest, the substrate was broken up, dried and reused
to inoculate new beds by inserting small masses of it into fresh manure. One of the
biggest problems of using blocks of manure for spawning was that they usually
contained a host of other microorganisms that could interfere with the development of
the mycelium and fruiting bodies of A. bisporus.
Later developments in mushroom growing included indoor cultivation and the use of a
pure culture spawn, which was produced from germinating spores and contained only
the living mycelium of the desired mushroom species. The use of pure culture spawns
extended quickly throughout Europe and the USA (Chang and Miles, 2004).
5
Chapter 2. Literature review
2.1.2 The Fungi kingdom
Fungi were historically ranked with plants. However, they are sufficiently and
significantly distinct to place them in an autonomous kingdom: Myceteae. The present
ranking is justified by an assemblage of specific and non-specific natural characteristics.
Characteristics specific to fungi are the presence of special chemicals which are rare
elsewhere, the diversity and complexity of their life cycles, nutrition by absorption and
the possibility of growth dependent entirely upon asexual reproduction. Fungi have
ultrastructural similarities with the plant kingdom, especially at the cellular level. Both
plants and fungi possess a cell wall, although there are differences in their composition.
The mode of nutrition of fungi is heterotrophic in respect of carbon (i.e. they lack the
ability to form organic compounds for their cellular growth from carbon dioxide in the
atmosphere and need to find it in their environment), as is the case in the animal
kingdom, but unlike animals, it is absorptive rather than digestive (Courtecuisse and
Duhem, 1995).
A mushroom is a macrofungus with a distinctive fruiting body which can be either above
or under ground. This broad definition of mushrooms includes edible and non-edible,
poisonous and also medicinal species (Chang and Miles, 2004).
In what follows the term mushroom will refer to A. bisporus fruiting body, unless
otherwise specified.
2.1.3 Morphology of mushrooms
The fruiting body of A. bisporus mushrooms has three differentiated parts, which are
represented in Figure 2.1:
6
Chapter 2. Literature review
•
The cap is fleshy and hemispherical and protects the reproductive tissues. Its
colour ranges from white to cream at first, and becomes brown with age and
damage.
•
The gills are situated underneath the cap and are responsible for spore
production. They comprise a series of radial lamellae whose colour is pink at first
and becomes brown as spores mature. A veil covers the gills in the early stage of
development and then breaks as part of the natural senescence process, leaving
the gills exposed.
•
The stipe is cylindrical and lifts the cap above the ground so that spores can be
released. The stalk is the link between the aerial fruiting body and the
underground vegetative structure of the mushroom (Courtecuisse and Duhem,
1995).
Figure 2.1 Diagram of mushroom tissues.
The vegetative structure of the mushroom is called the mycelium, which comprises
microscopic tubular structures called hyphae. These filaments grow only at the tip or at
specialised regions and form a system of branching threads and cordlike strands that
branch out throughout the soil, compost or other material on which the mushroom is
growing. After a period of growth, and under favourable conditions, the established
(matured) mycelium produces the fruiting body (Chang and Miles, 2004).
7
Chapter 2. Literature review
2.1.4 Mushroom growing
Good mushroom substrate (i.e. compost) and the right environmental conditions are the
two basic requirements for mushroom growing (MacCanna, 1984). As compost quality is
largely outside the control of mushroom growers, their main contribution to final product
quality is crop management. This task involves controlling temperature, relative
humidity, watering, ventilation and CO2 levels (Teagasc, 1994). Modern mushroom
houses are equipped with computerised environmental control systems for this purpose.
The steps involved in the mushroom growing process are:
•
Compost preparation: The objective of composting is to convert the raw materials
into a highly selective and nutritious medium for the growth of mushrooms
(MacCanna, 1984). Many agricultural by-products are used to make mushroom
substrate. Straw-bedded manure and hay or wheat straw are the common bulk
ingredient (Beyer, 2003). Conversion is achieved by the natural microorganisms
that inhabit these materials and which multiply in great numbers to attack them.
As the sole source of nutrients for the mushrooms, the quality of the compost
affects all aspects of mushroom production including yield, crop quality, crop
timing and disease susceptibility (MacCanna, 1984). Generally, the substrate is
prepared by specialised companies that supply the growers with ready-to-use
compost. This process can take as little as 10 days.
•
Spawn running: a pure culture of mushroom mycelium is added to the compost
and allowed to colonise it. The spawn produces heat so temperature becomes
an important factor to control during this period. High relative humidity and high
CO2 levels are important conditions too (Teagasc, 1994).
•
Casing: mushroom casing is a layer of organic material (usually neutralised peat)
which is applied to the surface of the spawn-run compost (Teagasc, 1994). The
function of a casing layer is to trigger the mushrooms to switch from vegetative
8
Chapter 2. Literature review
growth to reproductive or fruiting growth (Beyer, 2003). Temperature, relative
humidity, CO2 and watering must be controlled between the precise day of casing
and crop initiation (Teagasc, 1994).
•
Breaking: a major change in the growing conditions is needed to initiate the
development of mushrooms. When the mycelium is clearly visible in the crevices
of the casing, changes in air conditions, such as introducing fresh air, decreasing
the temperature, reducing the CO2 level and maintaining the relative humidity,
will cause breaking and subsequent pinning. A few days after the beginning of
the breaking, the pins will be visible as white clusters, which are formed by the
fusion of mycelial strands (Teagasc, 1994).
•
Harvest: the first flush of mushrooms occurs 2-3 weeks after casing and lasts for
3-5 days. The crop develops in four flushes in weekly cycles. The first two
flushes usually provide 70% of the total yield. Harvesting must occur when the
cap is at its maximum size before the veil has stretched and opened, exposing
the gills. Care must be taken not to remove excessive casing, which would
remove the pinheads required to form later flushes.
When the mushroom cycle is complete, the spent compost is killed off using steam. All
spent casing, compost and mushroom waste must be removed to reduce the chances of
contaminating the subsequent crops (Beyer, 2003).
There are two systems of growing crops in mushroom houses:
i) Single-layer bag-growing in tunnels (Figure 2.2.a). It is a simple and effective
method where specialised companies make and supply the mushroom producers
with ready-to-use compost bags that contain mushroom spawn mixed through it.
The large volume of air over the cropping surface and the deep layer of compost
in each plastic bag favour the growth of good quality mushrooms. The easy
disposal of cropping remains allows for efficient environmental control and better
9
Chapter 2. Literature review
hygiene practices. One of the disadvantages of the bag system is that it requires
a high manual labour input (Teagasc, 1994).
ii) Multi-layer systems on shelves (Figure 2.2.b), which can double, triple or
quadruple the output of a tunnel (Grogan, 2008a). This method allows
mechanised compost filling/emptying and facilitates the use of automated
harvesting equipment, which results in a reduction in labour input.
(a)
(b)
Figure 2.2 Mushroom growing systems.
(a) Single-layer bag-growing. (b) Multi-layer metal structure for growing in shelves.
The last official figures from the March 2006 Census showed that 34% (i.e. 44 out of
129) of Irish mushroom farms were shelf farms (Teagasc, 2006). Although there is no
later official data, mushroom experts have knowledge of an increase in the number of
shelf-based farms (Grogan, 2008b).
2.1.5 Irish mushroom industry
The development of the mushroom industry has been a big success for Irish horticulture
(Teagasc, 1994; Grogan, 2008a). From small beginnings, the industry has grown to the
stage where Ireland is now a major producer of fresh mushrooms and the third largest
fresh A. bisporus exporter in Europe, after Poland and the Netherlands (Van Horen,
10
Chapter 2. Literature review
2008). Over 45,000 tonnes of mushrooms, representing 80% of total mushroom
production in the country, are exported annually to the UK markets (Teagasc, 2007).
In the 1980s the Irish mushroom industry was the first to try out a new concept of
growing called the “satellite” system. It consisted of a small number of centralised
composting facilities that supplied compost in bags to many small and medium-sized
growers. The so-called “Phase 3” compost was spawn-run compost and was ready for
casing when delivered to the grower. The use of this substrate increased crop yield and
reduced overall cycle length. Mushrooms were then returned to a central depot and
marketed collectively, resulting in a very efficient production and marketing system
(Grogan, 2008a).
Between 2000 and 2004, Ireland was ranked third in the world for fresh-mushroom
exports, after China and the Netherlands (Grogan, 2008a). However, there has been a
dramatic increase in mushroom production in Poland in recent years, and it is now the
biggest producer and exporter in Europe (Van Horen, 2008).
The competition between Poland and the Netherlands has had a knock-on effect on
established mushroom industries throughout Europe, where the industry is consolidating
into larger, more productive farms with increased mechanisation and centralised
management. The numbers of farms has declined in most European mushroomproducing countries, including Ireland. There has been a decrease of 83% in the number
of Irish mushroom growers since 2000. Many small inefficient farms have closed and the
units remaining have generally increased in size. The farms continuing in production
have had to increase output and efficiency (Grogan, 2008a).
The only way to remain profitable is by producing quality mushrooms more efficiently. In
Ireland, this has been achieved by increasing production and improving quality. There
has been a major change from the small scale, labour-intensive system of growing in
bags in a single-layer to larger scale, multi-layer systems on shelves. The use of Phase
11
Chapter 2. Literature review
3 compost has increased yields and cut down cycles, which has increased the annual
production per tunnel.
The Irish industry needs to continue producing high quality mushrooms that will allow it
to compete in the export market. To continue to successfully compete for a significant
share of the European fresh mushroom market, Irish growers will also have to face
several challenges in the years ahead, such as increased quality and environmental
regulation, shortage of skilled labour, reduced availability of pesticides (i.e. disease
control with few or no chemicals) and increased costs of production (Grogan, 2008a).
Despite the difficulties of the past few years, growers are investing in improving the
growing system, which will allow the Irish industry to maintain a strong position in the
European export market place (Grogan, 2008a).
2.2 Key issues for the mushroom industry
2.2.1 Mushroom shelf-life
Whole mushrooms have a short shelf-life compared to most vegetables. Due to their thin
and porous cap cuticle, the respiration rate and water loss of mushrooms are relatively
high compared to other vegetables and fruits (Kader, 1985). Maturation and senescence
are phenomena which, together with microbial activity, contribute to natural postharvest
deterioration of fresh mushrooms (Beelman, 1987).
Mushrooms are very perishable products. Usually, their shelf-life is 1 to 3 days at
ambient temperature and 5 to 10 days under refrigeration conditions (Gormley, 1975;
Burton, 1989; Lee, 1999). Extending the shelf-life is desirable because a short life
restricts export potential. A few extra days of shelf-life would compensate for time in
12
Chapter 2. Literature review
transit and give more flexibility to processors, retailers and consumers (Brennan et al.,
2000).
Recent research in the area of mushroom shelf-life has focused on developing
postharvest technologies and improving packaging design for shelf-life extension.
Compost supplementation during cultivation and various chemical and physical
treatments during postharvest have been tested for extending the shelf-life of
mushrooms. Packaging, coating, refrigeration and dipping in sorbitol are the most
common methods used for extending the shelf-life of mushrooms (Roy et al., 1993;
Eissa, 2007; Aguirre, 2008). Appropriate packaging can delay development of
deterioration and senescence of mushrooms after harvest. Some researchers have
investigated the effect of modified atmosphere packaging (MAP) on fresh mushrooms
(Roy et al., 1995; Barron et al., 2002; Kim et al., 2006; Mahajan et al., 2008). Although
different films have been utilised, polyvinyl chloride (PVC) is still the most commonly
employed commercial packaging film for mushrooms. A recent study by Taghizadeh et
al. (2010) showed that perforated polyethylene terephthalate (PET) film is a viable
alternative to the conventional polyvinyl chloride (PVC) film, facilitating an increase in
mushroom shelf-life.
Although postharvest treatments delay mushroom senescence and extend shelf-life, the
product quality during storage is very much dependent on the quality of the harvested
mushroom (Flegg et al., 1985).
2.2.2 Variability
One of the main problems in mushroom technology (as with many fresh products) is the
uncontrollable effect that product variability has on the management of the product
(Aguirre
et
al.,
2008c).
Quality
control
13
during
postharvest
requires
precise
Chapter 2. Literature review
methodologies to assess the acceptability of fresh produce of varying batches, growers,
cultivation practices and postharvest treatments (Mohapatra et al., 2010).
A batch may be defined as all individuals sharing the same harvest date, grower and
cultivar (Schouten and Van Kooten, 1998). If all individuals in a batch were identical and
stored under the same conditions, they should all reach the quality limit at the same time
(Schouten et al., 2004). However, this is not the norm, as different units of an individual
batch tend to behave differently. Biological variation can be described as the composite
of biological properties that differentiate individual units of a batch (adapted from
Tijskens and Konopacki, 2003. Due to these differences between individuals, batches
behave differently, depending on the magnitude of variation in the product and on the
magnitude on the batch (Tijskens et al., 2003).
During growth, small local differences in growing conditions integrate into a considerable
variation in maturity at the moment of harvest (Tijskens et al., 2008). Further variation in
postharvest storage behaviour can be interpreted as the expression of the same generic
product behaviour, only the choice of time zero from which moment the produce is being
observed is randomly determined by the moment of harvest. As fruits and vegetables
are not harvested at such a homogeneous stage, variation will exist both at harvest and
during postharvest storage (Hertog et al., 2004).
The management of variability remains one of the greatest challenges for food
technologists. What differentiates the postharvest chain from most other handling chains
is i) the fact that postharvest deals with living tissue that is inherently changing with time
and ii) the large sources of biological variance. Postharvest product behaviour is
inherently affected by omnipresent biological variation. Postharvest management aims
at controlling the average batch behaviour and limiting biological variation as much as
possible. Sorting and grading the product at the different stages in the postharvest chain
are essential tasks for this purpose (Hertog et al., 2007).
14
Chapter 2. Literature review
Currently, the horticultural industry is not capable of differentiating its postharvest
treatments to such an extent that it can actually take into account batch specific handling
requirements. Much more insight is required into the dynamics of quality change in
general and the propagation of biological variance through the postharvest chain.
Identification and quantification of the different sources of variance is of utmost
importance to compare batches and to predict how an initial variance in quality attributes
propagates over time (Hertog et al., 2007).
Biological variance has also gained interest within the postharvest community. Data
generated in postharvest research are characterised by a large amount of biological
variance, which obscures the behaviour of interest, complicating the interpretation of the
data. Traditionally, postharvest scientists have tended to ignore the behaviour of the
individual objects and have focused on the averaged batch behaviour based on large
sample sizes. Only recently, awareness of the added value of reporting data from
individual objects has increased and new impetus has been given to include biological
variance into postharvest data analysis. Several (statistical) modelling techniques have
been applied in order to account for biological variance. These vary from models that
can be solved analytically to analyse the effect of experimental design parameters, via
hybrid techniques which combine analytical and numerical techniques, to pure numerical
models for dynamic simulations based on ordinary differential equations and/or partial
differential equations (Nicolaï et al., 2001; 2006b).
Aguirre et al. (2008b; 2008c) and Mohapatra et al. (2008) tackled the mushroom product
variability issue by incorporating a mixed-effects model approach into their modelling
strategy. Mixed effect models present a statistical framework that allows for a
simultaneous characterisation of the main effects that are going to influence an
experiment, together with the estimation of the different components that affect the
variability intrinsic to the problem being observed (Pinheiro and Bates, 2000). Therefore,
15
Chapter 2. Literature review
mixed-effects models may be useful for those cases where one has to deal with withinsubject as well as between-subject variability, such as harvested mushrooms.
2.3 Mushroom quality
Quality is a difficult term to define. It can be defined as the extent to which a product
meets the consumers’ expectations (Hewett, 2006). ElMasry et al. (2007) defined the
quality of a fresh product as a measure of the characteristics or attributes that determine
the suitability of the product to be eaten as fresh or stored for reasonable periods without
deterioration. Taking these two definitions into consideration, the quality of a horticultural
product could be considered to be a multi-concept (encompassing physical,
physiological, nutritional, and pathological attributes) that needs to meet several
consumers’ expectations.
Quality in mushrooms can be assessed as a combination of visual appearance,
freshness, colour, size, maturity stage, development stage, firmness, turgor, microbial
growth, clarity, lack of blemishes/blotches and weight loss (Burton, 1989; Carey and
O'Connor, 1991; Vízhányó and Felföldi, 2000). Consumers base their judgement on the
following factors: freshness, colour, size, texture, cleanliness, maturity and flavour
(Burton, 1986; 2004). High quality mushrooms are defined as fresh, white, unblemished,
firm, clean and with a closed veil (Berendse, 1984; Eastwood and Burton, 2002).
Mushroom quality determining attributes may be influenced by pre- and postharvest
factors. Pre-harvest aspects include strain, composting, casing, watering and
temperature regime, atmosphere composition and flush number. Postharvest factors
include picking, chilling, packaging and transportation. During this time, mushroom
quality becomes dependent on environmental factors such as temperature, relative
humidity and the gaseous composition of the storage atmosphere (Aguirre et al.,
2008a).The importance of systems for the commercial cooling of mushrooms, such as
16
Chapter 2. Literature review
vacuum cooling, has been reported in the past (Gormley, 1975; 1991). Vacuum cooling,
achieved by evaporating part of the moisture of mushrooms under vacuum conditions,
has been adopted commercially in Ireland and other parts of Europe to cool mushrooms
uniformly (Sun and Zheng, 2006).
The main processes which contribute to loss in quality after harvest are discolouration,
loss of closedness, weight loss and changes in the texture of mushrooms (Berendse,
1984; Burton and Noble, 1993). Loss of quality happens as a result of biochemical and
microbial degradative processes and improper handling during harvest, packaging and
transport. In the latter case, loss of quality is largely the result of bruising, initiated by
mechanical damage leading to the enzymatic oxidation of natural phenols.
Within all the parameters, the colour of the cap is the most important quality indicator of
fresh mushrooms, since it is the first characteristic consumers notice. Any brown
patches on white mushrooms are easily seen and viewed by consumers as an indicator
of low quality (Burton, 2004).
2.3.1 Mushroom colour as a quality parameter
Colour plays a major role in the assessment of the external quality of foods. It is
considered a fundamental physical property of most agricultural products and foods
(Mendoza et al., 2006). The agricultural industry uses measurements of colour mainly
for three reasons: i) colour serves as instant indicator of product quality, since it has
been widely demonstrated that it correlates well with other physical, chemical and
sensorial indicators of product quality, ii) colour measurements have been employed to
develop optimal storage policies with the aim to maintain appealing appearance through
the shelf-life and iii) the most important use of colour is to provide a buying criterion in
the purchase of products (Abdullah et al., 2004).
17
Chapter 2. Literature review
Colour is basically specified by the geometry and spectral distributions of three
elements: the light source, the reflectivity of the sample and the visual sensitivity of the
observer.
Human vision has three sets of sensors with peak sensitivities at light frequencies of 580
nm for red, 540 nm for green and 450 nm for blue. While a grayscale image typically
reflects the light intensity over a single band in the electromagnetic spectrum, a colour
image reflects the intensity over the red, green and blue bands of the spectrum. Light at
any wavelength in the visual spectral range from 400 to 700 nm will excite one or more
of these sensors. Our perception of which colour we are seeing is determined by which
combination of sensors are excited by how much.
In food research, colour is frequently represented using L*a*b* colour space, which was
specified by the Commission Internationale de l’Éclairage (CIE) in 1976. This colour
model (also known as CIELAB) permits an objective colour representation and its use is
essential for applications where the results must match those of human perception
(Sangwine, 2000). L* is the lightness component that goes from 0 (black) to 100 (white),
and parameters a* (from green to red) and b* (from blue to yellow) are the two chromatic
components, varying from -120 to +120, as shown in Figure 2.3.
Figure 2.3 CIELAB diagram.
18
Chapter 2. Literature review
Conventional colour specification instruments, such as colorimeters, can provide
readings in L*a*b* colour space. Although these instruments are generally suitable for
use with food materials, they are inherently unsuitable for assessing the colour
distribution and colour uniformity measurements of many whole foods (Mendoza et al.,
2006).
Alternatively, computer vision (CV) systems have been developed in response to the
lack of spatial resolution of the conventional instrumental techniques. These systems not
only offer a methodology for measurement of uneven colouration, but they can also be
applied to the measurement of other attributes of total appearance (Hutchings, 1999).
CV systems acquire images by means of digital or video cameras, which are powerful
enough for colour studies.
A digital colour image is represented in RGB form with three components (red, R;
green, G; blue, B) per pixel in the range of 0-255 and conventionally stored using eight
bits per colour component. These three intensity images are electronically combined to
produce a digital colour picture (Russ, 2005). Calibrated average colour measurements
using computer vision techniques have proven to closely correlate with those from visual
assessment (Mendoza and Aguilera, 2004).
Many image analysis-based quality assessment methods use grayscale imaging for
produce grading (Vízhányó and Felföldi, 2000). Grayscale values can be obtained from
RGB colour pictures by applying transformation algorithms (Aguirre et al., 2008b).
2.4 Mushroom discolouration
A healthy crop of growing mushrooms is expected to be white in colour. Loss of
whiteness during storage results in large economic losses every year and is a
particularly important issue for the mushroom industry (Burton, 1988). Considerable
19
Chapter 2. Literature review
research effort is required to increase understanding of the physical and physiological
factors involved in mushroom discolouration. Bruise prevention is necessary but only
possible when the factors responsible for bruise development are known.
There are several factors that induce mushroom discolouration:
i) Healthy mushrooms lose their colour due to two main reasons: a) natural
deterioration/senescence processes and b) mechanical damage leading to
bruising.
ii) Mushrooms can also develop brown colouration when subjected to pathogen
attack, which can be of either a) microbial or b) viral origin.
2.4.1 Why brown?
The brown colour is the result of melanins, i.e. chemicals made via a series of
biochemical and chemical reactions involving phenolic compounds and enzymes, which
are all naturally present in mushrooms (Jolivet et al., 1998).
The structure of phenol (C6H5OH) is that of a hydroxyl group (-OH) bonded to a phenyl
ring (-C6H5), as shown in Figure 2.4.
HO
Figure 2.4 Structure of phenol.
Polyphenolics are a multiplicity of different phenolic compounds, that is, compounds
composed of aromatic benzene rings(s) substituted with hydroxyl groups, including all
functional derivatives. The main natural melanogenous phenols present in A. bisporus
mushrooms are shown in Figure 2.5.
20
Chapter 2. Literature review
O
HN
O2
H2N
O
OH
glutaminyl-4-hydroxybenzene
HO
NH2
H2N
p-aminophenol
HO
NH2
O
OH
phenylalanine
OH
tyrosine
Figure 2.5 Natural melanogenous phenols present in A. bisporus.
Development of browning is the consequence of a series of complex reactions where
polyphenol oxidase (PPO) and peroxidase (POD) enzymes play an important role
(Jolivet et al., 1998; Bandyopadhyay et al., 2009).
The first two steps in the multipart browning pathway are:
i) Hydroxylation of monophenols into o-diphenols (known as monophenolase or
cresolase activity), as shown in Figure 2.6.
HO
R
R
O2
HO
HO
AH2
monophenol
A
o-diphenol
Figure 2.6 Hydroxylation of a monophenol (Cresolase activity).
ii) Oxidation of o-diphenols into corresponding o-quinones (known as diphenolase
or catecholase activity), as shown in Figure 2.7.
O
HO
R
R
O2
O
HO
o-diphenol
AH2
A
o-quinone
Figure 2.7 Oxidation of an o-diphenol (Catecholase activity).
Both oxidation steps happen at the expense of molecular oxygen. Quinones, which are
slightly coloured, undergo further reactions (i.e. oxidations and polymerisations), leading
to high molecular mass dark pigments called melanins (e.g. Figure 2.8).
21
Chapter 2. Literature review
O
O
NH
*
*
HN
n
Figure 2.8 Structure of a melanin.
PPOs are also knows as tyrosinases. In what follows, the terms PPO and tyrosinase will
be used interchangeably.
The pathways and mechanisms of enzymatic oxidation of phenols are complex. Two
non-exclusive explanations have been proposed for this multipart process:
i) The breakdown of intracellular membranes allows melanogenous substrates to
come into contact with tyrosinases. The location of fungal PPOs is not entirely
clear but they generally appear to be cytoplasmatic enzymes (Mayer, 2006). At
early stages of development, phenols and enzymes are separated by
intracellular membranes. A disruption of membranes within the cytoplasm results
in the mixing of phenols and tyrosinases, giving rise to a brown pigment. This
disruption could result from an internal disorganisation caused by environmental
stress, such as bruises and injuries derived from external damage to
mushrooms.
ii) The activation of inactive (latent) enzymes. Mushroom tyrosinases are known to
occur largely in a latent or inactive state. The mechanism of fungal tyrosinase
activation remains unclear (Mayer, 2006), although it is suggested that it is the
consequence of a conformational enzymatic change or a protein solubilisation
(Halaouli et al., 2006). The amount of mushroom enzymes present in the latent
state is still a matter of debate, ranging from less than 1% (Flurkey and
Ingebrightsen, 1989) to 30-40% (Burton, 1988). Proteinases have been shown to
22
Chapter 2. Literature review
fluctuate during the flushing pattern of mushroom cropping and a very large
increase in the level of a particular proteinase, serine proteinase, has been
observed in mushrooms after harvest (Burton et al., 1997). Serine proteinases
have been shown to activate tyrosinases in vitro (Espin et al., 1999). Therefore, it
is possible that the increase in serine proteinase activity leads to an activation of
tyrosinases and causes the increase in tissue browning after harvest.
Tolaasin, a toxin produced by Pseudomonas tolaasii (P. tolaasii) during infection
of mushrooms is also thought to activate tyrosinases. Studies by Soler-Rivas et
al. showed that tolaasin is involved in the tyrosinase activation process (SolerRivas et al., 1997; 2001).
Alternatively, the browning reaction in A. bisporus is perhaps a response to pathogen
infection. The browning reaction catalysed by tyrosinase may be involved in a defence
mechanism against bacterial infections (Soler-Rivas et al., 1997).
If a measurable proportion of tyrosinase is in the active form, then its activity is in great
excess relative to the phenolic compound levels and therefore the phenol levels would
be the limiting factor controlling the extent of browning (Burton et al., 1993). Jolivet et al.
(1995) compared tyrosinase and phenolic compound levels in two white A. bisporus
strains with different abilities to brown. The content of phenolic compounds in the
mushroom pileipellis (i.e. cap skin) was nearly sevenfold higher in the more sensitive
strain, whereas no significant difference in enzyme activity (neither active nor latent) was
observed between the two strains. These authors concluded that the phenol content
present in the upper part of the mushroom was the limiting factor of the natural browning
process occurring during storage and could largely explain the different susceptibilities
of the strains.
23
Chapter 2. Literature review
2.4.2 Causes of discolouration in healthy mushrooms
2.4.2.1 Senescence
When the mushroom is harvested, it is cut off from its source of food and water. The
mushroom responds quickly to harvesting, switching on stress genes as a response to
the wound damage. The nutrients and water within the fungi are mobilised and redistributed; the respiration rate is increased under water stress and nutritional limitations.
The cell reserves of carbohydrates and the protein content of the mushroom fall rapidly.
Different tissues respond differently. The stipe is used as a nutrient sink to fuel much of
the continued development in the cap and gills. A gene involved in cell wall
polysaccharide formation is believed to drive cap expansion, after which the mushroom
flattens and the protective veil breaks. The mushroom ages further and appears darker
and less firm. Under these conditions, the cellular structures start to break down. This
natural oxidative process is known as senescence (Eastwood and Burton, 2002; Burton,
2004) and involves degradation of the cellular and sub-cellular structures and
macromolecules, as well as mobilisation of the products of the degradation to other parts
of the fungi.
Tissue browning increases during post-harvest storage (Burton et al., 1987b). Cooling
reduces nutrient use and movement and it is a good way of minimising post-harvest
discolouration (Burton et al., 1987a). Modified atmosphere packaging, which involves
wrapping the mushrooms in semi-permeable plastic films, can reduce the availability of
oxygen and delay browning (Burton et al., 1987b).
Since Gormley and O’Sullivan (1975) reported the relationship between different quality
levels in A. bisporus mushrooms and Hunter L-value, colorimeters have been widely
used to assess the effects of senescence on the colour of mushrooms (Gormley, 1987;
Brennan and Gormley, 1998; Kim et al., 2006; Mohapatra et al., 2010). Recently, Aguirre
et al. (2008b) monitored the browning and brown spotting due to senescence with
24
Chapter 2. Literature review
inexpensive webcams. The RGB space of the images taken by these devices was
converted into an 8-bit grayscale space; average grayscale value was employed as a
measurement of the loss of whiteness and its standard deviation was used as a contrast
measurement to follow the appearance of brown spotting.
2.4.2.2 Mechanical damage
For healthy mushrooms, the second cause of browning is mechanical damage.
Mushrooms bruise very easily; a small amount of mechanical force can cause an
indentation in the mushroom surface which rapidly develops a dark brown discolouration
(Burton, 2004). During cultivation, some mushrooms may grow too close to each other
and come into contact, developing signs of bruising. Small insects can also damage the
cap tissue. Further damage to mushrooms is caused by human handling, through a
process known as “slip-shear”, which comprises a downwards force and sideways
movement (Burton et al., 2002). This occurs when the hand slips over the mushrooms
during picking. After harvest, bruising can result from mushrooms being rubbed against
each other and against other surfaces during packaging and transportation practices
(Burton et al., 2002; Burton, 2004). Mushrooms must be handled with extreme care at all
times if bruising is to be avoided.
Mushroom bruisability can vary from crop to crop and even within a crop. Sensitivity to
mechanical damage is not only determined by genetics but also by mushroom growing
conditions, such as agronomic and environmental (Burton, 2004). The strain and flush
number can have a major effect on bruisability and making the right choice is critical for
minimising detrimental effects. The major agronomic and environmental factors
influencing mushroom sensitivity to bruising include casing water content, humidity and
to a lesser extent calcium chloride irrigation. To reduce bruising, mushrooms should be
grown wet and allowed to dry out somewhat towards the end of the growth period and
prior to cropping (Burton et al., 2002). Calcium chloride irrigation (0.3-0.5%) can
25
Chapter 2. Literature review
effectively reduce the problem when bruising becomes a frequent problem on a farm
(Burton and Beecher, 2000). The compost composition has a minor effect (Burton,
2004).
Picking, handling, packaging and transport must be performed with extreme care in
order to avoid additional damage to the mushroom surface (Gormley, 1987). If the
mushroom does not detach easily from the soil during picking, it can be injured by the
force of the hand. This is particularly evident when two or three fingertips are used to
pull, so growers often use a whole hand grip (Knee and Miller, 2002). Mushrooms
should be picked gently, as dry as possible, into the tray from which they are going to be
marketed. Appropriate selection of the packaging technique and materials (tray and film)
are of considerable importance in minimising quality loss. An efficient distribution system
where mushrooms are transported under refrigeration conditions and where vibrations
are minimised is essential.
In past years, several studies investigated the ability of CV systems and various
spectroscopic techniques to detect bruises in agricultural products such as apples (Xing
and De Baerdemaeker, 2007; Xing et al., 2007; ElMasry et al., 2008), potatoes (Thybo
et al., 2004), cucumbers (Ariana et al., 2006), peaches and apricots (Zwiggelaar et al.,
1996).
Recent work by Gowen et al. (2008b) investigated the potential of a hyperspectral
imaging system for the detection of bruise damage on the mushroom surface. In this
study, mushrooms were damaged by controlled vibration to simulate damage caused by
transportation. The authors concluded that HSI combined with principal component
analysis was useful for this purpose, with correct classification ranging from 79% to
100%. Esquerre et al. (2009) used visible near-infrared (Vis-NIR) spectroscopy to detect
the chemical changes induced by controlled vibrational damage and modelled them
using multivariate data analysis. The authors were able to correctly classify undamaged
26
Chapter 2. Literature review
mushrooms from damaged ones with 99% accuracy. O’Gorman et al. (2010)
investigated the combined use of Fourier Transform Infrared Spectroscopy (FTIR) and
chemometrics to identify mushrooms subjected to mechanical damage. The authors
developed models that correctly classified mushrooms with >90% accuracy.
2.4.3 Causes of discolouration in diseased mushrooms
2.4.3.1 Microbial spoilage
Brown blotch or bacterial blotch disease is the most serious bacterial disease of the
cultivated white button mushrooms (Fermor et al., 1991), because of the large economic
losses associated with it. Other microbiologically induced diseases that cause
mushroom spotting include ginger botch, wet bubble, dry bubble, cob web,
Aphanochadium cap spotting, Hormiactics cap spot, Gliocadium diseases and
Trichoderma diseases (Fletcher and Gaze, 2008). Although bacterial spoilage does not
cause any health problems, it can result in a loss in consumer appeal in the market
place (Soler-Rivas et al., 1999).
Brown blotch is a bacterial disease caused by Pseudomonas tolaasii (P. tolaasii) (Paine,
1919). This bacterium was officially positioned in the taxonomic group of fluorescent
Pseudomonas biotype II. More recent studies have disagreed with this placement and its
position in the systematic classification is still unclear due to the high heterogeneity of
the group (Rainey et al., 1992; Wells et al., 1996).
Cutri et al. (1984) reported the existence of two variants of P. tolaasii for mushrooms: a
non-pathogenic and a pathogenic. Although the non-pathogenic form might present
different colonies, the pathogenic (also known as smooth form) presents small, semimucoid, opaque and non-fluorescent colonies in Pseudomonas agar F (Zarkower et al.,
1984). As a pathogen, P. tolaasii is not host specific, as it can infect almost all species of
mushrooms and also some vegetable and fruit plants. P. tolaasii can be identified based
27
Chapter 2. Literature review
on the Mushroom tissue block rapid pitting and White Line in Agar pathogenicity tests
(Wong and Preece, 1979).
Colonisation of mushroom caps by the bacterium results in brown discolouration,
appearing as brown/creamy lesions on the damaged area (pileus and stipe). The
symptoms (Figure 2.9) are induced by tolaasiin, a toxin produced by the pathogenic form
of P. tolaasii.
Figure 2.9 A. bisporus caps showing typical symptoms of bacterial blotch.
(Soler-Rivas et al., 1999).
Tolaasin was reported to be a membrane-permeabilising compound with both ion
channel-forming and biosurfactant properties (Brodey et al., 1991; Hutchison and
Johnstone, 1993). Thus, browning symptoms, assisted in their early stages of
development by the ability of the bacteria to attach to mycelial surfaces, result from a
membrane breakdown, allowing contact between tyrosinase and phenolic substrates,
leading to oxidation. Moreover, a study by Soler-Rivas et al. (1997) showed that tolaasin
is also involved in the tyrosinase activation process. After contact with bacteria or toxin
extract, the amount of active tyrosinase measured in the infected tissue increased in the
sensitive strains and remained unchanged in the resistant one. The authors suggested
the occurrence of a non-proteolytic factor in the toxin preparation was directly
responsible for the activation of the enzyme. Partially purified or purified tolaasin extract
induced a specific messenger RNA encoding tyrosinase when applied to A. bisporus
and activated tyrosinase (Soler-Rivas et al., 2001).
28
Chapter 2. Literature review
Many investigations have been carried out to find an adequate method to prevent or
control the disease. Manipulation of the environmental conditions (i.e. relative humidity,
temperature, ventilation, disinfection, etc.) has proved essential but not sufficient (Nair,
1974; Rainey et al., 1992). Calcium chloride and chlorinated compounds are currently
the most utilised chemicals for brown blotch disease (Royse and Wuest, 1980; Geels et
al., 1991; Solomon et al., 1991). The potential of several other chemicals, such as
disinfectants and antibiotics, was studied but none met the requirements of efficacy and
non-toxicity (Wong and Preece, 1985). Nowadays, other procedures for control are
being developed, such as selection and production of resistant strains (Soler-Rivas et
al., 1999). Advances in strain hybridisation, molecular biology and techniques for
production of transgenic mushrooms could contribute to the solution by developing
disease resistant A. bisporus strains.
Recent studies in the field of brown blotch detection include the work of Vízhányó and
Felföldi (2000), who tested the potential of a machine vision system to recognise and
identify brown blotch and ginger blotch diseases, both of which cause discolouration of
mushroom caps. Results from this study will be discussed in Chapter 7.
2.4.3.2 Viral infection
Mushroom virus diseases have been detected in many mushroom-growing countries all
over the world (Elibuyuk and Bostan, 2010). La France virus was the first viral pathogen
which affected mushroom cultivates in the USA and UK in the 1960s. Sinden and
Houser (1950) for the first time noticed the symptoms of this very infectious disease of
A. bisporus mushrooms leading to large scale losses in yield and brown and malformed
crops. Epidemiological research indicated that the disease was predominately spread
through spores from infected mushrooms and also by infected mycelium carried over
from one crop to the next (Schisler et al., 1967).
29
Chapter 2. Literature review
More recently, in 1996, a virus-like disease occurred on several mushroom farms in
Britain, which soon became more widespread (Grogan et al., 2004). A correlation was
made between the presence of novel double-stranded RNA and the occurrence of the
virus-like symptoms on the British mushrooms farms (Grogan et al., 2003), and the
disease was called Mushroom Virus X (MVX). MVX is an enigmatic disease which
causes crop delay, pinning disruption, poor quality and occasionally brown or offcoloured mushrooms (Figure 2.10). This browning is restricted to the surface of the cap.
Figure 2.10 A. bisporus caps showing symptoms of MVX disease.
(Aguirre, 2008).
It is proving difficult and time consuming to characterise the exact aetiology of MVXassociated cropping problems. MVX is unlike La France virus in that La France is most
problematical when virus infected spores or mycelium infect the crop at spawning and is
less problematical if infection takes place later. In contrast, epidemiological research
indicates that MVX infected mycelium can cause severe effects when introduced at
either spawning or casing. Further experiments indicate that the expression of the brown
mushroom symptom occurs within a narrow window of opportunity, as only when
infection takes place at the casing time are the brown symptomatic mushrooms
produced consistently. This suggests a complex interaction between the infective agent
and the actively differentiating tissue within the mushroom cap.
The absence of symptomatic mushrooms in an “infected” crop makes the ultimate
control of the problem more difficult, as growers will be unaware that there is a potential
30
Chapter 2. Literature review
problem on the farm. Improvements in farm hygiene aimed at preventing the
contamination of crops with potentially infected spores and mycelial debris at spawning
and casing are essential procedures to bring MVX disease under control. MVX was
brought under control in Britain through improvements in farm hygiene aimed at
preventing contamination. In Ireland, however, the brown mushroom symptom still
persists and the transient nature of its expression indicates that there is still a reservoir
of infected material within the industry (Grogan, 2008c).
2.5 Computer imaging for product quality measurement
2.5.1 The necessity for automating quality assessment
The assurance of quality is one of the most important goals of any industry. With
increased expectations for food products of high quality and safety, the need for
accurate, fast and objective methods for quality determination continues to grow. The
quality assurance methods used in the food industry have traditionally involved human
visual inspection. Such methods are tedious, laborious, time-consuming, inconsistent
and costly. In general, automation of a quality assessment operation not only optimises
quality assurance but, more importantly, it also helps in removing human subjectivity and
inconsistency. Moreover, automation usually increases productivity. If quality evaluation
is achieved automatically, production speed and efficiency can be improved drastically in
addition to increased evaluation accuracy, with an accompanying reduction in production
costs (Elmasry and Sun, 2010).
Measuring the physical properties of fruit and foodstuffs has always been of crucial and
increasing concern in the food industry. Traditionally, food engineers and technologists
have approached food property measurements in two ways: sensory and objective
evaluation. The development of instrumentation technologies means that some of the
31
Chapter 2. Literature review
measurements which were previously measurable via human senses can now be made
artificially.
2.5.2 Computer vision
Computer vision (CV) provides an automated, non-destructive and cost-effective
technique for food characteristic determination, and has found a variety of different
applications in the food industry. The technology aims to duplicate the effect of human
vision by electronically perceiving and understanding an image (Sonka et al., 1999). This
inspection approach, based on image analysis and processing, simulates artificially the
human thinking process and helps make accurate, quick and very consistent
judgements (Abdullah et al., 2004).
Since computer vision operates at visible wavelengths, it can only produce images
registering the external view of the object and not its internal view. CV is the construction
of explicit and meaningful descriptions of physical objects from images (Ballard and
Brown, 1982). Machine vision includes the capturing, processing and analysis of
images, facilitating the objective and non-destructive assessment of visual quality
characteristics in food products (Timmermans, 1998).
2.5.2.1. Components
Computer vision systems acquire images with an image sensor and use dedicated
computing hardware and software to analyse the images with the objective of performing
a predefined visual task.
The hardware configuration of computer-based machine vision systems is relatively
standard. Typically, a vision system is composed of the following basic components:
illumination device, a camera, a frame-grabber, computer hardware and software
(Abdullah, 2008), as shown in Figure 2.11 and discussed below:
32
Chapter 2. Literature review
•
Illumination: the provision of correct and high quality illumination is decisive.
Engineers and machine vision practitioners have long recognised lighting as
being an important piece of machine vision systems. However, choosing the right
lighting strategy remains a difficult problem, as there is no specific guideline for
integrating lighting into machine vision applications.
The intensity and homogeneity of the illumination system can greatly influence
the quality of the image and plays an important role in the overall efficiency and
accuracy of the system (Novini, 1995). In the vast majority of machine vision
applications, image acquisition deals with reflected light. Therefore, it is of utmost
importance to know how to control the reflection so that the image appears of a
reasonable good quality.
Traditionally, the most common illuminants have been fluorescent and
incandescent bulbs, even though other light sources (such as light-emitting
diodes (LEDs) and electroluminescent sources) are also useful. Fluorescent
bulbs are efficient and provide a more uniform dispersion of light from the
emitting surface, which means they normally do not require the use of diffusing
optics to disseminate the light source over the field of view.
The elimination of natural light effects from the image collection process is also
considered
of
importance,
with
most
modern
systems
having
built-in
compensatory circuitry (Brosnan and Sun, 2004).
•
Camera: there are many types of cameras, ranging from the older pick-up tubes
to the most recent solid-state imaging devices, such as Complementary Metal
Oxide Silicon (CMOS) cameras. The latter is the dominant technology for
cameras, and revolutionised the science of imaging with the invention of the
charge-coupled device (CCD). Both monochrome and colour cameras have been
used throughout the food industry for a variety of applications (Leemans et al.,
1998; Manickavasagan et al., 2008).
33
Chapter 2. Literature review
•
Frame-grabber: this is the component that converts pictorial images into
numerical form. In this process, called digitisation, an image is divided into a two
dimensional grid of small regions containing picture elements defined as pixels.
•
Computer hardware and software determine the image processing and image
analysis steps, which are considered to be the core of computer vision (Krutz et
al., 2000). Image processing aims to enhance the quality of an image in order to
remove defects associated with image acquisition. Image analysis is the process
of distinguishing regions of interest (ROIs) of an object from the background and
producing quantitative information.
Camera
Illumination system
Sample
Frame-grabber
Figure 2.11 Diagram of a typical computer vision system.
(Abdullah, 2008).
Owing to the imperfections of image acquisition systems, the images acquired are
subject to “defects” that will affect the subsequent processing. Therefore, it is preferable
to correct the images, after they have been acquired and digitised, by operations known
as image pre-processing steps. Following pre-processing, the images are segmented,
separating them into regions of similar image characteristics in an attempt to
automatically recognise the objects of interest in the images. After the segmentation, the
34
Chapter 2. Literature review
pixels in an image are analysed. These measurements are the core elements in a
computer vision system, because they carry the information that can be used for quality
evaluation and inspection. A great number of methods have been developed for the
acquisition of object measurements, including size, shape, colour and texture. The final
step will be to classify the objects into one of the finite sets of classes, which generally
involves comparing measured features with those of known objects.
2.5.2.2 Applications
Since its origin in the 1960s, the implementation of CV in industry has experienced
significant growth, with its applications expanding in diverse fields: medical diagnostic
imaging, factory automation, remote sensing, forensics and autonomous vehicle and
robot guidance (Brosnan and Sun, 2004). The main purpose of the application of CV
systems is to aid or enhance tasks normally performed by humans (Blasco et al., 2009).
Computer vision systems are being used increasingly in the food industry. These
systems offer the potential to automate manual grading practices thus standardising
techniques and eliminating tedious and sometimes subjective human inspection tasks.
In the last years, rapid scientific and technological advances have increasingly taken
place regarding the quality inspection, assurance, classification and evaluation of a wide
range of food and agricultural products.
•
CV has been used in bakery products where appearance is an important quality
attribute, correlating with product flavour and influencing the visual perceptions of
consumers. Computer vision has been used to evaluate the internal and external
appearance of bakery products (Sapirstein, 1995; Abdullah et al., 2000;
Davidson et al., 2001).
35
Chapter 2. Literature review
•
Visual inspection is used extensively for the quality assessment of meat products
applied to processes from the initial grading to consumer purchases (Basset et
al., 2000; Tan et al., 2000).
•
The use of CV systems in the fish industry has proved successful in reducing
human inspection labour and costs (Strachan and Kell, 1995) and in
discriminating between fish species (Zion et al., 1999; Storbeck and Daan,
2001).
•
Machine vision has been used in vegetable grading and inspection applications ,
as well as for quality assessment purposes (Howarth and Searcy, 1992; Hayashi
et al., 1998). It has also been used for classification, defect detection, quality
grading and variety classification of fruits (Brosnan and Sun, 2004).
•
Machine vision has also been applied successfully in the inspection of prepared
foods, cereal grains and food containers (Seida and Frenke, 1995; Sun, 2000;
Manickavasagan et al., 2008).
A recent book edited by Da-Wen Sun (2008) presents extensive coverage of the
application of computer vision in areas such as fresh and cooked meats, poultry,
seafood, agricultural products (e.g. apples, citrus, strawbery, table olives and potatoes),
grains (e.g. wheat, rice and corn) and other foods (e.g. pizza, cheese, bakery, noodles
and potato chips). With respect to the topic of this thesis, some of the most relevant
applications reported in that book are:
•
Defect recognition and classification in apples, including fungal attack, and
mechanical bounds (Leemans and Kleynen, 2008).
•
Defect recognition in citrus fruits, including mechanical damage (Moltó and
Blasco, 2008).
36
Chapter 2. Literature review
•
Bruise detection in strawberries (Shrestha et al., 2001).
•
Detection of bruises and necrosis on peeled potatoes (Marique et al., 2005).
2.5.2.3 Application in mushrooms
The application of image analysis for mushroom testing was initiated in the 1990’s, when
van den Vooren et al. (1992) developed a CV system and applied various image
analysis techniques to obtain morphological parameters from grayscale images. The
main purpose was to perform variety tests of mushrooms [Agaricus bisporus (Lange)
Imbach] to grant Dutch plant breeders’ rights. This system allowed for correct
identification of 80% of the cultivars tested.
Heinemann et al. (1994) investigated the potential of a monochrome camera for
mushroom grading (in terms of size, shape, colour, veil opening and stem cut). An
average misclassification rate of 20% was reported, which compared favourably to the
ability of human inspectors. The development of such a system helped standardise the
quality evaluation of mushrooms in Pennsylvania mushroom farms, which, up to that
time, had been performed by hand, making it a labour intensive, subjective and not very
consistent task. Reed et al. (1995) also used camera based technology to select
mushrooms by size for picking by a mushroom harvester. A CV system developed by
Van Loon (1996) measured the developmental stage of mushrooms and findings
indicated that cap opening correlated well with development.
Vízhányó and Felföldi (2000) used mushroom images recorded by a machine vision
system to recognise and identify discolouration caused by bacterial disease and to
differentiate it from senescence (i.e. natural browning) .
Aguirre et al. (2008b) used grayscale images to examine the influence of relative
humidity and temperature on the browning and brown spotting of mushrooms, finding
optimal storage conditions to minimise the rate of browning and appearance of spotting.
37
Chapter 2. Literature review
2.5.2.4 Limitations
CV has some drawbacks that make it unsuitable for certain industrial applications. It is
inefficient in the case of objects of similar colours, inefficient in the case of complex
classifications, unable to predict quality attributes (e.g. chemical composition) and it is
inefficient for detecting invisible defects (Elmasry and Sun, 2010). Since machine vision
is operated at visible wavelengths, it can only produce an image registering the external
view of the objects and not its internal view. Situations exist whereby food technologists
need to look inside the object in a non-invasive and non-destructive manner. Internal
structures are difficult to detect with relatively simple traditional imaging methods, which
cannot provide enough information for detecting internal attributes (Du and Sun, 2004).
Experience acquired over the experimental studies of this thesis revealed that the lack of
stability of some of the simplest sensors and the effect of external factors (e.g.
illumination conditions, camera location, sample morphology, etc) on the quality of the
acquired also limit the use and adoption of this technology.
2.5.3 Hyperspectral Imaging
Conventional CV systems may be useful for many food sorting operations, but they tend
to be poor identifiers of surface features sensitive to wavebands other than RGB. To
overcome this, spectral imaging systems (i.e. multispectral, hyperspectral and
ultraspectral) have been developed to combine images acquired at a number of narrow
wavebands, sensitive to the features of interest of the object (Gowen et al., 2010). In
particular, visible (Vis) and near-infrared (NIR) spectroscopy are established techniques
for chemical determination of foods.
Hyperspectral imaging, also known as chemical or spectroscopic imaging, is an
emerging technology that integrates conventional imaging and spectroscopy to attain
both spatial and spectral information from an object. Imaging is essentially the science of
38
Chapter 2. Literature review
acquiring spatial and temporal data from objects using a digital camera, whereas
spectroscopy is the science of acquiring and explaining the spectral characteristics of an
object to describe the light intensities emerging from its molecules at different
wavelengths and thus providing a precise fingerprint of that object. HSI was developed
to integrate spatial (imaging) and spectroscopic information -which otherwise cannot be
achieved with either conventional imaging or spectroscopic techniques (Elmasry and
Sun, 2010). Gowen et al. (2007) described this technology as an emerging process
analytical tool for food quality and safety control and reviewed its origins, technical
aspects and applications.
HSI was originally developed for remote sensing applications (Goetz et al., 1985) but
has since found application in diverse fields such as astronomy (Hege et al., 2003),
agriculture (Monteiro et al., 2007), pharmaceuticals (Gowen et al., 2008a) and medicine
(Zheng et al., 2004). HSI has more recently emerged as a powerful process analytical
tool for the analysis of food (ElMasry et al., 2007; Naganathana et al., 2008).
The non-destructive, rugged and flexible nature of HSI makes it an attractive process
analytical technology for identification of critical control parameters that impact on
finished product quality of an industrial process (Lewis et al., 2005). It is expected that
when HSI technology becomes more advanced technologically with cheaper instruments
and faster imaging speed, it will be increasingly adopted as a process analytical tool for
the food industry.
2.5.3.1 Hypercube formation
Hyperspectral image data consist of several congruent images representing intensities
at different wavelength bands and spatial positions. Hyperspectral images, also known
as hypercubes, are three-dimensional blocks of data, comprising two spatial and one
wavelength dimension (Figure 2.12). A hypercube consists of a series of contiguous
sub-images, each of which represents the intensity and spatial distribution of the tested
39
Chapter 2. Literature review
object at a certain wavelength. Equally, every pixel in a hyperspectral image contains
the spectrum (i.e. information about chemical composition) of that specific position along
the wavelength range. The resulting spectrum acts like a fingerprint, which can be used
to characterise the composition of that particular pixel.
Single wavelength image
Pixel spectrum
Figure 2.12 Hypercube structure.
Technically speaking, the hyperspectral data are characterised by the following features
(Elmasry and Sun, 2010):
•
Hypercubes are inherently high dimensional since they are, by definition,
composed of large numbers of spectral bands. Even though high dimensionality
offers access to rich information content, they also represent a dilemma for data
processing.
•
Data volumes are very large and suffer from collinearity problems, which means
large parts of the data might be redundant. This has implications on storage,
management, processing and analysis of the images.
•
The hypercube can be viewed in the spatial domain as images at different
wavelengths or in the spectral domain as spectral vectors at all the wavelengths,
as shown in Figure 2.13. For instance, if one hyperspectral image has
40
Chapter 2. Literature review
dimensions of 250 columns × 200 rows × 125 wavelengths, this hypercube can
be interpreted as 125 single channel images each with 50000 ( = 250 × 200)
pixels. Alternatively, the same image cube can be viewed as 50000 spectra,
each with 125 wavelength points.
m*n rows
m rows
K columns (wavelengths)
n columns
Raw hyperspectural data
with dimension of mxnxK
Unfolded hyperspectural data
with dimension of (mxn)xK
Figure 2.13 Unfolding of a hypercube.
(Elmasry and Sun, 2010)
It is currently not feasible to obtain information in all three dimensions of a hypercube
simultaneously; instruments can only capture one- or two- dimensional subsets of the
datacube, and thus require the temporal scanning of the remaining dimension(s) to
obtain the complete hypercube. Generally, two dimensions are obtained at a time, and
the third dimension is created by stacking several two-dimensional planes in sequence.
There are three ways to build one spectral image: area scanning, point scanning and
line scanning, all shown in Figure 2.14. Area scanning involves keeping the image field
of view fixed and obtaining images one wavelength after another. Point scanning
measures the spectrum of a single point and then the sample is moved and another
spectrum is taken. The sample is moved in the x and y directions point-by-point using a
computer-controlled stage. Line scanning or pushbroom systems construct the
41
Chapter 2. Literature review
hypercube by acquiring spectral measurements from a line of samples which are
simultaneously recorded by an array detector. This system also requires relative
movement between the object and the detector, but the sample is moved line-by-line in
this case. The resultant hypercube is stored in Band Interleaved by Line (BIL) format.
This method is particularly well suited to conveyor belt systems and may therefore be
more practicable than the former ones for food industry applications.
Area scanning
Point scanning
λ
λ
λ
y
y
x
Line scanning
x
y
x
Figure 2.14 Methods for acquiring three dimensional hypercubes.
Hypercubes contain two spatial (x,y) and one spectral (λ) dimensions. Arrows represent
scanning directions and gray areas represent data acquired at a time (Qin, 2010).
2.5.3.2 Components
Typical pushbroom hyperspectral imaging systems contain the following components:
objective lens, spectrograph, camera, acquisitions system, translation stage, illumination
and computer, as shown in Figure 2.15. Each of these components has its own
characteristics that influence the total accuracy of the system.
The camera, spectrograph and illumination conditions are the factors that determine the
spectral range of the system: Vis-NIR systems range between 400-1000 nm, and often
utilise CCD cameras. The sample is usually diffusely illuminated by a tungsten-halogen
or LED lamp. A beam of light reflected from the sample enters the objective lens and is
separated into its component wavelengths by diffraction optics contained in the
42
Chapter 2. Literature review
spectrograph; a two-dimensional image (spatial dimension × wavelength dimension) is
formed in the camera and saved on the computer. The sample is moved past the
objective lens on a motorised stage and the processes repeated; two-dimensional line
images acquired at adjacent points on the sample are stacked to form a threedimensional hypercube which may be stored on a computer for further analysis.
Slit
Spectrograph
Camera
M
irr
or
Objective
Figure 2.15 Diagram and image of a pushbroom hyperspectral imaging system.
(Gowen et al., 2010).
2.5.3.3 Analysis of images
The analysis of hyperspectral data aims to reduce the dimensionality of the data while
retaining important spectral information with the power to classify important areas of a
scene.
The following are typical steps in the analysis of hyperspectral images (Gowen et al.,
2007; Elmasry and Sun, 2010), as outlined in Figure 2.16:
43
Chapter 2. Literature review
•
Calibration is performed prior to sample data acquisition with the objective of
accounting for the background spectral response of the instrument and the “dark”
response of the CCD camera detector.
•
Pre-processing is usually carried out to remove non-chemical biases from the
spectral information and prepare the data for further analysis. Pre-processing of
spectral data is often of vital importance if reasonable results are to be obtained
from the spectral analysis step. Pre-processing includes spectral and spatial
operations.
o
There are a number of techniques used for spectral pre-processing, such
as normalisation, mean centring, standard normal variate (SNV),
multiplicative scatter correction (MSC) and differentiation (SavitzkyGolay). Some of these techniques will be described in section 2.6.
o
Spatial pre-processing is usually performed to remove noise (e.g. by
filtering) and enhance the spatial properties of an image.
Other steps usually carried out at the pre-processing stage include segmentation
operations, such as thresholding and masking, to differentiate the object of
interest from the redundant background information in the hypercube.
•
Spectral data analysis attempts to address what different components are
present in the sample and how they are distributed. Many chemometric tools fall
under this category. Data analysis is performed using multivariate analysis by
one or more chemometric tools, including correlation, classification and spectral
deconvolution techniques. Classification techniques, which can be qualitative/
quantitative and unsupervised/supervised, enable the identification of regions of
the hypercube with similar spectral characteristics. Section 2.7 of this thesis
includes descriptions of various classification techniques.
To build concentration maps for determining the estimated concentrations of
different components present in the tested sample and their spatial distribution, a
44
Chapter 2. Literature review
quantitative assessment should be performed using a standard analytical
method. Multivariate chemometric techniques can be used to build the calibration
models to relate spectral data to the actual quantitative data.
Moreover, with the aid of multivariate analysis, dimensionality and collinearity
problems of hyperspectral data can be reduced or eliminated by selecting the
spectral data at certain important wavelengths. Selection of important
wavelengths is an optional step but reduces the size of the required
measurement data while preserving the most important information contained in
the data space.
•
Image processing is performed to convert the contrast developed by the previous
steps into a picture showing component distribution. Grayscale or colour
mapping with intensity scaling is frequently used to display compositional
contrast between pixels in an image.
Reflectance calibration
Pre-processing
Spectral data analysis
Image processing
Figure 2.16 Schematic diagram of hyperspectral data analysis process.
2.5.3.4 Application of HSI to food quality and safety
HSI has recently emerged as a useful tool for quality analysis of consumer goods (e.g.
food and pharmaceutical products). Due to the current high cost of HSI systems amd
long scanning times, most food related research has been geared towards the
45
Chapter 2. Literature review
identification of important wavebands that will be employed in the development of low
cost multispectral imaging systems.
Hyperspectral imaging, like other spectroscopic techniques, can be performed in
reflectance, transmission or fluorescence mode. The majority of published studies have
been carried out in the reflectance mode (Gowen et al., 2007), which is usually
performed in the Vis-NIR (400-1000 nm) or NIR (1000-1700 nm) range.
Reflectance HSI has been used to detect defects and contaminants, as well as for
evaluating the quality of fruits, vegetables, meat and fish products, as described below:
•
Defect detection: Cheng et al. (2004) and Liu et al. (2005) conducted used a HSI
systems operating in the 430-930 nm region and reflectance mode to identify
chilling damage in cucumbers.
Ariana et al. (2006) investigated the use of NIR reflectance HSI (900-1700 nm)
for detection of bruises on pickling cucumbers, and later extended the study by
investigating the application of Vis-NIR transmittance HSI (450-1000) for the
same purpose (Ariana and Lu, 2008b).
Nicolaï et al. (2006a) developed a NIR reflectance HSI system (900-1700 nm) to
detect bitter pit defect in apples. The system was capable of identifying bitter pit
lesions invisible to the naked eye, although reduced luminosity at the image
edges caused some misclassification errors.
ElMasry et al. (2009) investigated the use of a Vis/NIR HSI system (400-1000
nm) and ANN for the detection of chilling injury in apples.
o
Bruise detection: Bruising is one of the most important and prevalent
surface defects in horticultural products, and a frequent cause of value
loss in the market. Xing et al. developed a Vis-NIR (400-1000 nm) HSI
system to identify bruises on apples and tested it on different varieties
(Xing et al., 2005; Xing and De Baerdemaeker, 2005; Xing et al., 2007). A
46
Chapter 2. Literature review
four wavelength multispectral system identified bruises with high
accuracy. The same research group also developed a multispectral
imaging system to discriminate between bruises and the stem-end/calyx
on apples (Xing et al., 2006), a well known problem in the apple sorting
technology world.
Elmasry et al. (2008) developed a system operating in the same
wavelength region but with a different apple variety, and found that, by
selecting a three wavelength multispectral system, bruised and sound
apples could be successfully distinguished.
o
Disease detection: Qin et al. (2009) developed a HSI system (450-930
nm) to detect citrus canker in grapefruits, which is a severe disease that
can affect the peel of most commercial citrus varieties. The use of the
Spectral Information Divergence (SID) classification method, which is
based on quantifying the spectral similarities by using a predetermined
canker reference spectrum, allowed for correct classification accuracy of
96.2%.
Gómez-Sanchis et al. (2008a) studied the feasibility of detecting rot
caused by Penicillium digitatum (fungi) in mandarins with a HSI system
operating in the 320-1100 nm range. The study concluded that the
minimum number of bands to optimise successful classification was 20
and obtained 91% correct classification.
•
Contaminant detection: The research group of Kim et al. (2001) developed a
laboratory-based HSI system with a spectral range of 430-930 nm to conduct
food quality and safety research. The system was used in reflectance HSI
experiments for the detection of defects and faecal contaminants on apple
surfaces (Kim et al., 2002; Mehl et al., 2004; Liu et al., 2007). Simple universal
47
Chapter 2. Literature review
algorithms were developed to detect faeces, although performance was poor
when the layers of contaminants were thin.
Park et al. (2006) investigated the performance of a Vis-NIR hyperspectral
reflectance imaging system for poultry surface faecal contaminant detection. The
system allowed for the selection of relevant wavelengths and led to the
construction of a multispectral imaging system. Further investigations (Park et
al., 2007) employed the same system to identify the type and source of faecal
contaminants.
•
Quality evaluation: Polder et al. (2002) showed that a hyperspectral reflectance
system in the spectral region 396-736 nm was more effective than RGB imaging
for discriminating ripeness level in tomatoes, regardless of the illumination
conditions.
Elmasry et al. (2007) used a Vis-NIR HSI system (400-1000 nm) for nondestructive determination of strawberry quality. Optimal wavelengths were
obtained from PLS analysis, and MLR was then used to predict moisture content,
total soluble solids content and pH. Similar systems were used to evaluate pork
quality and marbling level (Qiao et al., 2007; Liu et al., 2010).
Ariana and Lu (2008a; 2008c) developed and applied a HSI system operating
under simultaneous reflectance (400-675 nm) and transmittance (675-1000 nm)
modes for external and internal quality evaluation of pickling cucumbers. The
method was based on the premise that reflectance in the visible region is useful
for detecting external characteristics, while transmittance in the near-infrared
region is suited to internal quality evaluation.
The feasibility of a HSI system operating in the 1000-1700 nm region as an
accurate, objective and fast tool for apple fruit maturity determination was
investigated by Menesatti et al. (2008). Traditionally, the reference method for
measuring the maturity index of apples was the starch-iodine test, after which
48
Chapter 2. Literature review
experts assigned a starch index score to each sample by comparing them to
reference colour charts. The aim of the work was to study the relationship
between NIR spectral images and starch/starch free patterns, to indirectly
measure the maturity stage and avoid experts’ subjectivity in the assignment of
starch index. PLS-DA models were used to classify pixels using their reflectance
spectrum; > 80% and >66% correct classifications were obtained with models
built on individual apples and apple batches, respectively.
The use of hyperspectral imaging in fish production has recently been reported in
the literature. ElMasry and Wold (2008) used an online hyperspectral imaging
system (460-1040 nm) for quantitative measurements of moisture and fat
distribution of fish fillets. Menesatti et al. (2010) used a system working in the
400-970 nm spectral range to provide qualitative evaluation of fish freshness. An
objective technique based on a combination of HSI and geometric morphometric
tools was presented. This study represented an important methodological
evolution in the assessment of fish freshness, as it minimised the error
associated with the subjective choice of fish areas by operators.
2.5.3.5 Application of HSI to mushrooms
The application of HSI for quality prediction of A. bisporus mushroom slices during
storage was investigated by Gowen et al. (2008c). The instrument operated in
reflectance mode in the 400-1000 nm region. MLR models were developed based on
average spectral data to study the correlation between reflectance values at certain
wavelengths and various quality parameters, such as moisture content, colour and
texture. The ability of two and four parameter MLR models to predict the quality
parameters under study was compared and the latter was found to be better.
The same research group investigated the application of the system described above to
detect damage on the cap of whole mushrooms (Gowen et al., 2008b). As mushrooms
49
Chapter 2. Literature review
commonly exhibit surface browning due to physical impact during picking, packaging
and distribution, the mushrooms under study were damaged by controlled vibration to
simulate transport. In this case, two data reduction methods were investigated: in the
first, PCA was applied to the hypercube of each sample, and the PC 2 scores image
used for identification of bruise damaged regions of the mushroom surface; in the
second, PCA was applied to a dataset comprising of average spectra from regions of
normal and bruise damaged tissue. In this case, it was observed that normal and
bruised tissues were separable along the PC 1 axis. The second method performed
better when applied to a set of independent samples.
Due to the high amount of moisture content in mushrooms, storage at temperatures
below 0°C causes freezing damage and water loss after thawing and also leads to
enzymatic browning. A study was carried out to investigate the potential use of the
aforementioned system for identification of A. bisporus mushrooms subjected to freeze
damage before damage was visibly evident (Gowen et al., 2009). Mean reflectance
spectra of selected ROIs were corrected using SNV transformation. A procedure based
on PCA and LDA was developed for classification of both selected spectra and whole
mushrooms into undamaged and freeze damaged groups. It was found that after only 45
min thawing, when freeze-thaw damage was not visible to the naked eye, freeze
damaged mushrooms could be classified with high accuracy ( >95% correct
classification).
Moisture content is one of the most important parameters affecting mushroom quality
and shelf-life, and its measurement is also very important in the production of dried
mushrooms. Taghizadeh et al. (2009) investigated the potential of their HSI system to
predict moisture content of white button mushrooms. Mushrooms were subjected to
dehydration at controlled temperature for different time periods to obtain representative
samples at different moisture levels. Average reflectance spectra of the samples were
obtained and SNV transformation was found to be the best approach to build PLS
50
Chapter 2. Literature review
regression models to predict mushroom moisture content. The implemented method
highlighted contrast between areas of different moisture content to achieve better
knowledge of dehydration over the mushroom surface.
The shelf-life of mushrooms packaged using different polymer top-films and perforation
sizes was also investigated using the same HSI system to extract imaging data through
the packaging film (Taghizadeh et al., 2010). Quality indicators such as weight loss,
colour, maturity index and in-pack gas composition were measured and PLS regression
models were built to correlate HSI data with measured quality parameters. Results
demonstrated that HSI can be used for rapid evaluation of mushroom quality, facilitating
the non-destructive evaluation of the effect of the packaging systems on mushroom
shelf-life. The results obtained also showed that the PET film perforated with holes of 1
mm in diameter was superior in terms of maintaining overall quality. Perforated PET
packaging film proved a viable alternative to the conventional PVC film, facilitating an
increase in mushroom shelf-life from 10 to 14 days.
2.5.3.6 Advantages
Hyperspectral imaging offers many advantages over conventional analytical methods
(Gowen et al., 2007). It is a non-contact, non-destructive method for which no sample
preparation is required. It attains spectral and spatial information simultaneously. By
combining the chemical selectivity of spectroscopy with the power of image visualisation,
HSI enables a more complete description of constituent concentration and distribution
throughout heterogeneous samples. This ability to identify spatial distribution of chemical
and physical components in a sample makes HSI stand out over traditional analytical
methods. HSI is also a sensitive technique to detect minor components of samples and
accepts samples of various geometries. Its labour, time, reagent and waste
management costs are low. All these features make hyperspectral imaging the most
51
Chapter 2. Literature review
complete and versatile tool of all the imaging and spectroscopic techniques known to
date.
2.5.3.7 Limitations
There are a few major barriers to the widespread adoption of HSI in the food industry.
The first is the high purchase cost. Since this technology is emerging as a tool for food
applications, there are only a few commercial suppliers. The second limiting factor is
sample geometry. In the case of non-flat samples, signal differences may arise from
changes in the length from the detector to areas on the sample with variable height.
Non-destructive correction of such morphologically-induced spectral variability may be
obtained by developing application-specific post-processing algorithms. Another
limitation is the image acquisition time. Depending on sample size and image resolution,
acquiring an image can take from seconds to minutes, while processing and
classification time are largely dependent on computer capabilities.
2.6 Spectral data pre-processing techniques
In NIR spectroscopy, the reflected energy is a mixture of diffuse and specular reflections
which are dependant on the scattering nature and absorption characteristics of the
sample (Dhanoa et al., 1994). Scattering is the process by which light is deviated by
non-uniformities (e.g. size and nature of particles) in the sample. Without this effect, light
would be either transmitted through or absorbed by the sample, and reflectance analysis
would be impossible (Isaksson and Næs, 1988). Scattering of light is an important basis
for useful spectroscopic analysis; it is, however, a rather complicated phenomenon and
a major source of variation in NIR spectra.
Scattering effects are frequently encountered when obtaining diffuse spectra of solid and
semi-solid materials (Burger and Geladi, 2007). Scattering effects vary from sample to
52
Chapter 2. Literature review
sample and are both additive and multiplicative in nature (Dhanoa et al., 1994). Additive
effects cause vertical displacement or “shift”, whereas multiplicative effects results in a
non-unity slope for each spectrum when compared to an “ideal” or reference spectrum.
The difference in scattering from sample to sample can completely dominate the visual
inspection of spectra. Data pre-processing is useful to eliminate systematic variations
unrelated to analyte concentration in the sample (Azzouz et al., 2003). The methods of
standardisation or correcting for particle size effects in NIR spectroscopy include
standard normal variate (SNV) transformation, multiplicative scatter correction (MSC)
and differentiation. All these transformations are applied before entering the calibration
process and beneficial effects have been reported in the literature (e.g. simpler
calibration models are obtained, which are believed to be based on more chemical
information).
Some commonly employed techniques are described below. Figure 2.17 shows the
effects of different pre-processing techniques on a set of 500 mushroom spectra.
2.6.1 Standard normal variate
This pre-processing step removes the multiplicative interference of scatter (Barnes et al.,
1989). The correction is applied independently for each individual spectrum and at each
wavelength using Equation 2.1.
SNVij =
ri j − ri
l
∑ (r
j =1
ij
− ri ) 2 /(l − 1)
Equation 2.1
where SNVij is a single data point corresponding to the standard normal variate of
spectrum i at wavelength j, rij is the reflectance value of spectrum i at wavelength j,
ri is the average value of spectrum
i across all the wavelengths and
to the number of wavelengths in spectrum i .
53
l corresponds
Chapter 2. Literature review
2.6.2 Multiplicative scatter correction
This pre-processing method simultaneously corrects for both multiplicative and additive
scatter effects (Isaksson and Næs, 1988) by estimating the relation of the scatter of
each sample with respect to the scatter of a sample reference (Geladi et al., 1985).
Thus, the same level of scatter for all spectra is obtained.
First, a reference spectrum is selected, denoted by the vector r , which usually is the
spectrum mean of all the spectra in the calibration data set. Every spectrum is then
corrected using a linear equation in Equation 2.2:
ri = a1+br +ei
Equation 2.2
where
ri is a vector corresponding to spectrum i , the constant a is the bias indicating
a constant linear absorption additive effect with respect to the mean spectrum r ,
1 = (1,...,1)T is a vector of length l (i.e. the number of wavelengths in ri ) consisting of
ones, the constant
effects and
b is the slope indicating the influence of absorption multiplicative
ei is the residuals vector giving the differences between ri and r . These
residuals vectors are assumed to contain information about the chemical variance.
The new corrected spectrum r i is obtained by means of the expressions in Equation
2.3 and Equation 2.4:
ri − a1
e
=r+ i
b
b
e
rli = r + i
b
Equation 2.3
Equation 2.4
54
Chapter 2. Literature review
2.6.3 Differentiation: Savitzky-Golay
Differentiation takes differences between adjacent points in the spectrum and uses the
resulting differences as spectral variables (Isaksson and Næs, 1988). In this sense, the
additive part of the shift caused by the scatter effect is eliminated. Slope changes from
sample to sample can also be removed by taking second derivatives, i.e. taking
differences between adjacent derivative points. First and second order differentiation
reduce peak overlap and eliminate constant and linear baseline drift, respectively
(Osborne and Fearn, 1986). The Savitzky-Golay method fits a curve through a small
section of the spectrum and then finds the slope of the tangent of this curve at the
central point (Davies, 2007).
(a)
100
(b) 2
80
1
60
0
40
-1
20
-2
0
2800
3350
-3
2800
3900
Wavenumber (cm-1)
3350
3900
Wavenumber (cm-1)
(c) 120
(d) 3
2
100
1
80
0
60
-1
40
20
2800
-2
3350
-3
2800
3900
Wavenumber (cm-1)
3350
3900
Wavenumber (cm-1)
Figure 2.17 Effects of pre-processing techniques on mushroom spectra.
(a) Absorbance spectra; (b) Spectral pre-processing by SNV; (c) Spectral preprocessing by MSC and (d) 1st order Savitzky-Golay differentiation.
55
Chapter 2. Literature review
2.7 Chemometrics
Chemometrics are required to interpret spectral data which otherwise would be difficult
to understand. Chemometrics is a discipline using mathematical and statistical methods
for the selection of the optimal experimental procedure and data treatment for data
analysis (Massart et al., 1997).
Qualitative analysis of spectral data means classification of samples according to their
spectral characteristics. Once the classification of samples has been achieved, it can be
useful to know more precisely to what extent samples are different. Therefore, the
development of quantitative models appears necessary.
The subsections below will detail the most interesting statistical methods that have been
successfully employed in HSI. For a detailed list of chemometric tools applicable in
qualitative and quantitative analysis of spectral data, see Roggo et al. (2007).
2.7.1 Qualitative data analysis
The techniques for spectral data classification can be divided into two categories:
unsupervised and the supervised.
•
As regards unsupervised classification, samples are classified without prior
knowledge, except for the spectra. Then the clusters need to be associated to
different sample characteristics. Principal component analysis (PCA) is probably the
most popular technique of this type and will be described in this section. Other
unsupervised methods include hierarchical and non-hierarchical clustering analysis,
such as Gaussian mixture models and K-means.
•
Supervised techniques are the ones in which prior knowledge (i.e. the category
membership of samples) is required. Thus, the classification model is developed on
a training set of samples with known categories (Massart et al., 2003). Then the
model performance is evaluated by comparing the classification predictions to the
56
Chapter 2. Literature review
true categories of the validation samples. Partial least squares-discriminant analysis
(PLS-DA) falls into this category. Other supervised methods include correlation and
distances based methods (e.g. Euclidean or Mahalanobis), discriminant analysis, K
nearest neighbours (KNN), modelling methods (e.g. soft independent modelling of
class analogy (SIMCA)) and non-linear methods such as artificial neural networks
(ANN) and support vector machines (SVM).
2.7.1.1 Principal component analysis (PCA)
This unsupervised method is frequently used for data compression and exploratory
analysis in spectroscopy (Davies, 2005). It is a feature reduction method which forms
the basis of multivariate data treatment.
PCA is used to visualise the data. One of the most important PCA applications is the
reduction of the number of variables (scores) and the representation of a multivariate
data table in a low dimensional space (Martens and Naes, 1996). In HSI, PCA enables
the reduction of the many (typically>100) spectral dimensions of a hypercube to a
smaller number of principal component (PC) scores (typically<10) which capture the
maximum variation in the data. The new variables (loadings) are linear combinations of
the original ones and a matrix of loadings can be interpreted as the contribution of the
different wavelengths in the spectra towards a particular PC.
2.7.1.2 Partial least squares-discriminant analysis (PLS-DA)
This supervised method is an interesting alternative when dimension reduction is
needed and supervised classification is sought (Chevallier et al., 2006). The basics of
PLS-DA consist firstly in the application of a PLS regression (refer to section 2.7.2.3) on
variables which are indicators of the groups. This step identifies latent variables (LV) in
the features space which have maximal covariance within the predictor variables. The
second step of PLS-DA is to classify observations from the results of PLS regression on
57
Chapter 2. Literature review
indicator variables. The more commonly used method is simply to classify the
observations in the group giving the largest predicted indicator variable.
2.7.2 Quantitative data analysis
2.7.2.1 Multiple linear regression (MLR)
This regression allows establishing a link between a reduced number of wavelengths
and a property of samples (Roggo et al., 2007). The prediction of the searched property
can then be described with the following formula:
k
yi = b0 + ∑ bj x j + ei
j =1
Equation 2.5
where
yi is the observation of interest, such as chemical concentration, bj is the
computed constant coefficient for j = 0 to k,
wavelength and
xi the reflectance at each considered
ei is the prediction error.
Each wavelength is studied one after the other and correlated with the studied property.
The selection is based on the predictive ability of the wavelength. There are three
modes of selection, which are: forward, backward and stepwise.
2.7.2.2 Principal component regression (PCR)
This regression method is divided into two steps: first, the spectral data are treated with
PCA. Then, a MLR is performed on the scores as predictive variables (Roggo et al.,
2007).
The prediction equation is written as follows:
58
Chapter 2. Literature review
Ysampling = Tsampling b
Equation 2.6
where
T
is the matrix of the new dimensional coordinates, Ysampling the reference
values and
b is the coefficient vector.
One of the advantages of using this method is that PCA suppresses spectral collinearity.
One of its drawbacks is that there is no guarantee that the computed PCs will be
correlated to the property of interest.
2.7.2.3 Partial least squares regression (PLSR)
In this method, regressions are computed with least squares algorithms. The goal of
PLSR is to establish a linear link between the matrix of independent variables (i.e.
spectral data, X ) and the vector representing dependent variable (i.e. reference values,
y ). This technique is modelling both X and y in order to find out the variables in the
X
matrix that will best describe the
y vector
(Roggo et al., 2007). This can be
explained by the representation of the spectra in the space of wavelengths, in order to
show directions that will be a linear combination of wavelengths called factors which
best describe the studied property. In the case of hyperspectral imaging, PLS
implementation reduces the dimensionality of the experimental data (wavelengths) into
independent factors or latent variables (LV).
The PLS algorithm according to Geladi and Kowalski (1986b; 1986a), Haaland and
Thomas (1988a; 1988b) and Osborne et al. (1997) determines a set of orthogonal
projection axis
W, called PLS-weights, and wavelength scores T. Then, regression
coefficients b are obtained by regressing
y onto the wavelength scores T as follows:
59
Chapter 2. Literature review
y = XWa*b = Tab
W* = W(P' * W)−1
Equation 2.7
where y is the vector containing the predicted values of the attribute of interest, a is
'
the number of PLS factors, and P is the matrix of wavelength loadings.
PLSR and PCR are both methods to model a response variable when there are a large
number of predictor variables and those predictors are highly correlated or even
collinear (MathWorks, 1994-2010). Both methods construct new predictor variables as
linear combinations of the original predictor variables, and both require the determination
of the optimal number of components (Engelen and Hubert, 2005). If the components
are chosen too small, too much information in the data is lost and the regression model
will fit the data poorly. Assigning a large number of components leads to overfitting,
which implies that the model will fit the calibration data very closely, but it will probably
not be accurate in predicting the response value of new samples.
The way PCR and PLSR construct those components is different (MathWorks, 19942010). PCR creates components to explain the observed variability in the predictor
variables, without considering the response variable at all. On the other hand, PLSR
does take the response variable into account, and therefore often leads to models that
are able to fit the response variable with fewer components. PLSR has the advantages
of PCR without the drawbacks, given that the latent variables are selected according to
the covariance matrix between the data and the investigated parameter.
60
3. AIM AND OBJECTIVES
Chapter 3. Aim and objectives
The work presented in this thesis aims to investigate the use of visible and hyperspectral
imaging systems for the detection and discrimination of mechanical, microbiological and
viral damage of mushrooms.
To achieve this aim, the following objectives were set:
•
To develop injury protocols in order to produce controlled and reproducible
mechanical and microbiological damage to mushrooms.
•
To conduct image analysis experiments of mechanically damaged mushrooms
using a simple computer imaging device, in order to assess the ability of such
systems to monitor different levels of impact damage and discriminate between
them.
•
To carry out enzymatic assays on mushrooms subjected to various intensities of
impact damage, in order to study the effect of mechanical damage on the activity
of polyphenol oxidases.
•
To couple the results obtained in enzymatic assays with hyperspectral images, in
order to appraise the performance of hyperspectral imaging systems in the
prediction of enzyme activity.
•
To perform image analysis experiments of mushrooms subjected to microbial
damage using a hyperspectral imaging system, in order to evaluate the potential
of this technology to detect microbial spoilage and to differentiate it from impact
damage.
•
To conduct image analysis experiments of mushrooms infected with Mushroom
Virus X using a hyperspectral imaging system, in order to appraise the ability of
this technology to differentiate MVX infected mushrooms from sound ones.
62
4. EFFICACY OF A VISIBLE
IMAGING SYSTEM TO
DIFFERENTIATE AND MONITOR
THE BROWNING OF MUSHROOMS
Chapter 4. Visible imaging
4.1 Introduction
Horticultural products such as mushrooms are exposed to external mechanical forces
during their postharvest life (i.e. picking, packaging and distribution, Figure 4.1). Bruising
is the most common type of postharvest mechanical injury (Van Zeebroeck et al., 2007).
Mushroom bruising and cap discolouration are visible and direct consequences of
individuals impacting either each other or a hard surface (Siyami et al., 1988). Bruised
mushrooms are more likely to be rejected by consumers. There are two main colour
attributes of mushrooms which consumers use to accept or reject the products, i) the
general browning of the cap and ii) the appearance of brown spots (Burton, 2004).
Figure 4.1 Stages in mushroom harvesting and transportation.
Left to right: growing, harvesting, transportation (Gowen et al., 2010).
Finding standard damage methodologies is one of the most studied aspects of impact
damage analysis (Menesatti et al., 2002). Previous studies on hard fruit varieties (e.g.
pomaceae and drupaceae) employed free drop and instrumented pendulum devices to
simulate and standardise the damage incurred during produce transport and handling
(Menesatti et al., 1999; Menesatti and Paglia, 2001; Van Zeebroeck et al., 2003). A
study by Berardinelli et al. (2003) employed a vibrating table to reproduce stress during
transportation of more sensitive produce such as shell eggs.
Image analysis is a non-destructive technique that allows the kinetics of both individual
mushrooms and batches to be followed during postharvest storage of the product.
Computerised image analysis techniques offer a methodology to measure uneven
64
Chapter 4. Visible imaging
colouration of mushrooms and may offer several advantages over visual assessment,
such as being cost-effective and producing immediate objective results (Vízhányó and
Felföldi, 2000). Aguirre et al. (2008b) conducted image analysis experiments to examine
the browning and brown spotting processes in mushrooms. Grayscale information was
extracted from RGB images obtained by inexpensive webcams throughout storage and
the data were used to build models to study the dependence of the quality decay
process on storage conditions.
Mushroom bruising and browning are important for producers, as these deterioration
processes may be an influential factor in the acceptance/rejection of mushrooms by
retail stores and subsequent economical gain/loss. Together with identifying the origin of
mechanical impacts and determining ways of minimising them, developing damage
detection tools for the industry is essential in order to reduce losses. Designing
reproducible and controllable damage protocols which allow in-depth study of browning
kinetics could aid such developments. The objective of this work was to evaluate the
ability of a simple RGB imaging system to monitor the browning of mushrooms damaged
using simulated process and transport impact and to differentiate it from the natural
decay kinetics of undamaged control mushrooms. In this way, the aforementioned study
(Aguirre et al., 2008b) was expanded to consider mushroom bruising damage.
4.2 Materials and methods
4.2.1 Mushroom supply
A. bisporus mushrooms (strain Sylvan A15, Sylvan Spawn Ltd., Peterbourough, UK)
were grown in plastic bags and tunnels in Kinsealy Teagasc Research Centre (Kinsealy,
Co. Dublin, Ireland) following common practices in the mushroom industry. Spawn
running and casing took place throughout the six weeks prior to mushroom cropping.
65
Chapter 4. Visible imaging
Only uniform, undamaged, closed cap mushrooms from the first and second flush with a
cap diameter of 3-5 cm were hand-picked, placed in a metal grid and carefully
transported to the laboratory in purpose-built containers, to minimise mechanical
damage during transportation. Mushrooms arrived at the laboratory premises within 1 h
after harvesting; mushrooms which were free from defects were stored overnight at 4°C.
Before damaging, samples were cleaned with a paintbrush in order to remove any
compost/soil particles that could affect subsequent experiments, without causing
mechanical damage to the mushroom pilei.
4.2.2 Simulated process and transport damage protocol
4.2.2.1 Selection of means of damage
The study of damaged samples requires standardising the damage, which entails
selecting an appropriate means of producing the injury. Several devices, such as
penetrometer, Instron testing instrument, shaking sieve and shaking table (Figure 4.2)
were considered and tested for damage production purposes at initial stages of this
study. The aim was to produce controllable, uniform and reproducible damage whose
effect could be detected and measured by a simple RGB imaging device. The shaking
table (Figure 4.2.d) met all these requirements and was therefore selected. This device
produced damage by vibration and induced the formation of a uniform yellow-brown
colour film on mushroom caps. Different levels of damage could be obtained by varying
the shaking amplitude and/or shaking time.
66
Chapter 4. Visible imaging
Figure 4.2 Devices tested for inducing mechanical damage in mushrooms.
(a) Penetrometer; (b) Instron testing machine; (c) Shaking sieve and (d) Shaking table.
4.2.2.2 Description of damage protocol
Samples were subjected to vibrational damage to simulate crop handling and transport
from farm to retail store. Mushrooms were damaged in batches of 600 g (approx) units
inside polystyrene plastic boxes whose internal dimensions were 290 × 200 × 105 mm.
Mechanical damage was induced by using a Gyratory Shaker Model G2 shaking table
(New Brunswick scientific Co., Edison, N.J., USA) at 300 rpm amplitude and controlled
time. Different shaking times led to different damage levels, which resulted in different
colour differences. Damage time ranged from 1 min to 20 min.
4.2.2.3 Standardisation of damage protocol
In order to scale damage in an objective way, discolouration induced by different
shaking times was measured with a colorimeter. Colour was measured every 2 min
within the first 20 min of damage and every 5 min after that until 1 h of damage at 300
rpm amplitude. The colour was recorded by a HunterLab colorimeter (Colour Quest XE
Hunter Lab, Virginia, USA) which reported results in CIELAB colour space, i.e. lightness
variable L* and chromacity coordinates a* (greenness/redness) and b* (yellowness). The
colour of each mushroom was measured from the middle region of the cap and from two
side regions and average values were calculated. The average colour of four randomly
67
Chapter 4. Visible imaging
selected mushrooms was reported for each time point; 76 mushrooms were used in
total.
The damage protocol was standardised by calculating the following parameters:
i) Loss of lightness per unit of damage time.
ii) Gain of yellowness per unit of damage time.
iii) Total colour difference, defined as
ΔE =
(L
- L*i ) + ( a*0 - a *i ) + ( b*0 - b*i )
2
*
0
2
2
Equation 4.1
where the
0
subscript refers to colour measurements at time t0 and the
i
subscript
refers to colour measurements after shaking time ti.
4.2.3 Image acquisition system
Two light sources of 40W fluorescent lamps illuminated the inside of an environmental
incubator (MLR-350 HT, SANYO Electric Biomedical Co. Ltd., Japan). Four commercial
webcams (Live!Cam Notebook Pro Creative webcams, Dublin, Ireland) were installed
inside the chamber, each of which monitored a group of 12 mushrooms. All the cameras
had CMOS sensors and recorded information in VGA format (i.e. 640 × 480 pixel
resolution). Two circular polarising filters with a diameter of 27 mm (Hama, Dublin,
Ireland) were attached to each camera to prevent excessive light saturation of the
sensors.
For each experiment and camera, six mechanically damaged (MD) and six undamaged
(U) mushrooms were randomly displayed on an area of approx. 200 × 250 mm. Black
cardboard was used as background for image acquisition. Two 650 × 600 mm paper
colour standards were used as reference tiles and randomly placed in the field of vision
68
Chapter 4. Visible imaging
of the camera: Ice White (Canford Paper, Dublin, Ireland) whose colour coordinates
were L* = 93.33, a* = 2.05 and b* = 13.66 and Nutmeg (Murano Paper, Dublin, Ireland)
whose coordinates were L* = 61.35, a* = 8.76 and b* = 15.87. The colour of the tiles was
monitored throughout the duration of the study to ensure there was no colour
degradation affecting the tiles during the duration of the experiment.
4.2.4 Experimental design
Seven shaking times (1 min, 2 min, 3 min, 4 min, 5 min, 10 min and 20 min) led to seven
damage levels (MD 1, MD 2, MD 3, MD 4, MD 5, MD 10 and MD 20). A number of 336
samples were used following a full factorial design with two factors (damage and
camera) with seven and four levels, respectively. Damage and camera levels of each
experiment are listed in Table 4.1.
Experiment
number
Table 4.1 Experimental design.
MD 1-MD 20 indicate damage level.
1
2
3
4
5
6
7
C1
MD 1
MD 2
MD 3
MD 4
MD 5
MD 10
MD 20
Camera
C2
C3
MD 2
MD 3
MD 3
MD 4
MD 4
MD 5
MD 5
MD 10
MD 10
MD 20
MD 20
MD 1
M D1
MD 2
C4
MD 4
MD 5
MD 10
MD 20
MD 1
MD 2
MD 3
4.2.5 Image acquisition
Accelerated browning kinetics experiments were performed in an environmental
incubator at 25°C and 90% relative humidity controlled conditions. Pictures were taken
before injury (t0), immediately after damage (t0’) and then hourly (ti) over a period of 12 h.
Both mushrooms and tiles were monitored throughout the duration of each experiment.
RGB images were obtained with a resolution of 640 × 480 pixels and analysed using
69
Chapter 4. Visible imaging
image processing software (ImageJ, Image Processing and Analysis in Java, NIH,
Maryland, USA).
4.2.6 Image analysis
RGB images were converted into an 8-bit grayscale image using the transformation in
Equation 4.2.
Grayscale Value =
Red + Green + Blue
3
Equation 4.2
For each experiment and camera, all converted images were stacked in order of
increasing storage time prior to ROI selection. For each mushroom, the same ROI
comprising the whole cap area was selected throughout the stack. Similarly, ROIs
comprising the whole area of the tiles were selected.
4.2.6.1 Browning kinetics
The average Grayscale Value (GSV) was utilised as a measurement of general
mushroom cap browning. This parameter ranges from 255 (saturated white) to 0 (black);
loss of mushroom cap whiteness and subsequent discolouration produce a decrease in
the grayscale value. The GSV of the reference tiles was monitored to account for
variations in the lighting conditions and camera settings inside the chamber.
Other parameters were also calculated in order to cancel the different sources of noise
in the signal. Table 4.2 contains a summary of such parameters and the rationale behind
these parameters will be discussed in the “Results and discussion” section of this
chapter.
70
Table 4.2 GSV-derived parameters calculated in order to cancel the different sources of noise in the signal.
Abbrev.
Parameter name
Calculation
Equation no.
Equation 4.3
By measuring the difference between
images taken at t0 and at ti, this parameter
eliminates the effect of initial (i.e. prior de
damage) mushroom colour differences.
ΔGSVref = ΔGSV - ΔGSVtile
Equation 4.4
Cancel the noise effect of varying lighting
conditions inside the chamber during the
course of the experiment.
ΔGSVref .w = ΔGSV - weighttile ×ΔGSVtile
Equation 4.5
Eliminate the variations of sensitivities of
the different cameras at different locations.
Equation 4.6
Calculate the slope of the ΔGSVref.w curve
at each time point.
ΔGSV
Delta GSV
ΔGSV = GSVt0 - GSVti
ΔGSVref
Corrected ∆GSV
ΔGSVref.w
Weighed ∆GSV
DΔGSVref.w
ΔGSVref.w
derivative
Explanation
DΔGSVref .w =
d ΔGSVref .w
dt
=
ΔGSVref .wt − ΔGSVref .wt
i−1
i
ti − ti −1
71
GSVref
Corrected GSV
GSVref = GSV - GSVtile
Equation 4.7
This parameter is similar to ΔGSVref, except
that it does not contain information obtained
prior to damage.
GSVref.w
Weighed GSV
GSVref .w = GSV - weighttile × GSVtile
Equation 4.8
This parameter is similar to ΔGSVref.w,
except for the information obtained prior to
damage, which is not included.
Table 4.3 Limits of detection (LODs) set using the different GSV-derived parameters.
Abbrev.
Parameter used
Calculation
LODΔGSVref.w
ΔGSVref.w
LODΔGSV = ( ΔGSVref .w ) + 3 × σ ( ΔGSV
LODDΔGSVref.w
DΔGSVref.w
LODDΔGSV = ( DΔGSVref .w ) + 3 × σ ( DΔGSV
LODGSVref.w
GSVref.w
LODGSV = ( GSVref .w ) − 3 × σ (GSV
U
Explanation
Equation 4.9
Set the ΔGSVref.w value above which mushrooms
would be classified as damaged.
Equation 4.10
Set the DΔGSVref.w value below which mushrooms
would not be classified as damaged any more.
Equation 4.11
Set the GSVref.w value below which mushrooms
would be classified as damaged.
ref . w U
U
U
)
Equation no.
)
ref . w U
)
ref . w U
4.2.6.2 Limit of detection (LOD)
The minimum distinguishable signal of a device can be defined as the sum of the mean
blank signal plus a multiple of the standard deviations of the blank (Skoog et al., 1998).
Minimum distinguishable signal = blank + n×σblank
Equation 4.12
In this study, the blank or indistinguishable signal was the average of U mushroom
images and the standard deviation was multiplied three times (i.e. n = 3). Three LODs
were set using the minimum indistinguishable signal of three of the aforementioned
parameters, as shown in Table 4.3.
LODΔGSVref.w and LODDΔGSVref.w required imaging the mushrooms prior to damage and
LODGSVref.w did not.
4.2.6.3 Time to discrimination
The storage time after which each camera could discriminate between each level of
damage and no damage was established by performing analysis of variance (ANOVA)
between the LOD and the average signal of MD mushrooms. ANOVA was carried out
using R (R Development Core Team, 2007).
4.3 Results and discussion
4.3.1 Standardisation of damage protocol
The vibrational impact damage generated by the shaker induced uniform damage on
mushroom caps. The extent of the injury was confined to the outer layer of the cap,
72
which developed a yellow-brown colour film depending on the duration of the shaking
period. Damage intensity could be easily controlled by modifying the shaking time.
Brennan and Gormley (1998) identified lightness (L*) and yellowness (b*) as the most
important colour indicators of mushroom quality. Figure 4.3 illustrates how these two
parameters varied with increasing damage time. The average value of 12
measurements and their corresponding standard deviations are shown. L* value
decreased with increasing damage time and showed a linear behaviour at low damage
levels. b* value exhibited an increasing trend with damage time and behaved linearly at
low damage levels. The repeatability of the method was good; low standard deviation
indicated that mushrooms treated under similar time and amplitude conditions showed
CIELab colour scale
similar degrees of discolouration.
100
90
80
70
60
50
40
30
20
10
0
L*
b*
0
10
20
30
40
50
60
Damage time (min)
Figure 4.3 Average CIELAB colour coordinates as a function of damage time.
Red horizontal line indicates limit of acceptable mushroom quality.
Gormley and O’Sullivan (1975) correlated L* values with sensory analysis in order to
develop an objective mushroom grading scale. According to this criterion, L* >93,
“excellent” quality; 90<L*<93, “very good” quality; 86 <L*<89, “good” quality; 80<L*<85
“reasonable” quality; 69<L*<79, “poor” quality, not acceptable for wholesale; L*<69 “very
poor” quality, not acceptable for retail. The initial lightness of the samples collected for
73
this study was close to 100 and that could explain why relatively long damage times
were needed to get below the acceptable limit of quality (i.e. L* = 79, shown in red in
Figure 4.3). Average L* went below 80 only after 30 mins of damage. As this study
aimed at investigating mushrooms with the highest quality within the customer
acceptability range, only damage times of up to 20 mins were selected (i.e. MD 1 to MD
20) for kinetics studies. Figure 4.4 represents average lightness decrease and
yellowness increase over that period of time, where L* and b* parameters showed a
linear behaviour with R2 of 0.94 and 0.97, respectively.
100
CIELab colour scale
90
80
L* = ‐0.66 x Time + 99.81 R2 = 0.94
70
60
50
40
L*
b* = 0.60 x Time + 14.63 R2
30
= 0.97
b*
20
10
0
0
5
10
15
20
Damage time (min)
Figure 4.4 Average CIELAB colour coordinates and their calibration curves as a function
of damage time.
In this way, the effects of the damaging time on the colour parameters were set as
follows:
i) Loss of lightness: 0.66 units / min of damage.
ii) Gain of yellowness: 0.60 units / min of damage.
iii) ΔE values for damage times of up to 20 min are shown in Table 4.2.
74
Table 4.4 Total colour difference (ΔE) as a function of damage time.
Damage time (min)
2
4
6
8
10
12
14
16
18
20
ΔE
5.22
6.69
7.80
9.43
11.65
13.82
17.40
16.97
16.76
21.01
All of these values are specific to the model, operative amplitude and sample load of the
shaker, which in this study were Gyratory Shaker Model G2 (New Brunswick scientific
Co., Edison, N.J., USA), 300 rpm and load of 600 g (approx.) of mushrooms with 3-5 cm
cap diameter, respectively.
4.3.2 Browning kinetics
The whiteness of both MD and U mushrooms decreased with time as shown in Figure
4.5, where GSV (255 for white and 0 for black) is represented as a function of storage
time. Blue and pink lines correspond to MD and U mushrooms, respectively.
75
MD
U
0
C4
MD 1
5 10
0
C4
MD 2
C4
MD 3
5 10
0
C4
MD 4
C4
MD 5
5 10
C4
MD 10
C4
MD 20
GSV [0-255]
200
150
100
50
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
200
150
100
50
200
150
100
50
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
200
150
100
50
0
5 10
0
5 10
0
5 10
0
5 10
Storage time [h]
Figure 4.5 Grayscale Value (GSV) of mushrooms as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level.
All mushrooms showed a decreasing trend with time; the drop was very slight for U
mushrooms and more pronounced for MD mushroom. Loss of whiteness increased with
shaking duration. U and MD lines overlapped in some of the graphs and sample spread
was wide in most cases. The sources of these variations are thought to be i) biological
variability of fresh produce and ii) specific position of each mushroom in the chamber.
The relative position of the camera with respect to the light sources and the relative
76
position of a mushroom with respect to the camera resulted in lighting differences within
a group of mushrooms.
Delta Grayscale Value (∆GSV), which is the difference between the image taken at t0
(i.e. before damage) and the image at any other time point (i.e. ti), was calculated for
each mushroom and time point. In this way, natural colour variability among undamaged
mushrooms was eliminated and only discolouration due to damage was considered.
Figure 4.6 shows ∆GSV kinetics for all cameras and damage levels.
MD
U
0
C4
MD 1
5 10
0
C4
MD 2
C4
MD 3
5 10
0
C4
MD 4
C4
MD 5
5 10
C4
MD 10
C4
MD 20
ΔGSV [0-255]
100
50
0
-50
100
50
0
-50
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
100
50
0
-50
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
100
50
0
-50
0
5 10
0
5 10
0
5 10
0
5 10
Storage time [h]
Figure 4.6 Delta Grayscale Value (ΔGSV) of mushrooms as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level.
77
U and MD mushrooms appeared more separated for most plots in this figure. Data from
the pictures taken prior to damage were useful for separating U and MD lines, reducing
data variability and allowing for better visualisation for the purpose of assessing the
nature of the browning kinetics. ∆GSV resulted in being a better indicator than GSV for
visualisation of the damage browning kinetics of mushrooms.
4.3.2.1 Signal drift
When having a closer look at any of the plots in Figure 4.5 or Figure 4.6, considerable
signal noise was observed in all mushroom colour kinetic patterns. The occurrence of
the drift was random. Small increases of GSV in Figure 4.6 (and small decreases of
∆GSV in Figure 4.6) had no physical meaning and could not be justified from a biological
perspective, as mushrooms can only discolour with time (i.e. decrease and increase
their GSV and ∆GSV values, respectively). All the mushrooms within a batch (i.e.
monitored by the same camera) followed the same noise pattern.
Figure 4.7 demonstrates that these variations were also present in the patterns of the
two reference tiles, whose values should not vary in an ideal situation.
The similarity of all the mushroom and tile patterns monitored under the same camera
and experiment suggested a lack of stability of the chamber illuminants and/or the
webcam sensors.
The information contained in the reference tiles regarding illumination instability was
incorporated by calculating a new parameter, corrected ∆GSV (∆GSVref), which is the
difference between the ∆GSV of the mushroom and tiles monitored by the same
camera. ∆GSVref used the Ice White tile as a reference and its kinetics for all cameras
and damage levels are shown in Figure 4.8
78
MD
U
White tile
Nutmeg tile
C3
MD 20
100
ΔGSV [0-255]
80
60
40
20
0
0
2
4
6
8
10
12
Storage time [h]
Figure 4.7 Delta Grayscale Value (ΔGSV) of the MD 20 / C 3 experiment as a function
of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
Circles and lines show individual and average behaviour of all the mushrooms of each
type, respectively.
79
MD
U
0
C4
MD 1
5 10
0
C4
MD 2
C4
MD 3
5 10
0
C4
MD 4
C4
MD 5
5 10
C4
MD 10
C4
MD 20
ΔGSVref [0-255]
100
50
0
-50
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
100
50
0
-50
100
50
0
-50
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
100
50
0
-50
0
5 10
0
5 10
0
5 10
0
5 10
Storage time [h]
Figure 4.8 White-referenced Corrected ∆GSV (∆GSVref) of mushrooms as a function of
storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level.
Signal fluctuations were reduced, which can be observed in more detail in Figure 4.9,
where the reduction of the noise signal was particularly noticeable in U samples.
80
MD
U
C3
MD 20
100
ΔGSV [0-255]
80
60
40
20
0
0
2
4
6
8
10
12
Storage time [h]
Figure 4.9 White-referenced Corrected ∆GSV (∆GSVref) of the MD 20 / C3 experiment
as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
Lines show average behaviour of all of the MD and U samples monitored.
4.3.2.2 Normalisation of cameras within the imaging system
The importance of the reference tile in decreasing the illumination noise was found to be
different for each camera, which could be attributed to different effects of the light source
to each camera and different sensitivity between cameras.
A different “weight” could be assigned to each camera in the imaging system by
analysing the signal of all the undamaged control mushrooms. By doing this, all the
cameras could be normalised to the same units. A linear mixed effect model was used to
estimate the “weight” of each reference tile (i.e. the estimated “weight” of the effect of
the position and sensitivity of each camera), allowing for each individual U mushroom in
81
a camera to have a random intercept, with a common time slope and different reference
tile slopes for each camera.
The results of the estimation can be seen in Table 4.5.
Table 4.5 Estimation of the weight of the reference tile for each camera.
Camera
C1
C2
C3
C4
Estimate
0.52
0.90
0.90
0.90
Std. error
0.025
0.036
0.035
0.028
The differences in the values of the estimates were influenced by both the differences in
cameras and the variability between batches of mushrooms. C 1 presented different
average signal with respect to U mushrooms. This may point to differences in detection
of mushroom damage depending on the location and/or sensitivity of the camera.
The weight for each tile was incorporated into a new parameter, i.e. weighed ∆GSV
(∆GSVref.w), which is the difference between the ∆GSV of the mushrooms and that of the
tile multiplied by its weight. Figure 4.10 shows ∆GSVref.w kinetics for all cameras and
damage levels.
∆GSVref.w kinetics were very similar to those of ∆GSVref, as all the estimates of C 2, C 3
and C 4 are 0.90 (i.e. close to 1). In the case of C 1, the estimate was 0.52, which meant
the difference between ∆GSVref. and ∆GSVref.w would be slightly bigger.
The differences in the slope of ∆GSVref.w curves of damaged and undamaged
mushrooms at initial stages of storage led to some consideration on the calculation of
derivatives and their use as a tool for early discrimination of damage.
DΔGSVref.w, the time derivative of weighed ΔGSV, was calculated as described in
section 4.2.6.1 and is presented as a function of storage time in Figure 4.11.
82
MD
U
0
C4
MD 1
5 10
0
C4
MD 2
C4
MD 3
5 10
0
C4
MD 4
C4
MD 5
5 10
C4
MD 10
C4
MD 20
ΔGSVref.w [0-255]
100
50
0
-50
100
50
0
-50
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
100
50
0
-50
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
100
50
0
-50
0
5 10
0
5 10
0
5 10
0
5 10
Storage time [h]
Figure 4.10 Weighed ∆GSV (∆GSVref.w) of mushrooms as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level.
83
MD
U
0
dt
15
10
5
0
-5
-10
DΔGSVref.w =
dΔGSVref.w
[0-255]
C4
MD 1
5 10
0 5 10
C4
MD 2
C4
MD 3
C4
MD 4
0
C4
MD 5
5 10
C4
C4
MD 10 MD 20
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
C3
MD 10 MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
C2
MD 10 MD 20
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
C1
MD 10 MD 20
15
10
5
0
-5
-10
15
10
5
0
-5
-10
15
10
5
0
-5
-10
0 5 10
0
5 10
0 5 10
0 5 10
Storage time [h]
Figure 4.11 GSVref.w time derivative (D∆GSVref.w) of mushrooms as a function of storage
time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level.
84
Chapter 4. Visible imaging
First derivatives produced peaks where the original kinetics had maximum slope and
crossed zero when the original line had a peak, thus were not easy to interpret. As can
be seen in Figure 4.11, MD and U mushrooms showed different derivative values at
initial times of storage, when the slopes of ∆GSVref.w curves were different (refer to
Figure 4.10). On the contrary, MD and U mushrooms could not be separated at a later
stage by calculating derivatives, which could be explained by the homogeneity of the
slope of both mushroom types in Figure 4.10. This meant that the use of derivatives
would only be an effective alternative at initial stages of storage.
As described in section 4.3.2, having some previous knowledge (i.e. prior to damage) of
the state of each individual mushroom decreased noise signals and allowed for a more
precise inspection of the kinetics of mushroom browning. In this way, discrimination
between MD and U mushrooms was facilitated. Despite the fact that this information
would not available in a real retail situations, the study of these indexes helped to
ascertain the behaviour of the browning kinetics and therefore to identify appropriate
strategies for early browning detection. In this way, it was observed that the shape of the
noise-cancelled browning kinetics followed a parabolic trend, whose time derivative
peaked at short storage times and eventually cancelled.
In attempting to study the kinetics with the only information that would normally be
available in real retail situations, weighed GSV (GSVref.w) was calculated as indicated by
Equation 4.8. This parameter accounted for reference and webcam correction, but did
not include the information contained in images taken prior to damage. The kinetics of
GSVref.w for all cameras and damage levels are presented in Figure 4.12.
85
Chapter 4. Visible imaging
MD
U
0
C4
MD 1
5 10
0
C4
MD 2
C4
MD 3
5 10
0
C4
MD 4
C4
MD 5
5 10
C4
MD 10
C4
MD 20
GSVref.w [0-255]
100
50
0
-50
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
100
50
0
-50
100
50
0
-50
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
100
50
0
-50
0
5 10
0
5 10
0
5 10
0
5 10
Storage time [h]
Figure 4.12 Weighed GSV (GSVref.w) of mushrooms as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level.
It is believed that the calculation of some of the parameters described above could have
been avoided if more robust cameras, such as CCD cameras, had been employed.
Currently, low cost CCD cameras with better colorimetric properties than CMOS sensors
and enhanced image acquisition capacity are available in the market.
86
Chapter 4. Visible imaging
4.3.3. Limit of detection
The average signal of all of the undamaged mushrooms monitored by each camera was
employed to set the LOD of each device. For each of the last three parameters or
signals mentioned above (i.e., ∆GSVref.w, D∆GSVref.w and GSVref.w) and camera, a LOD
( LODΔGSVref.w , LODDΔGSVref.w and
LODGSVref.w , respectively) was calculated as described in
section 4.2.6.2. Each of these LODs was used as a critical value for MD and U
mushroom discrimination.
Table 4.6 contains LOD values for the different parameters and cameras. The minimum
and maximum signal observed among MD mushrooms is also displayed for each
parameter and camera.
Table 4.6 LOD values and minimum and maximum values of MD mushrooms for each
camera and parameter.
C1
LODΔGSVref.w
ΔGSVref.w range
Min
Max
LODDΔGSVref.w
DΔGSVref.w range
Min
Max
LODGSVref.w
GSVref.w range
Min
Max
Camera
C2
C3
C4
32.26
10.57
20.84
18.88
-22.99
127.95
-0.022
76.32
-8.69
99.14
4.85
76.54
3.52
-2.05
-3.40
-3.22
-2.28
16.88
-2.29
12.95
-1.88
16.46
-3.25
14.96
73.97
-28.01
6.81
-28.82
-3.97
119.62
-63.51
81.76
-62.1
61.06
-62.87
29.88
As can be seen in Table 4.6, the LOD of the four cameras in the imaging system differed
from each other, sometimes even in one order of magnitude. These differences could be
due to differences in location inside the environmental chamber, as luminosity showed
slight variations with location. The differences in LOD values could also be affected by
the variations in the detection sensitivity and range of the camera detectors. This
emphasised the need to calibrate camera sensors individually in the system and “train”
them with samples of both types (MD and U) before proceeding to discrimination.
87
Chapter 4. Visible imaging
A system for damage detection is something that could be applied in the management of
mushroom inventories (i.e. to establish a system of produce perishability in the store by
triggering batch disposal after its signal on the shelf is over the LOD). However, it is
necessary to establish how early this detection can happen. If the imaging system takes
more than 3 days to reject a sample, it is likely that its usefulness will be limited, as the
turnover of the product will easily compensate for damage signs after that time. On the
other side, if the cameras can detect the damage early enough, they may be employed
to discard the produce before it arrives at the retail store (or while on display), preventing
consumer dissatisfaction and preserving brand image.
Figure 4.13 shows ∆GSVref.w values of MD mushrooms (blue) for all cameras and
damage levels. The horizontal pink line represents the individual
LODΔGSVref.w for each
webcam, which separates the damaged and undamaged prediction regions of the plot
(i.e. below the
LODΔGSVref.w line represents undamaged; above represents damaged).
88
Chapter 4. Visible imaging
0 2 4 6 8 12
C4
MD 1
C4
MD 2
0 2 4 6 8 12
C4
MD 3
C4
MD 4
0 2 4 6 8 12
C4
MD 5
C4
MD 10
C4
MD 20
100
50
0
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
ΔGSVref.w [0-255]
100
50
0
100
50
0
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
100
50
0
0 2 4 6 8 12
0 2 4 6 8 12
0 2 4 6 8 12
0 2 4 6 8 12
Storage time [h]
Figure 4.13 Weighed ∆GSV (∆GSVref.w) of MD mushrooms as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level. The horizontal
line (pink) represents LOD∆GSV of each camera.
All the blue points which were above the pink line represented the mushrooms which the
camera could identify as being damaged. It was expected that the cameras would
perform better for highest damage levels, when the discolouration was more intense. As
the level of damage decreased, more points fell below the pink line, which means the
camera would not be able to detect the damage on those mushrooms. In some cases,
cameras did not show discriminative ability at early stages of storage but were able to
89
Chapter 4. Visible imaging
identify the damage as the storage time increased. C 1 showed the poorest overall
performance, as it was unable to discriminate between undamaged mushrooms and
mushrooms with low-intensity damage (i.e. MD<5). This may be explained by the low
sensitivity to U mushrooms (refer to weight estimates in Table 4.5).
Figure 4.14 shows D∆GSVref.w of MD mushrooms for all cameras and damage levels.
The horizontal pink line represents the individual
case, above the
LODDΔGSVref.w of each camera; in this
LODDΔGSVref.w line represents damaged and below represents
undamaged.
As D∆GSVref.w calculated the slope of the ∆GSVref.w signal at each time point, it was
expected that the cameras would lose their ability to discriminate between U and MD
mushrooms as the storage time increased. For low-intensity damage, only the first
D∆GSVref.w points fell above the line, which meant the cameras could identify the
damage only at very early stages of storage. As the level of damage increased,
D∆GSVref.w was able to discriminate for longer storage periods. D∆GSVref.w value might
be a good criterion to consider when ∆GSVref.w is not (i.e. early hours of storage), and
vice versa. The complementary use of both criteria might allow for uninterrupted
discrimination of MD mushrooms during storage.
90
Chapter 4. Visible imaging
0 2 4 6 8 12
C4
MD 1
C4
MD 2
0 2 4 6 8 12
C4
MD 3
C4
MD 4
0 2 4 6 8 12
C4
MD 5
C4
C4
MD 10 MD 20
dt
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
C3
MD 10 MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
C2
MD 10 MD 20
15
10
5
0
DΔGSVref.w =
dΔGSVref.w
[0-255]
15
10
5
0
15
10
5
0
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
C1
MD 10 MD 20
15
10
5
0
0 2 4 6 8 12
0 2 4 6 8 12
0 2 4 6 8 12
0 2 4 6 8 12
Storage time [h]
Figure 4.14 D∆GSVref.w derivative (D∆GSVref.w) of MD mushrooms as a function of
storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and MD 1-MD 20 indicate damage level. The horizontal
line (pink) represents
LODDΔGSVref.w of each camera.
Figure 4.15 shows weighed GSV (GSVref.w) of MD mushrooms (blue) for all cameras and
damage levels. The horizontal pink line represents the individual
camera; in this case, above the
LODGSVref.w of each
LODGSVref.w line represents undamaged; below
represents damaged.
91
Chapter 4. Visible imaging
0 2 4 6 8 12
C4
MD 1
C4
MD 2
0 2 4 6 8 12
C4
MD 3
C4
MD 4
0 2 4 6 8 12
C4
MD 5
C4
MD 10
C4
MD 20
GSVref.w [0-255]
100
50
0
-50
C3
MD 1
C3
MD 2
C3
MD 3
C3
MD 4
C3
MD 5
C3
MD 10
C3
MD 20
C2
MD 1
C2
MD 2
C2
MD 3
C2
MD 4
C2
MD 5
C2
MD 10
C2
MD 20
100
50
0
-50
100
50
0
-50
C1
MD 1
C1
MD 2
C1
MD 3
C1
MD 4
C1
MD 5
C1
MD 10
C1
MD 20
100
50
0
-50
0 2 4 6 8 12
0 2 4 6 8 12
0 2 4 6 8 12
0 2 4 6 8 12
Storage time [h]
Figure 4.15 Weighed GSV (GSVref.w) of MD mushrooms as a function of storage time.
MD (blue) and U (pink) represent damaged and undamaged mushroom, respectively.
C 1-C 4 indicate camera number and D 1-D 20 indicate damage level. The horizontal
line (pink) represents LODGSV of each camera.
As observed in the two previous figures, discrimination improved as the level of damage
increased. As can be seen in the above figure, almost all of the points corresponding to
low-intensity damage mushrooms appeared above the horizontal line, which meant
damage was not detected in the mushrooms. Overall, the performance of this parameter
was poorer than that of previous parameters, which confirmed that having information
prior to damage aided in the discrimination of bruised samples.
92
Chapter 4. Visible imaging
Once the limit of detection was established, it was necessary to determine at which point
the sensors would detect the different levels of damage. While there might be some
small level of damage which might fall outside the scope of the sensors, as the injury
became more intense, the signal would eventually reach the LOD at a certain time
during the monitoring of the storage of the mushrooms.
4.3.4 Time to discrimination
The time of MD/U discrimination for each damage level and camera was obtained as the
result of ANOVA performed between individual LOD values and the average behaviour
of MD mushrooms. Where some knowledge prior to damage was available, ANOVA
tested
LODΔGSVref.w against (ΔGSVref.w )MD ; in the case of the time derivatives, ANOVA
tested
LODDΔGSVref.w against (DΔGSVref.w )MD ; when t0 information was not available,
ANOVA tested
LODGSVref.w against (GSVref.w )MD .
Table 4.7 shows the times to discrimination when images were taken prior to damage.
Table 4.7 Time to discrimination of all cameras and damage levels, based on
Damage level
LODΔGSVref.w and (ΔGSVref.w )MD .
MD 1
MD 2
MD 3
MD 4
MD 5
MD 10
MD 20
Results show that the cameras
C1
10h
>12h
>12h
>12h
0h
2h
0h
C 2,
Camera
C2
C3
7h
>12h
0h
1h
0h
1h
0h
1h
0h
0h
0h
1h
0h
1h
C4
1h
0h
1h
0h
9h
0h
0h
C 3 and C 4 were able to detect damage
intensities higher than 2 min of vibrational bruising (i.e. loss of 1.34 units of L* value due
93
Chapter 4. Visible imaging
to damage) at an early stage of storage, i.e. immediately after damage or after 1 h of
storage. 1 min damage (i.e. loss of 0.67 units of L* value due to damage), on the
contrary, caused very slight discolouration and thus was harder to detect with the
webcams. The camera C 1 showed lower sensitivity than the rest and was not able to
differentiate between undamaged and low-intensity (i.e. MD<5, which corresponds to
loss of <3.34 unit of L value due to mechanical damage) damaged mushrooms
throughout a storage period of 12h. For higher damage intensities, C 1 was able to
discriminate within the first two hours of storage.
As can be seen from the results in Table 4.7, each camera recorded damage differently
depending on location and illumination conditions, and therefore the camera variable
was an important factor to consider.
Table 4.8 shows the time points after which each camera was not be able to
discriminate between MD and U mushrooms when using the DΔGSVref.w signal.
Table 4.8 Maximum time for MD/U discrimination of each camera and damage level,
LODDΔGSVref.w and D∆GSVref.w.
Damage level
based on
MD 1
MD 2
MD 3
MD 4
MD 5
MD 10
MD 20
C1
1h
1h
1h
2h
4h
5h
Camera
C2
C3
3h
4h
1h
3h
2h
5h
1h
3h
1h
6h
3h
3h
2h
C4
1h
1h
3h
2h
2h
5h
The empty cells indicate that derivatives were not useful at all for MD/U discrimination in
some cases. In general, derivatives were only useful for discrimination at very early
stages of storage, after which ∆GSVref.w could potentially start being used for
discrimination purposes.
94
Chapter 4. Visible imaging
Table 4.9 shows the time points after which each camera was able to discriminate
between MD and U mushrooms based on the GSVref.w signal, when no previous
knowledge was available.
Table 4.9 Time to discrimination of each camera and damage level, based on
Damage level
LODGSVref.w and (GSVref.w )MD .
MD 1
MD 2
MD 3
MD 4
MD 5
MD 10
MD 20
C1
>12h
>12h
>12h
0h
3h
0h
0h
Camera
C2
C3
>12h >12h
>12h >12h
7h
>12h
>12h >12h
>12h >12h
3h
>12h
0h
0h
C4
>12h
>12h
>12h
>12h
>12h
1h
4h
The results confirmed that the ability of the webcams to discriminate was significantly
lowered when using GSVref.w as a differentiating tool (i.e. when information prior to
damage was not available). In a real retail situation, most cameras would not be able to
detect low-intensity damage within the first 12 h of storage. The performance of the
cameras improved for higher damage intensities (i.e. loss of >6.70 units of L value due
to damage), with times to discrimination of 1 to 4 hours after damage in most cases.
4.4 Conclusions
It was previously found that colour change due to damage had a physiological basis
(enzyme activity together with mechanical damage promote browning) and that
discolouration increased with time. It was also reported that simple RGB imaging
systems were useful for the inspection of natural browning and brown spotting
processes in mushrooms.
95
Chapter 4. Visible imaging
The hypothesis tested in this chapter was that the colour change due to deliberately
induced browning may be monitored by visible imaging systems and employed to
identify mechanically damaged mushrooms.
Results presented in this chapter showed that such systems could be used to detect and
monitor browning of individual mushrooms in a batch. Under controlled illumination
conditions, the use of a reference tile minimised noise arising from fluctuations in the
illumination conditions.
Different image analysis criteria were tested based on i) Average Grayscale Value of the
mushroom caps (two different criteria, depending on the availability of information
previous to damage) and ii) Time Derivative of the Average Grayscale Value of the
mushroom caps. From the data obtained with undamaged control mushrooms, limits of
detection for individual devices and all criteria were estimated. These LODs were
employed to determine the time to discrimination of mechanically damaged mushrooms.
It was found that the use of an average grayscale value criteria was suitable to
discriminate mushrooms but required some time to pass after damage for discrimination
to be possible (over 12 h for low-intensity damage, and a maximum of 9 hours for highintensity damage). On the other hand, the use of time derivative based criteria enabled
the discrimination of damaged mushrooms soon after damage. However, the efficacy of
this criterion decreased after a certain storage time had passed.
Having information about the colour of mushrooms prior to purposely induced damage
allowed for earlier discrimination between damaged and undamaged mushrooms. The
time to discrimination was longer for almost all cameras and damage levels when this
information was not available.
Results from this study together with further investigations could potentially lead to the
development of an inexpensive and very simple monitoring system for the identification
of mechanically damaged mushrooms.
96
5. THE EFFECT OF MECHANICAL
DAMAGE ON THE ENZYME
ACTIVITY OF MUSHROOMS
Chapter 5. PPO kinetics
5.1 Introduction
Browning of mushrooms is the major cause of quality loss that accounts for reduction in
their market value. The development of brown colour is the consequence of a series of
complex reactions where enzymes such as polyphenol oxidases (PPOs) and
peroxidases (PODs) play a very important role (Bandyopadhyay et al., 2009).
The activity of PPOs has been reported to be more important than the activity of PODs
in mushroom browning (Mohapatra et al., 2008). PPOs are a family of copper containing
enzymes, which include laccases and catechol oxidases (also called tyrosinases). The
physiological importance of laccases in A. bisporus is very limited because of their low
levels (Turner, 1974), which suggests that catechol oxidases play a central role in the
enzymatic browning of mushrooms. Catechol oxidases (Figure 5.1) show cresolase
activity by hydroxylating some monophenols into o-diphenols, while their catecholase
activity is denoted by specifically oxidising o-diphenols into quinones (Jolivet et al.,
1998). Tyrosinases may be devoid of any cresolase activity; even when the latter is
present, its activity is usually much lower than the catecholase activity.
In the absence of extraneous oxidative enzymes, the measurement of the catecholase
or coupled cresolase-catecholase reactions may be colorimetric, since the o-quinone
product produces some colour (Mason, 1955; Mayer and Harel, 1979).
O
HO
R
R
O2
HO
R
O2
O
HO
monophenol
AH2
A
o-diphenol
AH2
A
o-quinone
Figure 5.1 Cresolase and catecholase activities of catechol oxidase.
These two reactions (i.e. hydroxylation and oxidation) are the first two steps in the
complex browning pathway that takes place in mushrooms when polyphenolic
substrates come into contact with PPOs. Although substrates and enzymes are
98
Chapter 5. PPO kinetics
originally separated by intracellular membranes, environmental stresses such as
mechanical damage can cause internal disorganisation and initiate the series of
reactions that lead to melanin formation. The higher levels of polyphenols and PPOs in
the pileipellis than in other parts of the fungi (Burton, 1988) explain why brown
discolouration resulting from mechanical damage is largely confined to the skin tissue of
the mushroom. Due to the large air spaces below the outer layer, this skin is only loosely
attached to the main flesh of the cap and its cells find very little support from the inner
layer of cells. Therefore, most of the energy of mechanical damage is absorbed by
pileipellis cells (Burton et al., 2002; Burton, 2004), resulting in a high degree of cell
disruption and discolouration of this tissue. The production of melanins constitute a
mechanism of defence and resistance to stresses such as free radicals, dehydration and
extreme temperatures (Halaouli et al., 2006), and contribute to the fungal cell-wall
resistance against hydrolytic enzymes in avoiding cellular lysis (Bell and Wheeler, 1986).
Consumer preference for fresh produce makes the management of PPO activity a
problem in the production, distribution and retail of fresh horticultural produce. An
appropriate knowledge of when this physiological response occurs, its intensity and its
characteristics is of importance to improve the shelf-life of these products. It is believed
that the increase in enzyme activity resulting in browning is produced as a physiological
response to environmental conditions. Studies in plants highlighted the role of PPO in
the resistance to stress (e.g. pathogen attack, wounding and herbivore attack). A major
difficulty is always to determine whether PPO is the direct reason for browning or
whether the browning reaction is a secondary result of other metabolic events (Mayer,
2006).
Many published papers report correlations between the levels of PPO activity and
environmental factors or changes during food processing or storage. In the case of
mushrooms, various studies modelled the postharvest browning as affected by the
storage conditions (Burton and Noble, 1993; Lukasse and Polderdijk, 2003; Bobelyn et
99
Chapter 5. PPO kinetics
al., 2006; Mohapatra et al., 2008).
Hughes (1958) reported that the response of
mushroom enzymes to environmental stress can be heightened when mechanical
damage has occurred. Gormley (1967) explained that vibrations taking place when
mushrooms are transported cause brown marks on the cap due to release of polyphenol
oxidase enzymes.
The aim of this study was to evaluate the effect of mechanical damage on the activity of
mushroom polyphenol oxidases. The interactive effects of mechanical damage
(Damage), storage temperature (Temperature) and storage time (Time) on the activity of
tyrosinase were also investigated.
5.2 Materials and methods
5.2.1 Sample preparation
For mushroom growing and harvest, refer to section 4.2.1. For details on the damage
protocol, refer to section 4.2.2.2.
After damage, mushrooms were placed on polystyrene trays in groups of 10 and overwrapped with PVC film. The trays were refrigerated (GRAM K400LU, Denmark).
Refrigeration conditions and damage levels will be described in section 5.2.2. At each
time point during storage, one tray of each damage level was randomly selected and
removed from storage approximately 15 min prior to analysis commencement. Wrapping
was removed and the mushrooms in each packet were randomly divided into two groups
of five mushrooms. Each of these subgroups was used to obtain one enzyme extract,
making two extract replicates per packet. Enzyme activity analysis was performed in
triplicate for each extract.
100
Chapter 5. PPO kinetics
5.2.2 Experimental design
Mushrooms of three damage levels [undamaged (U), 10 min damage (MD 10) and 20
min damage (MD 20)], stored at two different storage temperatures (5 ±1 °C and 10 ±1
°C), were monitored throughout five time points (days 0, 1, 2, 3 and 6 of storage). The
experiment was replicated, making two independent mushroom sets: Rep. 1 and Rep. 2.
549 mushrooms were used following a design with three factors (Damage, Temperature
and Time) with three, two and five levels, respectively. 114 experimental points were
obtained in total.
5.2.3 Extract preparation
Mushroom homogenates were prepared in duplicate from each sample tray, as follows:
5 g of the mushroom pileipellis (i.e. outer skin of the cap) were extracted using a sterile
sharp knife, chopped and placed in a Turrax homogeniser (ULTRA-TURRAX T25, Janke
& Kunkel IKA Labortechnik, Germany) in a 1:4 (w: v) ratio with 0.5 M phosphate buffer,
pH 6.5, containing 50 g/L polyvinylpirrolidone (Sigma-Aldrich). Homogenisation was
carried out for 1 min at 4°C and 8000 rpm. The homogenate was centrifuged (2K15
Laborzentrifugen, SIGMA, Germany) at 12,000 g for 35 min at 4°C. The supernatant
was collected by filtration through no. 1 Whatman paper and used as crude enzyme
extract. Extracts were kept at 4°C in the dark until spectrophotometric assay (within 2 h).
5.2.4 Spectrophotometric assay
PPO activity was measured spectrophotometrically by a modified method based on
those of Galeazzi et al. (1981) and Tan and Harris (Tan and Harris, 1995). The reaction
mixture contained 0.1 mL crude enzyme extract and 2.9 mL substrate solution (0.011
mol/L catechol (Sigma-Aldrich) as substrate in 0.05 mol/L phosphate buffer, pH 6.5).
The rate of catechol oxidation was followed at 410 nm (UV2 UV/vis Spectrometer,
101
Chapter 5. PPO kinetics
UNICAM, UK) and 25°C. The maximum slope of the straight-line section of the activity
curve was used to express the enzyme activity (EAU / g of fresh mushroom). A unit of
enzyme activity was defined as an increase of 0.001 absorbance units per minute.
Enzyme activity was calculated according to the following equation:
Activity =
Slope × 60 × V
v × 0.001× m
Equation 5.1
where V was the volume of filtrate obtained after centrifugation (mL), v was the
volume of filtrate in the cuvette (mL) and m was the initial mass of mushroom
pileipellis (g).
Enzyme activity was measured in triplicate for each mushroom extract and the average
value computed. Two activity values were obtained per mushroom tray.
5.2.5 Data analysis
The significance of the factors (Damage, Time) and their interactions (Damage*Time,
Temperature*Time, Damage*Temperature*Time) were tested by analysis of variance
(ANOVA). Differences were reported as significant to 95% Tukey’s HSD interval. R (R
Development Core Team, 2007) and SPSS (SPSS, 2006) were used for data analysis.
102
Chapter 5. PPO kinetics
5.3 Results and discussion
5.3.1 RGB images of undamaged and damaged mushrooms
Figure 5.2 shows representative RGB images of the three damage levels that were
under investigation in this study: (a) undamaged (U), (b) 10 min damage time (MD 10)
and (c) 20 min damage time (MD 20). The images were taken shortly after inducing the
damage.
(a)
(b)
(c)
Figure 5.2 Representative RGB images of various damage levels at day 0 of storage.
(a) undamaged mushroom; (b) mushroom damaged for 10 min (MD 10) and (c)
mushroom damaged for 20 min (MD 20).
U mushrooms were generally white in appearance, in spite of some unavoidable small
discoloured regions. Small regions of damage, typically caused by stress to the delicate
mushrooms surface, may have arisen during picking, packaging and transportation. MD
10 mushrooms showed some slight signs of browning on the mushroom surface, while
the mushrooms with the highest damage intensity (MD 20) exhibited a more uniform
brown discolouration over the entire mushroom surface.
As the storage time increased, slight signs of natural senescence became visible on
undamaged mushrooms, which exhibited isolated brown spots together with loss of cap
firmness. Textural changes are thought to be due to water loss and degradation of the
organic molecules forming the tissue membranes (Zhu et al., 2006). The slight brown
colour that damaged mushrooms exhibited just after being damaged developed further
103
Chapter 5. PPO kinetics
at initial stages of storage and became very evident by day 1-2 of storage. Damaged
mushroom caps were dark, dry and shrunken by the end of the storage period. To
illustrate these changes, representative RGB images of U (top row) and MD 10 (bottom
row) mushrooms stored at 5°C on days 0, 3 and 6 of storage are shown in Figure 5.3
(a)
(b)
(c)
Figure 5.3 Representative RGB images of various damage levels at different days of
storage.
Top row: undamaged (U) mushroom; Bottom row: mushroom damaged for 10 min (MD
10); (a) day 0 of storage at 5°C; (b) day 3 of storage at 5°C and (c) day 6 of storage at 5
at 5°C.
5.3.2 PPO kinetics
Figure 5.4 shows the PPO kinetics over the storage period, for all damage levels,
storage temperatures and repetitions.
The difference between damaged and undamaged mushroom enzyme levels is one of
the most obvious characteristics of these graphs. Enzyme activity levels of undamaged
mushrooms remained low (i.e. <20000 EAU / g) and relatively stable throughout the
experiments at both storage temperatures. This indicated that the storage conditions did
104
Chapter 5. PPO kinetics
neither induce stress nor trigger enzymatic expression over the course of 5 days storage
when no mechanical damage had occurred. The general pattern of tyrosinase activity of
damaged mushrooms, on the other hand, showed that activity increased with storage
time, reached a maximum level and then started to decrease gradually.
U
MD 10
MD 20
0
1
T = 10°C
Rep. 1
2
3
6
T = 10°C
Rep. 2
Enzyme Acvitity [EAU / g of pileipellis]
80000
60000
40000
20000
T = 5°C
Rep. 1
T = 5°C
Rep. 2
80000
60000
40000
20000
0
1
2
3
6
Storage time [days]
Figure 5.4 PPO enzyme activity as a function of storage time.
All damage levels, storage temperatures and repetitions are represented.
This trend was in agreement with results reported by Tao et al. (2007) and Mohapatra et
al. (2008), who monitored the enzyme activity of undamaged mushrooms over longer
storage periods. This showed that the enzymatic expression due to mushroom
105
Chapter 5. PPO kinetics
damaging was similar to the one that naturally occurs in senescent mushrooms, the
difference being that both the increase and the decrease of enzyme activity were
observed at earlier stages of storage in this study.
Results suggested that damaging induced enzyme activation/expression. The previously
reported stress-causing nature of mechanical damage (Hughes, 1958) may support this
finding. It is also well known that PPO levels of plants increase when these are exposed
to stress (Mayer, 1987). However, at this stage it is not clear whether the raised levels of
PPOs in the mushrooms were part of a response reaction to injury or simply the result of
cellular disruption. In any case, damaging may have caused the breakdown of
intracellular membranes, allowing phenols and enzymes to be in contact, starting a
series of reactions that give rise to brown colour. A study by Partington et al. (1999)
monitored the changes in the location of PPO in potatoes after impact injury and found
that the enzyme, initially located in starch grains and the cytoplasm of undamaged
tissue, became more distributed in the vacuolar region several hours after mechanical
bruising.
It is also unclear whether damage caused a transition from the latent to active state of
the enzymes or just an over expression of the active form of PPO. Providing an
explanation of the mechanism behind the results observed goes beyond the scope of
this investigation. Fundamental biochemical studies could provide this information.
As can also be seen in Figure 5.4, the first batch of mushrooms showed higher
enzymatic levels than mushrooms belonging to the second batch. ANOVA results
showed that the two mushroom sets (i.e. Rep.1 and 2) behaved significantly differently
(p<0.05). Possible reasons contributing to these variations are:
•
Biological variation: postharvest behaviour of horticultural products is known to
be affected by biological variation. Batch-to-batch variability was reported to be
an important issue in A. bisporus mushroom quality management, representing
106
Chapter 5. PPO kinetics
an average of 22% of the variability observed in various quality indicators
(Aguirre et al., 2008c). Burton (2004) reported that mushroom bruisability varied
from crop to crop, which could be due to genetics and/or mushroom growing
conditions. Mohapatra et al. (2008) observed 30% to 41% variability in enzyme
activity measurements and attributed it to batch-to-batch variability.
•
Enzyme assay: the assay of tyrosinase activity raises difficult questions, since
this enzyme catalyses two different reactions. Although cresolase activity has
usually been indicated as being much lower than catecholase activity, VamosVigyazo (1981) reported that catecholase/cresolase ratio can vary from 40:1 to
1:1. Monitoring PPO in crude extracts can be further complicated by the
presence of laccase and peroxidase activities in these extracts. Finally, owing to
their reactivity, o-quinones are prone to undergo side-reactions, resulting in the
non-linearity of plots pfabsorbance vs. quinone production.
Due to significant differences observed between the two mushroom sets, further
statistical analysis was performed separately for Rep.1 and Rep.2. The two sets will be
treated as different sample groups in further discussion.
5.3.3 The effect of Damage, Temperature and Day on enzyme activity
ANOVA of the factors and their interactions was performed as described in section 5.2.5
and results are shown in Table 5.1.
Table 5.1 ANOVA of Damage, Temperature and Time factors and their interactions on
Rep. 1 data.
Factor
Damage
Time
Damage*Time
Temperature*Time
Damage*Temperature*Time
Residuals
DF
2
4
8
5
10
30
Sum sq
1.81e+10
5.93e+09
2.66e+09
3.00e+09
1.79e+09
4.2027e+08
107
Mean Sq
9.05e+09
1.48e+09
4.57e+08
5.99e+08
1.79e+08
1.40e+07
F value
645.70
105.77
32.64
42.75
12.81
Pf(>F)
<2.2e-16
<2.2e-16
9.45e-13
9.23e-13
2.65e-08
Chapter 5. PPO kinetics
The significance of the Damage factor (Figure 5.5.a) confirmed that mechanical damage
had an effect on tyrosinase activity values. For all batches and storage temperatures,
higher activity levels of PPO enzymes were observed in damaged mushrooms than in
undamaged mushrooms. Possible explanations for this phenomenon have been
discussed in section 5.3.2. Pairwise comparisons among the different levels of the
Damage factor showed that the difference in PPO activity levels of MD 10 and MD 20
damage levels was not significant. This could imply that the stress caused by MD 10
damage level was sufficiently high to generate the maximum expression of enzyme
activity and further damage did not contribute to further activation of PPO. The Day
factor also had a significant effect on enzyme activity levels (Figure 5.5.b), which
confirmed that the changes in enzyme activity described in section 5.3.2 were
significant. Pairwise comparisons among the means of different days showed that
differences were significant between days 0 and 1 of storage, days 1 and 2 of storage
and days 3 and 6 of storage, but not between days 2 and 3 of storage. This was in
agreement with the shape of the kinetics of MD mushrooms, where days 2 and 3
appeared on top of the curve.
(b)
Enzyme Acvitity [EAU / g of pileipellis]
Enzyme Acvitity [EAU / g of pileipellis]
(a)
50000
40000
30000
20000
10000
U
MD 10
MD 20
45000
40000
35000
30000
25000
20000
15000
0
Damage
1
2
3
Day
Figure 5.5 Effect of first-order factors on the enzyme activity of Rep.1 mushrooms.
(a) Damage and (b) Day
108
6
Chapter 5. PPO kinetics
Second order Damage*Time interaction also had a significant effect on enzyme activity
levels, which meant that Damage influenced the kinetics of PPO significantly (i.e. the
extent of damage had a significant effect on the way time affected activity). This effect is
shown in Figure 5.6.
Enzyme Acvitity [EAU / g of pileipellis]
0
Damage : U
1
2
3
6
Damage : MD 10
Damage : MD 20
60000
50000
40000
30000
20000
10000
0
1
2
3
6
0
1
2
3
6
Day
Figure 5.6 Effect of second-order Damage*Time interaction on the enzyme activity of
Rep.1 mushrooms.
Day-to-day variation of U mushrooms was small in comparison to that of MD
mushrooms, but still significant. PPO activity of MD mushrooms increased at the
beginning of the storage period, reached a maximum level and started to decrease
gradually. The increase did not happen immediately after damage (mean PPO activities
of U and MD samples were not significantly different on day 0) but at a later stage during
storage. This suggests that mushroom physiology requires some time to activate and
respond to stress induced by damage. By day 1, all MD samples showed significantly
higher activity values than the U mushrooms, and this trend remained throughout the
duration of the experiment. The activity of MD mushrooms exhibited a maximum peak
on day 1, 2 or 3 of storage and decreased gradually until the end of the monitoring.
Presumably, after the peak, the metabolism would be more limited by the depletion of
nutrients and therefore the stress levels of enzyme activity might not be maintained.
109
Chapter 5. PPO kinetics
The significance of the Temperature*Time interaction showed that storage temperature
influenced PPO kinetics too (i.e. the temperature had a significant effect on the way time
affected activity). PPO activity values were higher when mushrooms were stored at
10°C, and the difference between the two storage temperatures was greater for
mushrooms belonging to the first batch. Mohapatra et al. (2008) found that enzyme
activity underwent a transition to “stressed” levels faster when the storage temperature
was increased, but did not show that the enzyme activity depended on storage
temperature. Their study concluded that storage temperature ranging from 3°C to 15°C
did not affect enzyme activity.
The
significance
of
the
third
order
interaction
of
the
three
factors,
Damage*Temperature*Time (shown in Figure 5.7), indicated that Temperature
influenced the way Damage affected the day-to-day evolution of PPO too (i.e. different
storage temperature influenced the way damage activated PPO kinetics differently). The
following figure serves as a summary of all the previously mentioned effects and
observations.
110
Chapter 5. PPO kinetics
0
1
2
3
6
Temperature : 10°C Temperature : 10°C Temperature : 10°C
Damage : U
Damage : MD 10 Damage : MD 20
Enzyme Acvitity [EAU / g of pileipellis]
80000
60000
40000
20000
Temperature : 5°C Temperature : 5°C Temperature : 5°C
Damage : U
Damage : MD 10 Damage : MD 20
0
80000
60000
40000
20000
0
0
1
2
3
6
0
1
2
3
6
Day
Figure 5.7 Effect of the third-order interaction (i.e. Damage*Temperature*Time) on the
enzyme activity of Rep. 1 mushrooms.
5.4 Conclusions
It was previously reported that tyrosinase plays a crucial role in mushroom browning. It
was also found that, during postharvest storage, the physiological response of
mushrooms to environmental stress can be more intense when mechanical damage has
taken place.
This piece of research studied the effect of mechanical damage on the enzymatic
activity of tyrosinase. Second and third order interaction with storage temperature and
time were also investigated.
Damage by vibrational shaking intended to simulate transport conditions was found to
cause an increase in the expression of polyphenol oxidase enzyme, which is the main
causal agent for the development of brown colour. The extent of damage to which
111
Chapter 5. PPO kinetics
mushrooms were subjected was not significant in this study. This suggested there was
an upper limit impact level where enzyme expression is not further affected by damage.
PPO kinetics of mechanically damaged mushrooms showed that the maximum enzyme
expression happened at least two days after injury, which showed that the physiological
response to stress was not immediate.
The storage temperature had a lower but still significant effect on mushroom kinetics,
which indicated that temperature needed to be controlled in the prevention of browning
development on mushroom caps.
Results from this study highlighted the importance of mechanical damage to the quality
of transported mushrooms, even when the effects of the injury are judged to be within
acceptable levels by the consumer. The physiological response happened not
immediately after damage but later during storage, and the control of storage
temperature might limit its extent but would not prevent it.
112
6. PREDICTION OF PPO ACTIVITY
USING Vis-NIR HYPERSPECTRAL
IMAGING ON MUSHROOM CAPS
Chapter 6. Prediction of PPO using HSI
6.1 Introduction
The role of PPO enzymes in the browning of mushrooms and the need to control their
activity was highlighted in previous chapters.
Current methods to perform tyrosinase assays include spectrophotometric, manometric,
radiometric and chronometric techniques (Falguera et al., 2010), as well as the use of
oxygen electrodes that follow the oxygen uptake throughout oxidation reactions (Jolivet
et al., 1998). All of these techniques require previous preparation of crude extracts,
which adds labour and time cost to the analysis. Rapid spectroscopic techniques show
potential for replacement of slow and/or expensive analytical measurements while
retaining sufficient accuracy (Zeaiter et al., 2005).
Hyperspectral imaging (HSI) combines conventional imaging with spectroscopy to
simultaneously acquire spatial and spectral information from an object. This technology
has recently emerged as a powerful process analytical tool for food analysis (Gowen et
al., 2007) and offers several advantages over traditional methods, such as minimal
sample preparation, its non-destructive nature and fast acquisition times (Elmasry and
Sun, 2010). Hyperspectral imaging data are three-dimensional blocks of data called
hypercubes, comprising two-spatial and one spectral dimension (Figure 6.1).
Hypercube classification enables the identification of regions with similar spectral
characteristics. Since regions of a sample with similar spectral properties have similar
chemical composition, hypercube classification allows for the visualisation of the
concentration and spatial distribution of biochemical constituents over the sample. Due
to the large size of hypercubes, multivariate analytical tools, such as stepwise multiple
linear regression (MLR), principal component regression (PCR) and partial least squares
regression (PLSR) are usually employed for hyperspectral data mining and identification
of key wavelengths for the development of automated multispectral sensors.
114
Chapter 6. Prediction of PPO using HSI
Figure 6.1 Schematic representation of a hyperspectral imaging hypercube.
Recent studies have demonstrated HSI to be a useful technology for the investigation of
mushroom quality deterioration, bruise and freeze damage detection, prediction of
moisture content and shelf-life evaluation (Gowen et al., 2008b; 2008c; 2009;
Taghizadeh et al., 2009; 2010). So far, hyperspectral imaging has not been employed to
study the activity of enzymes in mushrooms.
Short wavelength infrared hyperspectral imaging was recently used to predict α-amylase
activity at early germination stages in two classes of wheat kernels and R2 values of 0.54
and 0.73, respectively, were achieved (Xing et al., 2009). Given that PPOs play a key
role in the mushroom browning process and that current extraction activity measurement
techniques are time consuming (as an example, in this study, 1.5-2 hours were needed
to obtain an extract and measure its activity), it would be desirable to have a fast and
non-destructive system that could estimate enzyme activity on mushroom caps. The
development of a hyperspectral imaging system with the ability to make simultaneous
predictions on multiple mushroom caps could enable faster detection of produce likely to
lose market value and hence reduce economical losses in the industry.
The aim of the present study was to investigate the potential of Vis-NIR (445-945 nm)
hyperspectral imaging for the prediction of PPO enzymes activity on mushroom caps.
115
Chapter 6. Prediction of PPO using HSI
6.2 Material and methods
6.2.1 Sample preparation
Refer to section 5.2.1.
All the mushrooms in each package were scanned with the hyperspectral imaging
equipment prior to the enzyme extraction procedure. In this way, one scan per
mushroom and two extracts per package were obtained at each sampling point.
A total number of 549 mushrooms were scanned and 114 extracts were obtained.
6.2.2 Experimental design
Refer to section 5.2.2.
6.2.3 Hyperspectral imaging
6.2.3.1 Acquisition system
Hyperspectral images were obtained using a pushbroom line-scanning HSI instrument
(DV Optics Ltd, Padua, Italy, Figure 6.2). The instrument comprised a moving table,
illumination source (150 W halogen lamp source attached to a fibre optic line light
positioned parallel to the moving table), mirror, objective lens (16 mm focal length),
Specim V10E spectrograph (Spectral Imaging Ltd, Oulu, Finland) operating in the
wavelength range of 400-1000 nm (spectroscopic resolution of 5 nm), CCD camera
(Basler A312f, effective resolution of 580 × 580 pixels by 12 bits), acquisition software
(SpectralScanner, DV Optics, Padua, Italy) and a personal computer. A cylindrical
diffuser was placed in front of the fibre optic line light to produce a diffuse light source. In
this study, only spectral data within the wavelength range of 445-945 nm were used, as
116
Chapter 6. Prediction of PPO using HSI
beyond this range the noise level of the camera detector was high and affected the
signal efficiency.
Lens
Spectrograph
Mirror
CCD Camera
Sample
Light source
Optical fibres
Moving table
Figure 6.2 Pushbroom hyperspectral imaging system
(image courtesy of P. Menesatti, CRA-ISMA, Rome, Italy).
6.2.3.2 Reflectance calibration
Reflectance calibration was carried out prior to mushroom image acquisition in order to
account for the background spectral response of the instrument and the “dark” camera
response. The bright response (‘W’) was obtained by collecting a hypercube from a
uniform white ceramic tile; the dark response (‘dark’) was acquired by turning off the light
source, completely covering the lens with its cap and recording the camera response.
The corrected reflectance value (‘R’) was calculated from the measured signal (’I’) on a
pixel-by-pixel basis as shown in Equation 6.1 (Ariana et al., 2006):
Ri =
(Ii − darki )
(Wi − darki )
Equation 6.1
where i is the pixel index, i.e. i =1,2,3,…,n and n is the total number of pixels.
117
Chapter 6. Prediction of PPO using HSI
6.2.4 Enzyme extraction
Mushroom homogenates were prepared in duplicate from each mushroom tray. Refer to
section 5.2.3.
6.2.5 Image processing and data analysis
Reflectance data were saved in ENVI header format using the acquisition software and
then exported to MATLAB (MathWorks, 2007).
6.2.5.1 Masking
A masking step was carried out to separate the mushroom pixels from the background.
The mask was created by thresholding the mushroom image at 940 nm, where a pixel
threshold value of 0.2 was used to segment the mushroom from the background. All
background regions were set to zero and the non-zero elements of the image were used
to extract one mean spectrum for each mushroom.
6.3.5.2 False RGB images
False RGB images were obtained by extracting mushroom images at wavelengths 460
nm (blue), 545 nm (green) and 645 nm (red) and stacking them.
6.2.5.3 Model building
One of the main challenges involved in building predictive models with hyperspectral
image data is that such images contain a vast amount of spectral data, whilst only one
or a few measurements of the variable of interest can be taken for each sample studied.
In this particular study, the reference method for enzyme extraction involved using the
pileipellis of three to five mushrooms to obtain one single enzyme extract. Consequently,
three to five hyperspectral images were matched with one single enzyme activity value
in regression modelling.
118
Chapter 6. Prediction of PPO using HSI
When developing regression models with hyperspectral data, it is common practice to
extract the mean spectrum of each sample and use it to build a prediction model to
estimate an attribute (Burger and Geladi, 2006). With that approach in mind, two
different modelling strategies were used:
i) Strategy 1: The first strategy extracted the mean spectrum of each mushroom
and assigned the same enzyme activity value to all the mushrooms used in
obtaining one particular extract. A training set of nTRAIN_1=280 and a test set of
nTEST_1=269 were used for this strategy.
ii) Strategy 2: The second strategy computed the mean spectra of all the
mushrooms used to obtain one enzyme extract and assigned the enzyme activity
value of that extract to the resulting spectrum. A training set of nTRAIN_2=60 and a
test set of nTEST_2=54 were used for this strategy.
Both modelling strategies are shown in Figure 6.3.
EXPERIMENTAL
DATA
1 EAU value
j spectra
Day
Damage
Package
Extract
PPO
activity
Mushroom
λ1
λ2
λ3
…
λi
1
R1,1
R1,2
R 1,3
…
R1,i
0
U
1
1
EAU1
…
…
…
…
…
…
j
Rj,1
Rj,2
Rj,3
…
Rj,i
R1,i
0
U
1
2
EAU2
1
R1,1
R1,2
R 1,3
…
…
…
…
…
…
…
k
Rk,1
Rk,2
R k,3
…
Rk,i
Mushroom
λ1
λ2
λ3
…
λi
STRATEGY 1
Day
Damage
Package
Extract
PPO
activity
EAU1
1
R1,1
R1,2
R1,3
…
R1,i
EAU value * j times
j spectra
0
U
1
1
EAU1
…
…
…
…
…
…
0
U
1
2
EAU1
j
Rj,1
R j,2
Rj,3
…
Rj,i
EAU2
1
R1,1
R1,2
R1,3
…
R1,i
EAU2
…
…
…
…
…
…
EAU2
k
Rk,1
Rk,2
Rk,3
…
Rk,i
STRATEGY 2
Day
Damage
Package
Extract
PPO
activity
Mushroom
λ1
λ2
λ3
…
λi
1 EAU value
1 average spectrum
0
U
1
1
EAU1
1-j
R1
R2
R3
…
Ri
0
U
1
2
EAU2
1-k
R1
R2
R3
…
Ri
Figure 6.3 Diagram showing modelling strategies 1 and 2.
119
Chapter 6. Prediction of PPO using HSI
The following spectral pre-processing methods were used in order to remove nonchemical biases, such as scattering effects and variations arising from mushroom
surface curvature, from the spectral information: standard normal variate (SNV) (Barnes
et al., 1989) and multiplicative scatter correction (MSC) (Geladi et al., 1985). The
rationale behind these pre-treatments was introduced in section 2.6. MSC aims to
reduce the effects of scattering in a set of spectra by performing linear regression on a
“target” spectrum. Two different target spectra led to two different MSC methods:
i) “Set MSC”, where the mean spectrum of each mushroom was corrected using
the mean spectrum of the data set as the target spectrum.
ii) “Sample MSC”, where the spectrum of each pixel in a mushroom was corrected
using the mean spectrum of that mushroom as the target spectrum. The mean
sample MSC corrected spectrum for each mushroom was obtained and used for
the model.
To improve normality of the distribution of the reference variable, enzyme activity values
were transformed into natural logarithmic units and mean centred.
Three regression methods were used to build models for enzyme activity prediction:
i) Multiple linear regression (MLR): this was the most simple approach for model
building. Optimal wavelengths for enzyme activity prediction were selected by the
“forward” method in best subsets stepwise linear regression using the “leaps”
package in R (R Development Core Team, 2007). Forward stepwise regression
is based on the procedure of sequentially introducing wavelengths into the
model, one at a time (Chong and Jun, 2005). Multicollinearity of predictor
variables is problematic for MLR models based directly on spectroscopic values,
resulting in unstable model predictions (Fekedulegn et al., 2002). The variance
inflation factor (VIF) measures how much the variance of an estimated
regression coefficient is increased because of collinearity. It is an index
120
Chapter 6. Prediction of PPO using HSI
commonly used to measure the collinearity between variables in regression
models: typically, predictor variables with VIF>10 are considered to be highly
correlated. In order to test the predictor wavelengths for multicollinearity, the VIF
of each predictor was calculated using the “DAAG” package in R (R
Development Core Team, 2007).
ii) Principal component regression (PCR): principal component analysis (PCA)
reduces the dimensionality of spectral data by transforming them into principal
component scores in order of decreasing variance. The autoscaled matrix of
spectral values was transformed into PC space by representing the original data
in the directions defined by orthogonal eigenvectors using R (R Development
Core Team, 2007). PCR models were developed using PC space scores instead
of wavelength space values. Analysis of variance (ANOVA) was employed using
R (R Development Core Team, 2007) to compare models with increasing
number of PCs. The decision on the number of PCs to be taken for each model
was made based upon ANOVA test results. Only significant components
(p<0.05) were included in the model.
iii) Partial least squares regression (PLSR): this technique is commonly used when
predicting a response from many measured variables which may be collinear.
PLSR was applied using the “pls” package in R (R Development Core Team,
2007). Leave-one-out (LOO) cross-validation was used on the training set. LOO
takes out one of the samples in the set, does the calibration on the remaining
samples and uses the calibration to predict the sample that was left out. The
exercise is repeated leaving out another sample and getting another prediction
and it is repeated until all the samples have been left out and predicted.
Performance of the prediction models was evaluated using the root of the mean
of the sum of squared differences between predicted and measured enzyme
activity values of the training set (RMSECV, Equation 6.2) and the number of
latent variables required (# LV). The optimal number of latent variables for
121
Chapter 6. Prediction of PPO using HSI
inclusion in the PLSR models was estimated using the method described by
Martens and Dardenne (1998).
∑(
n
RMSECV =
i =1
y − y
i
i
)
2
df
Equation 6.2
where y i is the predicted enzyme activity value of the i th observation, yi is
the measured enzyme activity value of the i th observation and df are the
degrees of freedom.
The experiment was carried out two times, making two independent mushroom sets: a
training set (nTRAIN_1=280 mushrooms and nTRAIN_2=60 extracts) and a test set
(nTEST_1=269 mushrooms and nTEST_2=54 extract). Overall, 549 mushrooms were used to
obtain 114 extracts in total. All of the models were built on training sets and then applied
to independent test sets of samples. The ratio of percentage deviation (RPD), which is
the ratio of the standard deviation of the laboratory measured (reference) data to the
root-mean-square of cross-validation (RPDTRAIN) or root-mean-square error of prediction
(RPDTEST) (Williams, 1987), was used to assess model performance.
122
Chapter 6. Prediction of PPO using HSI
m
n
SDTRAIN =
∑ ( yi − yTRAIN )2
i =1
SDTEST =
n
Equation 6.3
∑(y
j =1
j
− yTEST ) 2
m
Equation 6.5
∑(
m
RMSEP =
j =1
y − y
j
j
)
2
df
Equation 6.6
RPDTRAIN =
SDTRAIN
RMSECV
RPDTEST =
SDTEST
RMSEP
Equation 6.7
Equation 6.4
where yTRAIN is the average enzyme activity value of all the observations ( n ) in the
training set; m is the number of samples in the test set; y j is the measured enzyme
activity value of the j th observation; yTEST is the average enzyme activity value of all
the observations ( m ) in the test set and l
y j is the predicted enzyme activity value of
the j th observation.
Twenty four models (2 strategies × 4 pre-treatments × 3 regression methods) were
classified in terms of their transferability, following the criteria outlined by Viscarra
Rossel et al. (2007), based on which RPDTEST<1.0 indicates very poor model/predictions
and their use is not recommended; 1.0<RPDTEST<1.4 indicates poor model/predictions
where only high and low values are distinguishable; 1.4<RPDTEST<1.8 indicates fair
model/predictions that may be used for assessment and correlation; 1.8<RPDTEST<2.0
indicates good models/predictions where quantitative predictions are possible;
2.00<RPDTEST<2.5 indicates very good, quantitative model/predictions and RPDTEST>2.5
indicates excellent model/predictions.
123
Chapter 6. Prediction of PPO using HSI
6.2.5.4 Prediction maps
The two models whose performance was found to be best were selected and applied to
each pixel in the hypercube data of individual mushrooms. This enabled the generation
of virtual prediction images for enzyme activity.
6.3 Results and discussion
6.3.1 Spectral variation arising from mushroom shape
Curvature inherent in the morphology of a sample introduces spectral variability into its
hyperspectral image, and mushrooms are a good example of that. This can be seen in a
typical hyperspectral image of a mushroom cap, as shown in Figure 6.4. In order to
assess the effect of curvature, the mean hyperspectral image of this mushroom (Figure
6.4.a) was grouped into regions of similar spectral characteristics using k-mean
clustering (Gowen and O'Donnell, 2009). The resultant regions, shown in Figure 6.4.b,
formed concentric ovals and exhibited a decreasing trend in reflectance intensity from
the centre of the mushroom to its edge. Mean and standard deviation spectra from each
region are shown in Figure 6.4.c. As can be seen in this figure, the curvature of the
mushrooms surface caused a scaling difference in the spectra: the intensity of the
spectra decreased as the mushroom edge is approached. The overall spectra profile
was similar for all regions.
124
Chapter 6. Prediction of PPO using HSI
(a)
(b)
(c)
0.8
0.7
Reflectance [ ]
0.6
0.5
0.4
0.3
0.2
0.1
0
450
500
550
600
650
700
750
800
850
900
Wavelength [nm]
Figure 6.4 Typical hyperspectral image of the surface of a mushrooms.
(a) mean intensity image; (b) segmentation of the previous into regions of similar light
intensity using k-mean clustering; (c) corresponding mean and standard deviation
reflectance spectra for each region, showing the effect of curvature on spectral
response.
This effect is thought to be partly due to the relative difference in path length from
different points of the curved mushroom surface to the detector (Gowen et al., 2008b):
points on the mushroom surface that are nearer to the detector result in higher
reflectance values than points that are further from the detector, such as those on the
edge. Non-uniform lighting over the curved surface adds to spectral variation in regions
of similar composition (Aleixos et al., 2002).
The inherent curvature of samples is particularly problematic when surface classification
is sought (e.g. defect or contaminant detection), as regions of similar composition but
different location over the surface could potentially be classified as different due to the
differences in their spectral amplitude. Several solutions have been proposed in the
literature to overcome spectral variations over spherical surfaces (Gómez-Sanchis et al.,
2008b), which include improvement of lighting conditions, adding image pre-processing
steps and eroding the outline of the spheres.
Following common practice in the development of regression models with HSI data
(Burger and Geladi, 2006), this study obtained the mean spectrum of all the pixels in a
mushroom. Such a solution does not compensate for spectral differences due to
125
Chapter 6. Prediction of PPO using HSI
curvature but assumes that the effect of the morphology will be similar in all the
samples.
6.3.2 Spectra
Mean reflectance spectra obtained from the hyperspectral imaging data of undamaged
mushrooms (U), mushrooms damaged for 10 min (MD 10) and mushrooms damaged for
20 min (MD 20) are shown in Figure 6.5.a. The mean reflectance of damaged samples
was lower than the mean reflectance of non-damaged mushrooms over the entire
spectral region. Bruising due to mechanical damage was expected to have led to loss of
whiteness and lightness (L*) and therefore lower reflectance values. A remarkable
difference in intensity was observed between U and MD 20 mushrooms, whereas the
intensity of MD 10 spectra was intermediate between U and MD 20. Broad spectra in the
visible-near infrared wavelength range are characteristic of undamaged mushrooms,
corresponding to their white appearance (Gowen et al., 2008c). The greatest differences
in shape between bruised and non-bruised samples arose in the 600-800 nm region,
where undamaged mushrooms exhibited broader spectral features than the damaged
mushrooms. The spectral differences mentioned above could be related to the formation
of brown pigments (Gowen et al., 2009) mainly melanins, which derive from enzymecatalysed oxidation products called quinones.
6.3.2 Enzyme activity
The average polyphenol oxidase enzyme activity of each mushroom group is shown in
Figure 6.5.b. The higher activity values observed in bruised mushrooms are justified by
the effect mechanical damage has on enzyme expression (refer to Chapter 5).
126
Chapter 6. Prediction of PPO using HSI
(b)
(a)
1
12
0.9
11.5
Reflectance [ ]
ln(PPO activity [EAU / g])
U
MD 10
MD 20
0.8
0.7
0.6
0.5
0.4
0.3
0.2
10.5
10
9.5
9
8.5
8
7.5
0.1
0
450
11
500
550
600
650
700
750
800
850
7
900
Wavelength [nm]
U
MD10
MD 20
Damage level
Figure 6.5 Summary of spectral and enzymatic characteristics of the data set.
(a) Mean raw reflectance spectra for mushroom at different damage levels. (b) Mean
±standard deviation of PPO activity as a function of damage level.
6.3.3 Modelling
VIF was greater than 10 for every MLR model built with more than two wavelengths.
Therefore, MLR models that used only two wavelengths were considered for further
analysis. In the case of PCR models, the inclusion of the third PC was not always
significant (p<0.05) so 2 and 3 PC models were considered for further sections. For all
PLSR models, 2 was the optimal number of latent variables to include in the model.
Previous studies in the field employed models that performed well using a low number of
wavelengths (Gowen et al., 2008c), principal components (Gowen et al., 2008b; 2009)
or PLS latent variables (Esquerre et al., 2009).
Model performance in terms of RPD is shown in Table 6.1. RPDTRAIN is a measure of
model performance within the model training data set and RPDTEST indicates how the
model performed when applied to an independent model testing data set. RPDTEST was
considered to be more adequate to assess model performance and further sections of
this paper will focus only on RPDTEST values.
127
Table 6.1 Ratio percentage deviation (RPD) for the different model strategies, spectral pre-processing techniques and chemometric
methods under investigation.
MLR
Strategy Pre-treatment
1*
2*
128
None
SNV
Set MSC
Sample MSC
None
SNV
Set MSC
Sample MSC
λ (nm)
450, 945
835, 560
835, 545
465, 945
470, 945
450, 465
450, 575
495, 945
PCR
RPDTRAIN RPDTEST # PCs
1.87
1.02
1.52
1.91
1.28
1.22
1.15
1.35
1.47
1.06
1.14
1.43
1.16
1.07
0.89
1.22
3
2
2
3
2
1
1
2
PLSR
RPDTRAIN
RPDTEST
2.01
1.71
1.65
2.01
1.27
1.17
1.17
1.35
2.13
1.84
1.77
2.13
1.30
1.20
1.16
1.27
# LVs RPDTRAIN RPDTEST
2
2
2
2
2
2
2
2
1.95
1.63
1.62
1.95
1.25
1.17
1.17
1.33
1.16
1.22
1.20
1.14
0.97
0.85
0.81
1.22
MLR: multiple linear regression; PCR: principal component regression; PLSR: partial least squares regression; SNV: standard normal
variate; MSC: multiple scatter correction; # PCs: number of principal components; # LVs: number of latent variables.
RPDTEST<1.0 = “very poor”; 1.0<RPDTEST<1.4 = “poor”; 1.4<RPDTEST<1.8 = “fair”; 1.8<RPDTEST<2.0 = “good”; 2.00<RPDTEST<2.5 = “very
good” and RPDTEST>2.5 = “excellent”.
Chapter 6. Prediction of PPO using HSI
6.3.3.1 Strategy
Overall, more transferable models were obtained when strategy 1 was employed. As
can be seen in Table 6.1,
for any pre-processing and chemometric technique
combination, the RPD obtained under model strategy 1 (i.e. when the mean spectrum of
each mushroom was extracted and the same enzyme activity value was assigned to all
the mushrooms used for one extract) was higher than the RPD obtained under model
strategy 2 (i.e. when the mean spectra of all the mushrooms used to obtain one enzyme
extract was computed and the enzyme activity value of that extract was assigned to the
resulting spectrum). In fact, strategy 2 only gave “poor” or “very poor” predictive models,
whose RPDTEST ranged from 0.81 to 1.3. This could be because when the mean
spectrum was computed for an extract under strategy 2, some features arising from the
original spectral variability of the mushrooms within that extract might have been lost.
This would result in partial loss of their transferability and a decrease in RPDTEST values.
6.3.4.2 Pre-treatment
For MLR, raw reflectance spectral data and sample MSC corrected reflectance spectra
led to better performance models than SNV or set MSC spectra. The better models were
“fair” and the worse ones were “poor” (according to the previously mentioned RPD
classification) and therefore discarded. Similar trends were observed in PCR models,
where “very good” models were obtained with raw reflectance and sample MSC
corrected reflectance spectra (RPDTEST=2.13 with 3 PCs), a “good” model with SNV pretreated reflectance data (RPDTEST=1.84 with 2 PCs) and a “fair” model with set MSC
corrected reflectance spectra (RPDTEST=1.77 with 2 PCs). The number of PCs was lower
in the case of SNV and set MSC but adding a third one did not significantly improve
model performance or RPDTEST. For PLSR models, all pre-treatments resulted in “poor”
models, whose highest RPDTEST was 1.22.
129
Chapter 6. Prediction of PPO using HSI
6.3.4.4 Regression method
Under strategy 1, PCR models performed better than MLR or PLSR models for all of the
pre-treatments. This happened for both training and test sets. The performance of MLR
and PLSR models for the test set was not as good as it was for the training set, but that
did not happen for PCR models, where RPDTEST values were higher than RPDTRAIN
values.
Under model strategy 2, all chemometric methods performed similarly for the training
set. For the test set, PCR models performed better than MLR or PLSR but still “poor”
predictions (RPDTEST<1.3) were obtained.
PCR models developed on raw reflectance and sample MSC corrected reflectance data
under model strategy 1 were selected as the best models and used in further analysis.
The
coefficient
of
determination
and
root
mean-squared
error
of
cross-
validation/prediction for these models were: R2TRAIN_1=0.75, RMSECV=0.38 [ln(EAU/g)],
R2TEST_1=0.78 and RMSEP=0.30 [ln(EAU/g)]. Root mean-squared errors of crossvalidation/prediction are frequently used to assess the performance of the regression
and low values indicate good models.
In Figure 6.6, enzyme activity values predicted by one of the selected models (model
strategy 1, PCR, raw reflectance data) are plotted against experimental enzyme activity
values, for (a) training and (b) test sets, respectively.
130
Chapter 6. Prediction of PPO using HSI
(b)
11.0
10.0
RMSEP = 0.30
9.5
11.0
8.5
9.0
9.5
10.0
RMSECV = 0.38
2
RTEST = 0.78
9.0
ln(Predicted PPO Activity[EAU/g])
2
RTRAIN = 0.75
8.5
ln(Predicted PPO Activity[EAU/g])
(a)
8.5
9.0
9.5
10.0 10.5
11.0
8.5
11.5
9.0
9.5
10.0 10.5
11.0
11.5
ln(PPO Activity[EAU/g])
ln(PPO Activity[EAU/g])
Figure 6.6 Predicted PPO vs measured PPO activity.
Strategy 1, raw spectra, 3 PC PCR model. (a) training data set and (b) test data set.
The range of measured reference values was wider in the training set than in the test
set, where PPO activity levels were, in general terms, lower and confined to a narrower
range of values. This scenario is not optimal for model testing but it is common when
dealing with horticultural products, whose postharvest behaviour is known to be affected
by biological variation. This variation was previously discussed in section 5.3.2. Some
vertical scattering can be seen in this figure too, indicating variability in predicted values
for mushrooms with similar reference enzyme activities. This would explain the relatively
low values of the coefficients of determination obtained (R2TRAIN_1 =0.75 and R2TEST_1
=0.78). The horizontal scattering is mainly attributable to mushroom to mushroom
variability.
6.3.4.5 Prediction maps
Hyperspectral imaging has the ability to map the spatial distribution of components on a
sample. The two selected models (model strategy 1, PCR, non-treated reflectance and
sample MSC corrected reflectance) were applied to each pixel in the hypercube data of
131
Chapter 6. Prediction of PPO using HSI
individual mushrooms and that enabled the generation of virtual prediction images for
enzyme activity. In such images, the grayscale intensity is related to the value of the
predicted enzyme activity in different regions of the mushroom cap: the lighter the
colour, the higher the predicted activity value.
Figure 6.7 and Figure 6.8 show the predicted distribution of enzyme activity over the cap
of undamaged (U) and damaged (MD 20) mushroom samples, respectively. Each figure
shows (a) false RGB images, (b) prediction maps based on the raw reflectance model
and (c) prediction maps based on the sample MSC pre-treated reflectance model of four
mushroom caps whose skin was processed together to obtain one single enzyme
extract. The mean and standard deviation (SD) of the predictions, both in [ln(EAU/g)],
are displayed below each map in (b) and (c). The values below the false RGB images
correspond to the activity measurement obtained experimentally for each extract, which
is the same for all of the mushrooms within each figure.
The main difference between the prediction images of U and MD 20 was the grayscale
intensity. The dark gray tonality in Figure 6.7.b and Figure 6.7.c indicated that the
models predicted low activity values for U mushroom caps. MD 20 predictions, on the
contrary, showed much lighter colours in Figure 6.8.b and Figure 6.8.c, which revealed
higher predicted values for enzyme activity. At the time of scanning, damaged
mushrooms looked different from undamaged ones and the corresponding extracts
exhibited much higher enzyme activity, for which it was expected that the models would
generate very different prediction images according to damage level.
132
(a)
(b)
(c)
Measured= 9.33
Mean=9.08 ; SD=0.85 Mean=9.08 ; SD=0.26
Measured = 9.33
Mean=9.15 ; SD=0.81
(a)
(b)
(c)
Measured = 10.98
Mean=11.31 ; SD=0.23 Mean=11.30 ; SD=0.13
Measured = 10.98
Mean=10.76 ; SD=0.32 Mean=10.72 ; SD=0.13
Measured = 10.98
Mean=10.82 ; SD=0.35
Mean=10.80 ; SD=0.13
Measured = 10.98 Mean=11.13 ; SD=0.30
Mean=11.11 ; SD=0.14
Mean=9.14 ; SD=0.085
133
Measured = 9.33
Measured = 9.33
Mean=9.26 ; SD=0.74 Mean=9.24 ; SD=0.087
Mean=9.41 ; SD=0.73 Mean=9.42 ; SD=0.073
Figure 6.7 Undamaged mushroom caps.
(a) false RGB image; (b) prediction maps by raw reflectance
model and (c) prediction maps by sample MSC corrected
reflectance model.
Figure 6.8 Damaged mushroom caps.
(a) false RGB images, (b) prediction maps by raw reflectance
model and (c) prediction maps by sample MSC corrected
reflectance model.
Chapter 6. Prediction of PPO with HSI
For all of the mushrooms in Figure 6.7 and Figure 6.8, the mean predicted values by raw
reflectance and sample MSC corrected reflectance models (displayed under each image
in columns (b) and (c)) were very similar. This pointed to the result that both raw
reflectance and sample MSC corrected reflectance models performed very similarly in
terms of quantitative prediction. This was in agreement with the similarities observed
previously in the coefficient of determination and the root-mean-square error of both
models. However, the very different appearance of predictions maps in (b) and (c)
means these two models had some dissimilarities too.
•
In the raw reflectance predicted images (Figure 6.7.b and Figure 6.8.b), the
distribution of enzyme activity prediction was uneven throughout the cap. The
relatively high standard deviation values under each map revealed this
heterogeneity too. As clearly seen in Figure 6.7.b, the highest predicted values
concentrated around the mushroom edges, (i.e the region showing higher level
of bruising on false RGB images (Figure 6.7.a). This could be partly due to
increased presence of brown coloured pigments at edge regions, which may be
derived from PPO-mediated reaction products, but spectral differences related to
mushroom curvature might also have affected the performance of the model
differently in different regions of the cap. It was difficult to estimate the extent of
such phenomena at this point. The lack of shading effects in Figure 6.8.b where
predicted values did not show any clear morphological trend, suggested that the
effect of sample curvature on the reflectance model may not be observable when
the levels of damage and browning were high.
•
However, it is interesting to note that all the images within Figure 6.7.b and
Figure 6.8.b revealed the ability of this model to point out the regions that looked
“different” in false RGB images. The model captured the spectral variation arising
from surface bruises/marks (e.g. confined regions which showed browner colour
in false RGB images) and reflected it onto the prediction maps. For undamaged
134
Chapter 6. Prediction of PPO with HSI
mushrooms, Figure 6.7.b exhibited lighter grayscale tonality (indicating higher
predicted values) on the small regions that showed signs of brusing in Figure
6.7.a. Similarly, for damaged mushrooms, Figure 6.8.b presented darker colour
(indicating lower predicted value) on those regions where browning had yet not
developed in Figure 6.8.a.
•
Sample MSC corrected reflectance predicted images, on the other hand,
appeared smoother than raw reflectance predictions. All the pixels within one
sample MSC corrected reflectance prediction image had similar predicted values
and therefore the grayscale intensity was very uniform and the SD values are
low. The MSC correction estimates the relation of the scatter of each pixel with
respect to the target spectrum (in this case, the mean spectrum of all the pixels)
(Geladi et al., 1985). Thus, a similar level of scatter was obtained for all spectra
and the effect that the morphology of the sample (i.e. mushroom curvature) could
have on the model was diminished too.
Figure 6.9 shows the enzyme activity prediction of imaginary lines drawn through the
centre of each mushroom cap, shown in red in Figure 6.9.a. Figure 6.9.b shows how the
raw reflectance model predicted the pixel values on those lines; the pixels that form the
line are represented in the x axis, while the predicted enzyme activity values are shown
in the y axis. The line in Figure 6.9.c corresponds to the prediction of the sample MSC
corrected reflectance model.
135
Chapter 6. Prediction of PPO with HSI
(b)
11.5
11
10.5
10
9.5
9
8.5
8
7.5
ln(PPO Activity[EAU/g]) = 9.33
0
20
40
60
80
100
120
140
12
11.5
11
10.5
10
9.5
9
8.5
8
7.5
160
0
20
40
Pixel number
12
11.5
11
10.5
10
9.5
9
8.5
8
7.5
0
20
40
60
80
100
Pixel number
60
80
100
120
140
160
Pixel number
ln(Predicted PPO Activity[EAU/g])
ln(Predicted PPO Activity[EAU/g])
ln(PPO Activity[EAU/g]) = 10.98
(c)
12
ln(Predicted PPO Activity[EAU/g])
ln(Predicted PPO Activity[EAU/g])
(a)
120
140
160
12
11.5
11
10.5
10
9.5
9
8.5
8
7.5
0
20
40
60
80
100
120
140
160
Pixel number
Figure 6.9 Prediction of an imaginary line drawn through the centre of mushroom caps.
Top row: undamaged mushroom cap; bottom row: damaged mushroom cap.
(a) False RGB image (imaginary line drawn in red); (b) corresponding prediction by raw
reflectance model and (c) corresponding prediction by sample MSC corrected
reflectance model.
For an undamaged mushroom (see top row), the curved shape of the prediction line in
(b) indicated that pixels from the centre and edge regions of the cap were predicted
differently; the activity was low in the central region of the mushroom and increased
gradually towards the edges. This was in agreement with what was observed in Figure
6.7.b and could be because the enzyme activity distribution was not uniform along the
mushroom cap surface (which was not possible to prove in this study) or because this
model was not able to deal with spectral differences arising from mushroom cap surface
curvature. The line in (c), predicted by the sample MSC corrected reflectance model,
was much flatter than the one in (b), which indicated that predictions along the imaginary
line were more homogeneous and suggested enzyme activity was equally distributed
over the mushroom cap. Despite the fact that both models predicted similar mean
activity values (9.91 [ln(EAU/g)] and 9.94 [ln(EAU/g)], respectively), differences in pixel
136
Chapter 6. Prediction of PPO with HSI
distribution suggested that the ability of each model to overcome spectral variability due
to sample morphology was different. For damaged mushrooms (see bottom row), the
line predicted by the reflectance model (b) was uneven but, as opposed to what was
observed in the undamaged mushroom, it did not have a clear curved shape. In this
case, the variation of predicted enzyme activity values across the imaginary line could
be related to the level of damage/browning, whereas the relationship between predicted
values and pixel position/surface curvature was not as clear as for undamaged
mushrooms. The line in (c) was flatter than in (b), as observed for undamaged
mushrooms. Raw reflectance and sample MSC corrected reflectance models predicted
very similar mean enzyme activity values (10.36 [ln(EAU/g)] and 10.37 [ln(EAU/g)],
respectively) and their distribution across the pixel line was more similar than in the case
of undamaged mushrooms.
6.4 Conclusions
Current methods employed for tyrosinase assay are time consuming. It was previously
reported that rapid spectroscopic techniques may replace analytical measurements. The
ability of a HSI system to predict PPO activity on mushroom caps was assessed in this
study.
Twenty four models employing different modelling strategies, pre-treatments and
regression methods were developed and tested in terms of their ability to predict the
PPO activity on mushroom caps. Two models, both three component Principal
Component Regression (PCR) models built under strategy 1 (i.e. when the mean
spectrum of each mushroom was extracted and the same enzyme activity value was
assigned to all the mushrooms used to obtained one extract), one on non-pretreated
reflectance data and the other one in sample MSC corrected reflectance data, were
found to perform best at predicting enzyme activity. These models gave R2=0.75 and
137
Chapter 6. Prediction of PPO with HSI
RMSECV=0.38 [ln(EAU/g)] for the training set and R2=0.78 and RMSEP=0.30
[ln(EAU/g)] for the test set
PPO activity prediction maps were generated to gain an understanding of i) the
distribution of the enzyme activity over the mushroom cap and ii) the effect of sample
MSC pre-treatment on the predictive ability of the model. Non-pretreated reflectance
model prediction maps showed that enzyme distribution was not uniform over the
samples, suggesting that activity might be higher in the edge regions of the mushrooms.
The effect of spectral variability due to sample curvature might have conditioned such a
result. Sample MSC corrected reflectance model prediction maps did not exhibit any
clear morphological trends and distributed predicted activity values uniformly over the
mushroom surface. These results suggested that correcting original spectra using a
common target spectrum could compensate for variability due to pixel location and
generate models with enhanced transferability.
Results revealed some potential of Vis-NIR hyperspectral imaging as a tool to estimate
the activity of enzymes responsible for mushroom browning. The mushroom industry
could benefit from such a tool for rapid identification of mushrooms of short shelf-lives
and reduced marketability.
138
7. Vis-NIR HYPERSPECTRAL
IMAGING FOR THE DETECTION
AND DISCRIMINATION OF BROWN
BLOTCH ON MUSHROOM CAPS
Chapter 7. Detection of brown blotch with HSI
7.1 Introduction
Cultivated mushrooms are susceptible to a variety of pests and diseases. Among them
is Pseudomonas tolaasii (P. tolaasii), the causal agent of brown blotch disease (Paine,
1919) and the most important pathogenic bacterium of A. bisporus (Nair and Bradley,
1980). This disease has been detected and described worldwide and affects not only the
button mushroom market but the mushroom market in general (Soler-Rivas et al., 1999).
Brown blotch can cause a general loss of crop yield of 10% and a decrease in quality of
another 10% (Rainey et al., 1992).
The most typical symptoms of brown blotch are pitting and browning of mushroom
tissues, induced by a watersoluble toxin, tolaasin (Moquet et al., 1996). This
extracellular toxin is produced by the pathogenic form of P. tolaasii (Nutkins et al., 1991).
The colonisation of mushroom caps by P. tolaasii results in the appearance of
unappealing brown spots on the mushroom cap and stipe. Lesions are slightly concave
blemishes, sometimes small, round or spreading in many directions (Olivier et al., 1978).
When the damage is more intense, the spots are darker and sunken. Browning affects
only the external layers of the cap tissue and is restricted to 2-3 mm below the surface of
the cap. Lesions may coalesce to cover the entire mushroom surface.
The mushroom industry is in need of objective evaluation methodologies to ensure that
only high quality produce reaches the market (Heinemann et al., 1994). Studies in the
field of brown blotch detection include the work of Vízhányó and Felföldi (2000), who
tested the potential of a machine vision system to recognise and identify brown blotch
and ginger blotch diseases, both of which cause discolouration in mushroom caps. A
vectorial normalisation method was developed to decrease the effect of the natural loss
of whiteness of the mushroom surface and increase the differences in the image caused
by the disease. The method showed an ability to discriminate between discolouration
caused by microbial disease and other sources of discolouration, such as natural
140
Chapter 7. Detection of brown blotch with HSI
senescence. However, no attempt was made to discriminate brown blotch from bruises
induced by mechanical stress, which is also an important source of discolouration and
quality loss in the mushroom industry (Jolivet et al., 1998).
Hyperspectral imaging (HSI) has been applied at various levels in the assessment of
safety and quality of food, including constituent analysis, quality evaluation and detection
of contaminants and defects (Gowen et al., 2007). Additionally, a number of researchers
have reported the potential of HSI for identification of microorganisms of concern in food
(Dubois et al., 2005; Escoriza et al., 2006). Recent advances in the detection of skin
damage of other products include work by Ariana et al. (2006) with cucumbers and
Nicolaï et al. (2006a) and ElMasry et al. (2009) with apples. As regards damage of
microbial origin, Gómez-Sanchis et al. (2008a) proposed a hyperspectral imaging
system for the early detection of rot caused by Penicillium digitatum (fungi) in
mandarins. This method’s success in classifying rotten fruit was above 91% and it
represented an alternative to the operationally inefficient sorting system previously used
in the citrus industry. While evidence from the literature points to its feasibility, HSI has
not been used to detect damage of bacterial origin in horticultural products.
The objective of this study was to investigate the potential application of Vis-NIR HSI for
brown blotch identification on mushroom caps and for its discrimination from mechanical
damage injuries.
7.2 Materials and methods
7.2.1 Sample preparation
For mushroom growing and harvest, refer to section 4.2.1.
Mushrooms were harvested two times, making training (ntrain = 144) and test (ntest = 108)
data sets. For each set of mushrooms, samples were divided in 3 groups (undamaged
141
Chapter 7. Detection of brown blotch with HSI
(U), mechanically damaged (MD) and P. tolaasii inoculated mushroom (PT)) of equal
size (ntrain,i = 48 and ntest,i = 36, where i = U, MD, PT).
Each mushroom class was treated as follows:
•
U: No treatment.
•
MD: damaged as described in section 4.2.2.2. Damage time was 10 min (in this
chapter, MD=MD 10). Samples were stored in the environmental incubator at
25°C and 90% RH for 24 h prior to imaging.
•
PT: samples were obtained by inoculating 4 drops of 10 µL/each of a solution of
pathogenic P. tolaasii onto each clean cap at 4 × 106 cfu. Samples were stored in
the incubator for 48 h at 25°C and 90% RH prior to imaging, to encourage
appearance of brown blotch symptoms on the mushroom caps.
A total number of 252 mushrooms were used in this experiment.
7.2.2 Preparation of pathogenic P. tolaasii solution
Freeze-dried culture (DMS no. 19342) was purchased from Deutsche Sammlung von
Mikroorganismen und Zellkulturen GmbH (DSMZ), Germany, resuspended in nutrient
broth (NB, Scharlau, Dublin) and incubated at 25°C for 24 h. The pure culture was
transferred into nutrient agar plates (NA, Oxoid, Dublin) and incubated at optimal
conditions to obtain isolated colonies.
Mushroom tissue block rapid pitting and White line in agar pathogenicity tests were
carried out following the procedure of Wong and Preece (1979) to confirm culture
pathogenicity on the inoculated mushrooms.
142
Chapter 7. Detection of brown blotch with HSI
7.2.2.1 Mushroom tissue block rapid pitting test
The outer skin of a mushroom was peeled off and mushroom cap tissue blocks of
approx. 15 × 15 × 5 mm were cut. Bacterial isolates were grown on Pseudomonas agar
base (PAB, Oxoid, Dublin) at 25°C for 24 h and suspended in sterile distilled water ( 108
cfu mL-1 approx). Mushroom blocks were placed in duplicate on Petri dishes containing
sterile water-moistened filter paper. The bacterial solution was inoculated onto the cut
surface of one of the mushroom blocks and incubated at 25°C. Sterile water was
inoculated onto the surface the other mushroom blocks for negative control. Pitting of
the cut surface of mushroom tissue revealed pathogenicity of P. tolaasii on mushrooms.
7.2.2.2 White Line in Agar test
Pseudomonas reactans (P. reactans) was streaked out in a line, directly from agar slope
culture, across PAB in a Petri dish. This strain had been isolated and provided by
Kinsealy Teagasc Research Centre.
The P. tolaasii isolate to be tested was streaked immediately after on to the plates at
right angles to the reacting organism (i.e. P. reactans). Plates were incubated at 25°C
for 24 h for the production of a white line with the reacting organism.
White line production in the agar between P. reactans and P. tolaasii was interpreted as
positive interaction between colonies and a positive response for pathogenicity test.
A loopfull of pathogenicity confirmed working culture was transferred to a sterile 0.8%
saline solution (Sigma, Dublin). 4 droplets of 10 µL/each of a 108 cfu mL-1 solution were
inoculated onto the cap, resulting in inoculation concentration of 4 × 106 cfu.
143
Chapter 7. Detection of brown blotch with HSI
7.2.3 Hyperspectral imaging
Refer to section 6.2.3. In this case, the focal length of the objective lens was 25 mm.
HSI images of U mushrooms were acquired on day 0 of the experiment. MD mushrooms
were scanned after 24 h of storage. PT mushroom images were taken after 48 h of
storage to allow for full development of browning and blotching.
Reflectance data were saved in ENVI header format using the acquisition software.
7.2.4 Confirmation of P. tolaasii
After image acquisition of PT mushrooms on day 2 of storage, 0.5 g of the pileipellis
10% of each mushroom class were extracted with a sharp sterile knife. Skins were
suspended separately in 10 mL of a 0.8% saline solution. Samples were homogenised
in a stomacher (Seward BA 7020, Seward, UK) for 60 s at high intensity. Serial dilutions
of each suspension were prepared and transferred onto NA and PAB plates, which were
incubated for 48 h at 25°C to obtain isolated colonies. The same procedure had been
carried out after image acquisition of U mushrooms on day 0 and resulting colonies were
used as negative controls. Serial dilutions of the pathogenic P. tolaasii solution that had
been used to inoculate PT mushrooms were also prepared and transferred onto NA and
PAB plates; the resulting colonies were used as positive controls.
Mushroom tissue block rapid pitting and White line in agar tests were carried out to test
for pathogenicity of isolated colonies.
7.2.5 Image processing
For each mushroom hyperspectral image, 175 characteristic (i.e. U, MD or PT,
depending on mushroom class) regions of interest (ROIs) were selected by using an
interactive selection tool (“ROI tool”) available in the acquisition software. The ROIs
144
Chapter 7. Detection of brown blotch with HSI
were 3 × 3 pixels in size and were selected from the central region of the mushroom
cap, where possible. Selecting spectra from analogous surface areas in all the
mushrooms aimed at minimising the scaling differences caused by mushroom surface
curvature (Gowen et al., 2009). The mean reflectance spectrum (“R”) of each ROI was
obtained by averaging the pixel spectra of the region. Spectral data of each mushroom
set were used to build two-dimensional matrices, where each row represented the
spectrum of one ROI.
Prior to the development of multivariate models for damage class prediction, spectra
were pre-processed using the Standard Normal Variate (SNV) transformation to reduce
spectral variability due to non-chemical biases (Barnes et al., 1989).
Training set matrices (raw and SNV-corrected) contained 8400 spectra and test set
matrices (raw and SNV) contained 6300 spectra.
7.2.6 Partial least squares- discriminant analysis (PLS-DA)
Partial least-squares discriminant analysis was applied to the training set matrices (raw
and SNV-corrected, n=8400) using MATLAB 7.0 (MathWorks, 2007). The aim was to
build models that would enable maximum separation of sample spectra into different
classes depending on their physical condition. A two step model approach was taken for
each set of spectra (i.e. raw and SNV): one model (namely “U/Dam” model) was
developed to discriminate between undamaged (U) and damaged (Dam) spectra and
another model (namely “MD/PT” model) was built to discriminate between the two
classes of Dam, i.e. mechanical (MD) and microbiological (PT). Overall, four models
were built: U/Dam_raw, U/Dam_SNV, MD/PT_raw and MD/PT_SNV.
For this purpose, a dummy response variable, named Y, was constructed using
analytical contrasts and assigned to each spectrum. For U/Dam models, Y = 0 for U
spectra and Y = 1 for Dam spectra. For MD/PT models, Y = 0 for MD spectra and Y = 1
145
Chapter 7. Detection of brown blotch with HSI
for PT spectra. In both cases, a cut-off value of 0.5 was used to classify spectra, as
suggested by Esquerre et al. (2009); spectra with a predicted dummy variable <0.5 were
identified as belonging to class 0, while those with predicted Y-value ≥0.5 were classified
as belonging to class 1.
Spectra from the training data set were split into 10 sections and continuous blocks
cross-validation was performed. The decision on the number of latent variables (# LV) to
select for each model was made based on the root-mean square error of crossvalidation (RMSECV), which is the mean of the sum of squared differences between the
actual and the predicted value of the dummy variable.
The models were also applied to the test set matrices (raw and SNV-corrected, n=6300),
which represented independent sets of sample spectra. Performance of the
classification models was evaluated on the basis of their sensitivity (number of spectra
of a given type correctly classified as that type) and specificity (number of spectra not of
a given type correctly classified as not of that type) on training and test sets.
7.2.7 Prediction maps
An important feature of hyperspectral imaging is the ability to map the distribution of
components/attributes on samples. In this case, developed PLS-DA models could be
applied to entire hypercubes of mushrooms to form two dimensional prediction images
where the damage class of each pixel as predicted by the PLS-DA models would be
represented by its intensity (“I”). With this in mind, the following step-by-step procedure
was carried out on all mushroom hypercubes:
1. Masking. This step was performed to separate the mushroom pixels from the
background. The mask was created by thresholding the mushroom image at 940
nm, where a pixel threshold value of 0.1 was used to segment the mushroom
146
Chapter 7. Detection of brown blotch with HSI
from the background. All background regions were set to zero and only the nonzero elements of the image were used in further steps.
2. Erosion. As all the spectra collected to build the models corresponded to
interactively selected ROIs of the central part of the mushrooms, the image
outline (i.e. edge) of the mushrooms was eroded to spectra showing differences
due to sample curvature. This was done by eroding the masks using disk-shaped
structuring elements (SE, whose radii increased from 0 – i.e. no erosion - to 40
pixels, in 10 pixel gaps), prior to the application of PLS-DA models. The effect of
varying the radius of the SE on i) the area of the mask, ii) the pixel distribution of
the predicted maps and iii) the performance statistics of the two models at a pixel
level, based upon ANOVA results obtained using R(R Development Core Team,
2007), was studied.
3. Application of developed PLS-DA models. The U/Dam model was applied to
eroded hypercubes and following this classification, the MD/PT model was
applied only to the pixels previously classified as Dam. Three binary images
(BinU, BinMD and BinPT, one for each damage class, where 1 indicated class
membership and 0 indicated non-membership) were generated after classifying
each pixel as “U” (IU/Dam<0.5), “MD” (IU/Dam>0.5 and IMD/PT<0.5) or “PT” (IU/PT>0.5
and IMD/PT>0.5) tissue.
4. Closing. As enzymatic browning was expected to develop uniformly across MD
mushroom caps and bacterial lesions were expected to appear as brown spots of
visible size, BinMD images were closed (Rocha et al., 2010) in order to avoid
noise in the form of isolated PT pixels in such maps. The Closing morphological
operator performs dilation followed by erosion; this was done using a diamondshaped SE with a radius of 3 pixels. The effect of omitting/incorporating Step 4
on i) the pixel distribution of the prediction maps and ii) performance statistics,
based upon ANOVA results, was investigated.
147
Chapter 7. Detection of brown blotch with HSI
5. Concatenation: the three binary maps (i.e. BinU, BinMD and BinPT) were
concatenated to build false colour maps where U, MD and PT classified pixels
were represented in green, red and blue, respectively.
7.2.8 Mushroom classification
Based on the percentage of pixels of each damage class on the prediction map, a
decision tree (shown in Figure 7.1) was used to allocate each mushroom to one of the
three mushroom classes.
As the main objective of this work was to identify P. tolaasii inoculated mushrooms, the
PT pixel percentage of the prediction maps was selected as the discrimination criteria
and a cut-off value was established by exploring the pixel histograms of the BinPT
images of U, MD and PT mushrooms (Figure 7.2).
148
Prediction map
Count the number of U, MD and PT pixels in the prediction image
Is the percentage of PT pixels greater than 2% ?
Yes
Classify the hypercube as PT
No
149
Is the percentage of MD pixels greater than the percentage
of U?
No
Classify the hypercube as U
Figure 7.1 Decision tree for mushroom hypercube classification.
U = undamaged; MD = mechanically damaged and PT = P. tolaasii inoculated.
Yes
Classify the hypercube as MD
Chapter 7. Detection of brown blotch with HSI
(b)
80
80
70
Number of mushrooms
Number of mushrooms
(a)
90
70
60
50
40
30
20
50
40
30
20
10
10
0
60
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0
0
1
2
Percentage of PT pixels
3
4
5
6
7
Percentage of PT pixels
(c)
5
Number of mushrooms
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
0
10
20
30
40
50
60
Percentage of PT pixels
Figure 7.2 Histograms showing number of mushrooms as a function of the percentage
of pixels in the BINPT binary images.
(a) undamaged (U), (b) mechanically damaged (MD) and (c) P. tolaasii inoculated (PT)
mushrooms. All the mushrooms (i.e. training and test set mushrooms) of each class
were plotted together, making a total of 86 samples per mushroom class.
A PT pixel percentage of 2% appeared to be a reasonable cut-off point, as all of the U
mushroom predictions and almost all of the MD mushroom predictions exhibited lower
values and almost all of the PT mushroom predictions showed higher values. After
visual inspection of the two PT mushrooms that were below the cut-off value, it was
observed that these two mushrooms did not develop any brown blotch on their caps, for
which they could be left out for cut-off establishment purposes. As can be seen in the
figure, when the number of PT pixels in the prediction image was higher than 2%, the
mushroom was classified as PT. If the number of PT pixels was lower than 2% and the
150
Chapter 7. Detection of brown blotch with HSI
amount of MD pixels was higher than the amount of U pixels, the mushroom was
classified as MD. Finally, if the number of PT pixels was lower than 2% and the amount
of MD pixels was lower than the amount of U pixels, the mushroom was classified as U.
Sensitivity and specificity of the classification procedure were computed after the
application of the decision tree to all of the mushroom hypercubes.
7.3 Results and discussion
7.3.1 RGB images
Figure 7.3 shows representative colour images of the three mushroom classes under
investigation in this study.
Figure 7.3 Representative RGB images of the mushrooms under investigation.
(a) undamaged (U); (b) mechanically damaged (MD) and (c) P. tolaasii inoculated (PT).
Mushrooms labelled as U (Figure 7.3.a) were white in general appearance, although
some of them showed some signs of natural discolouration. By day one of storage, MD
samples (Figure 7.3.b) exhibited uniform browning over the entire mushroom surface. By
day two of storage, P. tolaasii had colonised the cap of most PT mushrooms (Figure
7.3.c), which exhibited slightly concave brown-coloured spots, the typical symptoms of
brown blotch disease.
151
Chapter 7. Detection of brown blotch with HSI
7.3.2 Confirmation of P. tolaasii
The prevalence of Pseudomonas in mushroom surfaces is high, but only pathogenic P.
tolaasii is capable of causing brown blotch. Two different types of colonies, whose
colours were i) creamy and ii) green, were found in PAB plates of PT mushrooms, while
only one type appeared in PAB plates of U mushrooms and the P. tolaasii inoculum; the
colour of these colonies was creamy and green, respectively. Isolates of the two types
found in PT mushroom plates were obtained by re-streaking representative colonies
onto fresh PAB plates. Figure 7.4 shows growth in PAB plates of the two types of
colonies of PT mushrooms (Figure 7.4.a, creamy colonies and Figure 7.4.b, green), the
creamy colonies of U mushrooms (Figure 7.4.c) and the green colonies of the P. tolaasii
inoculum (Figure 7.4.c).
(a)
(b)
(c)
(d)
Figure 7.4 Colonies in PAB plates.
(a) Creamy colonies in PAB plate of PT mushrooms, (b) Green colonies in PAB plate of
PT mushrooms, (c) Creamy colonies in PAB plate of U mushrooms and (d) Green
colonies in PAB plate of P. tolaasii inoculum.
152
Chapter 7. Detection of brown blotch with HSI
The creamy colonies of PT mushrooms (Figure 7.4.a) were found to be similar to those
found in U plates (Figure 7.4.c), and both gave a negative response to the Mushroom
tissue block rapid pitting and White line in agar tests. The green colonies of PT
mushrooms (Figure 7.4.b) were similar to those found in P. tolaasii inoculum plates
(Figure 7.4.d) and both had a positive response to the two pathogenicity tests. These
results confirm that the pitting observed in PT mushrooms was due to pathogenic P.
tolaasii.
7.3.4 Spectra
Mean spectra of the various spectra classes are shown in Figure 7.5:(a) non-pretreated
reflectance spectra and (b) SNV-corrected reflectance spectra.
(b)
(a)
1.5
0.9
1
Reflectance [ ]
0.8
0.7
0.6
0.5
0.4
0.3
U
MD
PT
0.2
0.1
0
450
500
550
600
650
700
750
800
850
SNV Reflectance [ ]
1
0.5
0
-0.5
-1
-1.5
U
MD
PT
-2
-2.5
-3
450
900
Wavelength [nm]
500
550
600
650
700
750
800
850
900
Wavelength [nm]
Figure 7.5 Mean spectra of the mushroom classes under investigation.
(a) Mean raw reflectance spectra and (b) Mean SNV-corrected reflectance spectra of
selected regions of undamaged (U), mechanically damaged (MD) and P. tolaasii
inoculated (PT) mushroom caps.
In Figure 7.5.a, signal intensity and shape differences between U and MD spectra are
remarkable. The mean MD spectrum exhibited lower reflectance values over the entire
spectral region, as expected after bruising had led to loss of whiteness of the caps. The
greatest differences in shape between MD and U spectra arose in the 600-800 nm
153
Chapter 7. Detection of brown blotch with HSI
region, where the mean U spectrum exhibited broader features than the mean MD
spectrum. Broad spectra in the visible-near infrared wavelength range are characteristic
of undamaged mushrooms, corresponding to their white appearance. The spectral
differences mentioned above could be related to the formation of brown pigments,
mainly melanins, which derive from enzyme-catalysed oxidation products called
quinones. The mean PT spectrum appeared to be more similar in shape to the mean
MD spectrum, although its slope was not as linear as MD’s was in the 600-800 nm
region.
Spectral differences mainly arise from differences in sample composition, but differences
in illumination conditions, sample height and curvature may also affect the spectral
response of an object. Spectra pre-processing methods such as Multiplicative Scatter
Correction (MSC) (Geladi et al., 1985) and SNV (Barnes et al., 1989) can be used to
compensate for spectral variability caused by these external factors (Azzouz et al.,
2003).
The mean spectra of SNV-corrected reflectance spectra of U, MD and PT spectra are
shown in Figure 7.5.b. Comparing U and MD spectra, MD exhibited higher SNVcorrected reflectance values in the 450-500 nm region and lower SNV-corrected
reflectance in the 500-750 nm region. The oxidation of polyphenolic compounds and
subsequent formation of brown-coloured pigments in the MD mushrooms might be partly
responsible for this dissimilarity. In the 800-950 nm region, MD spectra showed higher
values than U mushrooms. Overall, the shape of the mean spectrum of PT spectra was
somewhat intermediate between the mean of U and MD spectra in the wavelength range
of study. The visible end of the mean PT spectrum looked more similar to MD than to U,
whereas its shape in the >700 nm region was very similar to the shape of U mushrooms.
154
Chapter 7. Detection of brown blotch with HSI
7.3.5 PLS-DA analysis
Figure 7.6 shows RMSECV and performance statistics (i.e. sensitivity and specificity) of
the four PLS-DA models developed, as a function of the number of latent variables (from
1 to 10). For both U/Dam models (Figure 7.6.a and Figure 7.6.b), RMSECV exhibited a
“corner” (pointed with a red dash circle) at 2 LV. The performance statistics, which were
very poor at 1LV, increased at that point and remained at similar levels thereafter.
Consequently, 2 was considered to be the optimal # LV for models discriminating
between U and Dam spectra. Both MD/PT models seemed to perform best when 4 LV
were selected; RMSECV did not decrease significantly after that and performance
statistics remained high.
(b)
(a)
1
1
U/Dam_raw
0.8
Sensitivity_train
0.6
Sensitivity_train
0.6
Specificity_train
0.4
U/Dam_SNV
0.8
Specificity_train
0.4
Sensitivity_test
0.2
Sensitivity_test
0.2
Specificity_test
0
0
2
4
6
8
Specificity_test
0
RMSECV
10
RMSECV
0
2
4
# LV
6
8
# LV
(d)
(c)
1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
MD/PT_raw
0.8
Sensitivity_train
0.6
Specificity_train
0.4
Sensitivity_test
0.2
Specificity_test
0
RMSECV
0
2
4
10
6
8
10
MD/PT_SNV
Sensitivity_train
Specificity_test
Sensitivity_test
Specificity_test
RMSECV
0
# LV
2
4
6
8
10
# LV
Figure 7.6 RMSECV and performance statistics of PLS-DA models as a function of the
number of LV.
(a) U/Dam_raw model, (b) U/Dam_SNV model, (c) MD/PT_raw model and (d)
MD/PT_SNV model. RMSECV: root mean square error of cross-validation.
155
Chapter 7. Detection of brown blotch with HSI
Numeric values of performance statistics of the selected models are shown in Table 7.1.
Table 7.1 Performance statistics at spectra level of PLS-DA models built on reflectance
spectra.
Model
U/Dam_raw
U/Dam_SNV
MD/PT_raw
MD/PT_SNV
# LV
2
2
4
4
Trainig set
Sensitivity Specificity
0.997
1
0.973
0.999
0.988
0.983
0.963
0.998
Test set
Sensitivity Specificity
0.832
1
0.825
0.999
0.661
0.984
0.641
0.998
When the models built with raw spectra were applied to the training set of spectra,
almost perfect classification was achieved in the case of the U/Dam model (sensitivity =
0.997 and specificity = 1.000). The model performed slightly worse when built on SNVcorrected spectra (sensitivity = 0.973 and specificity = 0.999). In both cases, almost all
of the Dam spectra were classified as such and none or only a few U were misclassified
as Dam. When the MD/PT model was applied to the damaged spectra, the sensitivity of
the raw model (sensitivity = 0.988) was slightly higher than that of the SNV-corrected
model (sensitivity = 0.963), whereas the specificity of the MD/PT_raw model was lower
than the MD/PT_SNV model’s (0.983 and 0.998, respectively). These results showed
that almost all of the spectra of the mushrooms that had been inoculated with P. tolaasii
were classified correctly and only a few or none of the spectra of the MD samples were
misclassified as PT.
When the models were applied to the test set of spectra, the sensitivity of the
U/Dam_raw model was lower (sensitivity = 0.832) but still none of the U spectra were
misclassified as Dam (specificity = 1.000). Performance statistics were quite similar for
the U/Dam_SNV model (sensitivity = 0.825 and specificity = 0.999). When the MD/PT
model was applied to the damaged spectra of the test set, a smaller percentage of raw
PT spectra were classified correctly (sensitivity = 0.661) but almost none of the MD
spectra were misclassified as PT (specificity = 0.984). As observed for the previous
156
Chapter 7. Detection of brown blotch with HSI
model, the sensitivity of the MD/PT_SNV model was slightly lower (sensitivity = 0.641)
and the specificity was higher (specificity = 0.998).
7.3.6 Prediction maps
Figure 7.7 shows examples of prediction maps (with no erosion applied in Step 2) of (a)
U, (b) MD and (c) PT mushrooms as a result of the application of raw (top row) and
SNV-corrected (bottom row) PLS-DA models to the data hypercubes. Overall,
predictions by models built on raw reflectance spectra appeared to be more appropriate
than predictions by models built on SNV-corrected reflectance spectra: for each
mushroom class, the corresponding pixel class was the main pixel class and pixels were
distributed in an even manner.
(a)
(b)
(c)
Figure 7.7 Prediction images of mushrooms (after no erosion in Step 2).
(a) undamaged (U); (b) mechanically damaged (MD) and (c) P. tolaasii inoculated (PT).
Top row: PLS-DA models built on raw reflectance spectra.
Bottom row: PLS-DA models built on SNV-corrected reflectance spectra.
157
Chapter 7. Detection of brown blotch with HSI
In the example shown above, on the top row (i.e. predictions by models built on raw
reflectance spectra), neither the map of the U mushroom nor the central region of the
prediction of the MD mushroom showed misclassification, whereas most edge pixels of
the latter were misclassified as U. Considering that all the spectra selected for model
building belonged to central regions of the mushrooms, this misclassification could be
related to the inability of the models to account for spectral differences due to mushroom
surface curvature. Fewer pixels were misclassified in the prediction map of the PT
mushroom, where some pixels were classified as MD. On the bottom row of Figure 7.7
(i.e. prediction maps by models built on SNV-corrected reflectance spectra), the maps of
all
mushroom
types
showed
misclassification.
For
U
and
PT
mushrooms,
misclassification happened mainly but not only on the edges, where many U pixels were
classified as MD. For MD mushrooms, misclassified pixels were distributed evenly along
the mushroom surface. In this case, MD pixels were misclassified as PT.
Considering that the main focus of this work lies in the identification of PT mushrooms,
after visual inspection of the prediction maps (Figure 7.7), PLS-DA models built on nonpretreated spectra were considered more appropriate for this purpose. Models built on
SNV-corrected reflectance spectra were therefore discarded and further sections of this
paper will focus only on models built on the raw data.
Effect of varying the radius of the SE in Step 2
Erosion of a binary image is a basic operation to wear the boundaries of regions away.
This can be done to overcome the problem introduced by the edge effect, by which
variability in reflection of light is introduced by spherical surfaces (Blasco et al., 2003). In
this particular case, where PLS-DA models were built on spectra selected from central
regions of the mushrooms, it was expected that these models would perform better on
central areas of the mushrooms than on edge regions. For this reason, masks defining
the mushroom region were eroded using SEs before the models were applied (see Step
158
Chapter 7. Detection of brown blotch with HSI
2 in Prediction maps section). This led to a decrease in misclassified pixels, which
typically belong to edge regions and are shown in green in Figure 7.7.b, top row.
Figure 7.8.a shows the effect that increasing the size of the structuring element used in
this step (i.e. Step 2) had on the mask (top row) and on the prediction map (bottom row)
of a MD class mushroom.
(a)
no erosion
disk radius = 10 pixels
disk radius = 20 pixels disk radius = 30 pixels
(c)
100
1.000
90
0.980
0.960
0.940
0.920
80
Sensitivity [ ]
Relative mask area [%]
(b)
disk radius = 40 pixels
70
60
50
40
30
20
U/Dam_train
0.900
0.880
0.860
0.840
MD/PT_train
U/Dam_test
MD/PT_test
0.820
0.800
10
0
0
10
20
30
0
40
10
20
30
40
Radius of structuring element [pixels]
Radius of structuring element [pixels]
Figure 7.8 Effect of varying the radius of the SE in Step 2.
(a) Binary masks (top row) and prediction maps (bottom row) of a mechanically
damaged (MD) mushroom, (b) Average relative mask area ± SD and (c) Sensitivity of
PLS-DA models, as a function of the radius of the structuring element (SE) used for
erosion in Step 2.
As the radius increased (from left to right, from 0 -no erosion- to 40 pixels), the mask
became smaller. Consequently, the number of MD class pixels that were misclassified
(as U class) decreased progressively. Figure 7.8.b shows the decrease of the average
relative area of the mushroom region as a function of the radius of SE, where the
159
Chapter 7. Detection of brown blotch with HSI
relative area of each mask at a certain SE radius value is displayed as a percentage of
the area of the mask when the radius was zero (i.e. when no erosion was applied) and
the average and standard deviation values were obtained by considering all the
mushrooms in the training and test data sets. Figure 7.8.c shows the sensitivity of the
PLS-DA models built on raw reflectance spectra applied at a pixel level as a function of
the radius of the SE. The sensitivity of the U/Dam model reached its maximum
(sensitivity = 1) at a radius value of 20 pixels when applied to the training set and at a
radius value of 40 (sensitivity = 0.972) when applied to the test set. The sensitivity of the
MD/PT model was not affected by the radius and remained at its maximum (sensitivity =
1) for the training set, whereas, for the test set, it increased progressively until it reached
its maximum (sensitivity = 0.944) at a radius value of 40 pixels. Modifying the radius of
the SE did not affect (p>0.05) the specificity of the PLS-DA models (results not shown).
Effect of omitting/incorporating Step 4: closing
As the existence of isolated PT pixels in the prediction map had no physical meaning
(brown blotch lesion on mushroom caps are detectable by the human eye), a closing
step (see Step 4 of Prediction maps section) was incorporated to the prediction map
routine. This step performed dilation followed by erosion on the BinPT images. Figure
7.9.a shows how the final prediction of a MD class mushroom looked when i) Step 4 was
omitted and ii) Step 4 was incorporated in the routine.
As can be observed in the figures, Step 4 removed PT class isolated pixels (blue) in the
prediction image and converted them into pixels of class MD (red). The effect of such
conversion on the performance statistics of the models at a pixel level was studied and
only the specificity of the MD/PT_raw model on the test set was found to change
significantly (p<0.05). Figure 7.9.b shows how specificity changed as the radius of the
SE of Step 2 increased, when i) Step 4 was omitted (round marker) and ii) Step 4 was
incorporated (square marker) in the routine. The specificity of the MD/PT_raw model
160
Chapter 7. Detection of brown blotch with HSI
improved when this step was incorporated, which means more MD class mushrooms
were correctly classified. This figure suggests that adding a closing step was important
to achieve good levels of classification.
(b)
(a)
(i)
(ii)
1.000
0.900
Specificity [ ]
0.800
0.700
0.600
0.500
0.400
Omitting Step 4
0.300
Incorporating Step 4
0.200
0.100
0.000
0
10
20
30
40
Radius of structuring element [pixels]
Figure 7.9 Effect of the omitting/incorporating Step 4.
(a) the pixel distribution in the prediction map of a mechanically damaged (MD)
mushroom when (i) Step 4 was omitted and (ii) Step 4 was incorporated.
(b) the specificity of the MD/PT_raw model on the test set as a function of the radius of
the structuring element (SE) used for erosion in Step 2.
After studying the two effects, a disk radius of 40 pixels was selected for the SE of Step
2 and it was decided to include Step 4 in the generation of prediction maps. Further
sections of this paper will only focus on results based on the use of the aforementioned
steps.
7.3.7 Mushroom classification
The application of PLS-DA models built on raw spectra to the totality of the entire
hypercubes led to the performance statistics shown in Table 7.2.
161
Chapter 7. Detection of brown blotch with HSI
Table 7.2 Performance statistics at pixel level of PLS-DA models built on raw
reflectance spectra.
Model
U/Dam_raw
MD/PT_raw
Training set
Test set
Sensitivity Specificity Sensitivity Specificity
1
1
0.972
1
1
0.979
0.944
0.972
For the training set mushrooms, both the sensitivity and the specificity of the U/Dam_raw
model were 1, which means there was no misclassification at all. For the same samples,
the sensitivity of the MD/PT_raw model was 1 and its specificity was 0.979. Only 1 out of
48 MD mushroom was misclassified as a PT mushroom. The models performed quite
similarly for the mushroom hypercubes of the test set: for the U/Dam_raw model,
sensitivity = 0.972 and specificity = 1. Only 2 out of 72 Dam mushrooms were
misclassified as U, and none of the U was misclassified as Dam. For the MD/PT_raw
model, sensitivity = 0.944 (only 2 out of 36 PT mushrooms were not classified as such)
and specificity = 0.972 (only 1 MD mushroom was misclassified as being PT).
These results show the models performed well when applied at a pixel level and confirm
they could be used to classify independent sets of mushrooms with high levels of
accuracy. Overall, the correct classification of the models presented in this study is
higher than the classification of the algorithms by Vízhányó and Felföldi (2000), which
correctly classified 81% of the diseased areas of test mushrooms using conventional
computer imaging. While those algorithms discriminated diseased spots from healthy
senescent mushroom parts, the models developed in this paper discriminate microbial
spoilage from both undamaged and mechanically damaged samples. The correct
discrimination between PT and MD mushrooms ensure no misclassification of samples
whose colour analysis might be similar and hence avoid “false positives”.
162
Chapter 7. Detection of brown blotch with HSI
7.4 Conclusions
It was previously reported that brown blotch is an important bacterial disease in the
mushroom industry. A conventional computer imaging system was found useful for the
discrimination between diseased and sound mushrooms. In order to study the potential
of HSI technology, the ability of a HSI system to discriminate brown blotch diseased
mushrooms from undamaged and mechanically damaged mushrooms was assessed.
Results presented in this work show that raw reflectance data of mushroom caps could
be used to classify mushrooms according to their damage class (i.e. undamaged,
mechanically damaged or brown blotch diseased). PLS-DA models were developed to
initially sort mushrooms into undamaged or damaged classes and to further classify the
damaged into mechanically damaged or microbiologically diseased classes. The
application of the models at a pixel level together with the use of a decision tree allowed
for correct classification of >95%. This study demonstrates the potential use of
hyperspectral imaging as an automated tool for detection of brown blotched mushrooms
and for their discrimination from mechanically damaged mushrooms.
Knowledge gained in this research could be incorporated towards the development of a
sensor to detect and classify mushroom damage of different sources. Such a system
could aid the industry in increasing quality control standards by correctly identifying low
quality produce. However, further research and evaluation at industrial scale are
required to facilitate its adoption.
163
8. HYPERSPECTRAL IMAGING
FOR THE DISCRIMINATION OF
MUSHROOMS INFECTED WITH
MUSHROOM VIRUS X
Chapter 8. Discrimination of MVX with HSI
8.1 Introduction
Mushroom Virus X (MVX) appeared for the first time in Britain in 1996 (Grogan et al.,
2004). The main symptoms associated with it are: pinning disruption, crop delay,
premature opening of the cap, brown coloured mushrooms, malformations and loss of
yield (Adie et al., 2004). In Ireland, the main symptom observed is the appearance of
brown and off-coloured mushrooms.
One of the problems this infection presents is that, while mushrooms may be harvested
in good condition, the quality may deteriorate rapidly during postharvest storage, with
the mushrooms going brown or developing an off-colour. Mushroom browning during
storage and distribution may present a risk to the mushroom grower in terms of brand
image, consumer perception of Irish mushrooms and rejected produce.
As part of a collaborative work with the project “Mushroom Virus X disease:
Understanding the factors which trigger brown mushroom symptom expression as a
means to improved diagnosis and control” (Stimulus funded project RSF 07-547),
coordinated by researchers in Kinsealy Teagasc Research Centre, the Agri-Food and
Biosciences Institute (AFBI, Loughall) and Dublin Institute of Technology, the potential of
hyperspectral imaging for the discrimination of mushrooms infected with MVX was
investigated.
Previous results from this project indicated that MVX infection occurs very rapidly in the
farm, but that the typical symptoms (creamy/brown mushrooms) occur only under very
specific conditions. The testing methods for MVX infection involve the detection of
double-stranded RNA (Grogan et al., 2003) as part of a costly and labour intensive
determination procedure. Thus, it would be very useful to develop rapid methods for
early detection of MVX infected mushrooms not showing the typical symptoms. Such
methods would facilitate fast and early monitoring of the infection, allowing for the
control of the disease before production was affected.
165
Chapter 8. Discrimination of MVX with HSI
The objective of this work was to study the possibility of using visible (Vis) and near
infrared (NIR) hyperspectral imaging (HSI) for the discrimination of mushrooms infected
with MVX.
8.2 Materials and methods
Experiments involved scanning mushrooms that had been infected with various strains
of MVX at different concentrations and stages of mushroom cultivation.
8.2.1 Sample preparation
For mushroom growing and harvest, refer to section 4.2.1.
Following arrival in the laboratory, mushrooms were stored at 4°C for 3 days.
8.2.2 Experimental design
8.2.2.1 Study of the effect of MVX strain
This experiment aimed at investigating whether different MVX strains cause different
symptoms in mushroom. The various strains used for mushroom infection are listed in
Table 8.1
Table 8.1 Experimental design of the first study.
Mushroom class
Control (n = 30)
A (n = 30)
B (n = 30)
C (n = 30)
Specifications of the infection procedure
Strain
Concentration
Time
A 15
0.01 %
End of spawn-run
MVX 2735
0.01 %
End of spawn-run
MVX 4569
0.01 %
End of spawn-run
MVX 4614
0.01 %
End of spawn-run
166
Chapter 8. Discrimination of MVX with HSI
8.2.2.2 Study of the effect of infection time/MVX concentration
This experiment aimed at investigating whether the timing and concentration of the MVX
infection caused different symptoms within the mushrooms. Table 8.2 shows the
different treatments used to infect the mushrooms.
Table 8.2 Experimental design of the second study.
Mushroom class
Control (n = 30)
D (n = 30)
E (n = 30)
B (n = 30)
Specifications of the infection procedure
Strain
Concentration
Timing
A 15
0.01 %
End of spawn-run
MVX 4569
25 %
Spawning
MVX 4569
0.01 %
Spawning
MVX 4569
0.01 %
End of spawn-run
Both studies were carried out using two consecutive flushes (2 replicates).
8.2.3 Image acquisition
8.2.3.1 Vis-NIR imaging
Refer to section 6.2.3.
8.2.3.2 NIR imaging
A pushbroom line-scanning instrument (refer to section 6.2.3.1) operating in the
wavelength range of 880-1720 nm (spectroscopic resolution of 7 nm) was employed.
HSI images were acquired using both instruments on days 0, 1, 2 and 3 of storage.
8.2.4 Data analysis
Mean reflectance and sample MSC spectra were extracted from each mushroom cap.
SNV and Savitzky-Golay 1st and 2nd derivative corrections were applied to raw spectral
data.
167
Chapter 8. Discrimination of MVX with HSI
Table 8.3 shows the chemometric tools which were applied to Vis and NIR data
separately.
Table 8.3 Chemometric tools employed and main purpose of utilisation.
Analysis
PCA
PCA
PLS-DA
PLSR
Study of the Vis/NIR correlation
Main purpose
To discriminate between mushroom classes
To discriminate between storage days
To predict mushroom class
To predict mushroom age
To discriminate between mushroom classes
Each analysis was applied to all the spectral pre-treatments described above.
8.3 Results and discussion
At the time of harvest, most infected mushrooms were asymptomatic or only showed
slight signs of browning to the naked eye. No evidence of browning developed during
storage time.
Researchers in Kinsealy confirmed that L* measurements of MVX-infected mushrooms
harvested for this experiment showed less variation in colour than mushrooms harvested
for previous experiments, i.e. there were less symptomatic mushrooms than in previous
experiments. It was previously reported that the expression of brown symptoms occurs
within a narrow window of opportunity and that it is difficult to produce consistently
symptomatic mushrooms. The conditions that need to be recreated for MVX-infected
mushrooms to show symptoms are unknown, which makes the study of this
phenomenon difficult.
Table 8.4 summarises the results of the classification of MVX asymptomatic mushrooms
using HSI and the different classification methods.
168
Chapter 8. Discrimination of MVX with HSI
Table 8.4 Summary of results obtained.
Analysis
PCA
PCA
PLS-DA
PLSR
Vis/NIR correlation
Main purpose
To discriminate between classes
To discriminate between days
To prediction mushroom class
To predict mushroom age
To discriminate between classes
Results
No separation between clusters
No separation between clusters
Very poor predictions
Poor predictions
Similar correlations
Figure 8.1.a shows PCA scores plot for Vis MSC-corrected reflectance data of control
mushrooms (green) and MVX infected mushrooms (red), on day 3 of storage. As can be
seen in the figure, the two clusters did not separate.
Figure 8.1.b shows PCA scores for Vis MSC-corrected reflectance data of mushrooms
infected by different MVX strains, on day 3 of storage. As can be seen in the figure, the
three clusters did not separate.
Neither Vis-NIR nor NIR HSI could differentiate between control and asymptomatic
MVX-infected mushrooms or between different mushroom strains.
(a)
(b)
0.08
0.08
Control
MVX
0.04
0.04
0.02
0.02
0
0
-0.02
-0.02
-0.04
-0.04
-0.06
-0.06
-0.08
-0.08
-0.1
-1
-0.8
-0.6
Class A
Class B
Class C
0.06
PC 2
PC 2
0.06
-0.4
-0.2
0
0.2
0.4
0.6
0.8
-0.1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
PC 1
PC 1
Figure 8.1 PCA scores plot for Vis MSC-corrected spectra representing control
mushrooms and mushrooms infected with different strains.
(a) Control and MVX-infected mushrooms. Day 3 of storage.
(b) Mushrooms infected with different MVX strains. Day 3 of storage.
Figure 8.2.a shows PCA scores plot for Vis raw reflectance data of control mushrooms
(green) and MVX infected mushrooms (red), on day 3 of storage. As can be seen in the
figure, the two clusters did not separate.
169
Chapter 8. Discrimination of MVX with HSI
Figure 8.2.b shows PCA scores for Vis raw reflectance data of mushrooms infected at
different timing/concentrations, on day 3 of storage. As can be seen in the figure, the
three clusters did not separate.
Neither Vis-NIR nor NIR HSI could differentiate between control and asymptomatic
MVX-infected mushrooms or between different infection timing/concentrations.
(a)
(b)
0.1
0.1
Control
MVX
0.05
0
0
PC 2
0.05
D
E
B
-0.05
-0.05
-0.1
-0.1
-0.15
-1
-0.5
0
0.5
-0.15
-1
-0.5
0
0.5
PC 1
Figure 8.2 PCA scores plot for Vis raw reflectance spectra representing control
mushrooms and mushrooms infected at different timing/concentrations.
(a) Control and MVX-infected mushrooms. Day 3 of storage.
(b) Mushrooms infected at different timing/concentrations. Day 3 of storage.
Figure 8.3.a shows PCA scores plot for NIR MSC-corrected reflectance data of control
mushrooms on different days of storage. As can be seen in the figure, the four clusters
did not separate.
Figure 8.3.b shows PCA scores for NIR MSC-corrected reflectance data of mushrooms
infected with MVX 4569 strain at 0.01% at the end of spawn-run. As can be seen in the
figure, the four clusters did not separate.
Neither Vis-NIR nor NIR HSI could differentiate between control and asymptomatic
MVX-infected mushrooms or between different infection timing/concentrations.
170
Chapter 8. Discrimination of MVX with HSI
(b)
0.15
0.1
0.1
0.05
0.05
0
0
PC 2
PC 2
(a)
0.15
-0.05
-0.1
-0.15
-0.2
-0.8
-0.05
-0.1
Day
Day
Day
Day
-0.6
0
1
2
3
Day
Day
Day
Day
-0.15
-0.4
-0.2
0
0.2
0.4
0.6
-0.2
-1.2
-1
0
1
2
3
-0.8
PC 1
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
PC 1
Figure 8.3 PCA scores plot for NIR MSC-corrected reflectance spectra representing
different storage days.
(a) Control mushrooms.
(b) Mushrooms infected with 4569 strain at 0.01% at the end of spawn-run.
In summary, none of the spectral pre-treatments, modelling strategies and wavelength
ranges proved useful in the discrimination of mushroom infected with MVX.
8.4 Conclusions
The potential of HSI technology with regard to discriminating MVX infected mushrooms
when symptoms are present should not be disregarded, given that there are some areas
of the spectrum that seem to point to small differences. Future studies with symptomatic
mushrooms are recommended.
Alternative spectroscopy techniques in other areas of the spectrum with richer metabolic
activity, and the use of imaging techniques to study different tissue parts and locations in
the mushrooms might offer successful alternatives to RNA analysis. In this way
recommendations to use FTIR or Raman spectroscopy or microscopy might prove
successful in the study of the propagation of this disease.
171
9. OVERALL CONCLUSIONS
AND RECOMMENDATIONS
Chapter 9. Conclusions and recommendations
This study employed visible and hyperspectral imaging systems for the detection and
discrimination of mechanical, microbiological and viral damage.
Various imaging experiments were conducted involving undamaged mushrooms,
mushrooms subjected to various levels of impact damage, mushrooms inoculated with a
solution of pathogenic Pseudomonas tolaasii and mushrooms infected with Mushroom
Virus X.
In summary, this study draws the following conclusions:
i) The employment of a visible imaging system allowed for the monitoring of the
browning due to mechanical stress and the identification of samples subjected to
damage.
ii) It was found that an average grayscale value could be useful to discriminate
between undamaged and mechanically damaged mushrooms. Incorporating the
information contained in images acquired prior to damage induction allowed for
faster differentiation between mushroom types.
iii) Attention should be paid to external variables such as changes in the illumination
conditions, location of the devices and sensitivity of the sensors, all of which can
affect the performance of inexpensive visible cameras.
iv) Vibrational damage induced in mushrooms caused an increase in the activity of
the enzymes responsible for mushroom browning.
v) At the levels of study, the extent of damage did not affect significantly the
increase in enzyme activity, which suggested there could be a threshold above
which enzyme expression would not be further affected by the level of damage.
vi) The increase in mushroom activity did not happen immediately after damage but
during storage time, reaching its maximum on day 1, 2 or 3 of storage depending
on the level of damage and the storage conditions.
173
Chapter 9. Conclusions and recommendations
vii) Increasing the storage temperature increased the levels of browning related
enzyme activity.
viii) Vis-NIR hyperspectral imaging reflectance data combined with chemometric
tools resulted in models with an ability to predict enzyme activity in mushrooms.
ix) The predictive ability of these models could be enhanced by incorporating
corrections that compensate for spectral variation due to sample morphology.
x) Hyperspectral imaging reflectance data could be used to classify mushrooms
according to the type of damage (i.e. microbiological or mechanical) with high
levels of accuracy.
xi) Hyperspectral imaging was unable to identify mushrooms infected with
Mushroom Virus X that were not showing typical browning symptoms. Further
studies with symptomatic mushrooms to clarify whether this technique has
potential for the detection of MVX viral infection in mushrooms are
recommended.
Overall, results from this study showed Vis-NIR HSI has potential to supplement and/or
replace classical analytical/microbiological assays, with a view to reducing time and
labour costs in the mushroom industry.
The industry could also benefit from results presented in this study by employing
hyperspectral imaging as a tool for the identification of mushrooms of reduced
marketability. Additional research and trials at industrial scale are necessary to facilitate
adoption of such techniques by industry. With a view to extending the work presented in
this thesis, the following recommendations are made:
i) Pilot studies using conventional imaging techniques in real retail and packaging
line situations would help to confirm the suitability of this technique for the
mushroom industry. Such investigation could also aid further implementation of
other imaging systems (i.e. hyperspectral imaging systems) in the future.
174
Chapter 9. Conclusions and recommendations
ii) A systematic study of the tyrosinase activity in the different mushroom strains
used in the industry, with the scope of identifying the steps along the
production/retail chain where increases in enzyme activity would most likely
occur, would help to avoid mushroom browning.
iii) Hyperspectral imaging experiments at laboratory scale using i) mechanically
damaged produce classified by current grading standards and ii) mushrooms
belonging to confirmed microbiological outbreaks would validate the application
of hyperspectral imaging for early detection of problems associated with
mushroom production. In this way, a collaborative assay involving the
conventional P. tolaasii analysis in Kinsealy and hyperspectral imaging of
mushrooms from outbreaks in Ireland would help to establish the usefulness of
hyperspectral imaging in the early detection of this microbial disease.
iv) The possibility of developing a combined computer vision-hyperspectral imaging
systems should be considered. Multispectral systems, which operate at <10
wavelengths, are faster and more affordable than hyperspectral systems. Data
obtained in this study could be employed to perform variable selection studies
(e.g. iterative PLS, general algorithms) in order to reduce spectral dimensionality
and determine the wavebands to include in the development of the new systems.
v) A reproducible technique to produce MVX infected mushrooms which exhibit
a) symptomatic and b) asymptomatic characteristics would aid in the preparation
of further hyperspectral imaging experiments.
vi) A study of the metabolome of MVX infected mushrooms and the possible
markers susceptible to being detected using spectroscopic and imaging
techniques would aid in the early identification and infection control of this
disease.
175
REFERENCES
References
Abdullah, M. Z., Guan, L. C., Lim, K. C. and Karim, A. A. (2004). The applications of
computer vision system and tomagraphic radar imaging for assessing physical
properties of food. Journal of Food Engineering, 61(1), 125-135.
Abdullah, M. Z. (2008). Image acquisition systems. In Sun, D.-W. (Eds), Computer vision
technology for food quality and evaluation, pp. 3-35. Elsevier.
Abdullah, Z. M., Aziz, A. S. and Dos-Mohamed, A. M. (2000). Quality inspection of
bakery products using a colour-based machine vision system. Journal of Food
Quality, 23(1), 39-50.
Adie, B., Choi, I., Soares, A., Holcroft, S., Eastwood, D., Mead, A., Grogan, H., Kerrigan,
R., Challen, M. and Mills, P. (2004). MVX disease and dsRNA ellements in
Agaricus bisporus. In Proceedings of the 16th International Congress on the
Science and Cultivation of Edible and Medicinal Fungi, Miami, USA.
Aguirre, L. (2008). Combination of natural variability estimation with real time
measurement for mushrooms shelf life assessment. Dublin Institute of
Technology. Dublin, Ireland. PhD thesis.
Aguirre, L., Frias, J., Barry-Ryan, C. and Grogan, H. (2008a). Some like it cool.
TResearch, 3(4), 27-29.
Aguirre, L., Frias, J. M., Barry-Ryan, C. and Grogan, H. (2008b). Modelling browning
and brown spotting of mushrooms (Agaricus bisporus) stored in controlled
environmental conditions using image analysis. Journal of Food Engineering,
91(2), 280-286.
Aguirre, L., Frias, J. M., Barry-Ryan, C. and Grogan, H. (2008c). Assessing the effect of
product variability on the management of the quality of mushrooms. Postharvest
Biology and Technology, 49(2), 247-254.
Aleixos, N., Blasco, J., Navarrón, F. and Moltó, E. (2002). Multispectral inspection of
citrus in real-time using machine vision and digital signal processors. Computers
and Electronics in Agriculture, 33(2), 121-137.
177
References
Ariana, D. P., Lu, R. and Guyer, D. E. (2006). Near-infrared hyperspectral reflectance
imaging for detection of bruises on pickling cucumbers. Computers and
Electronics in Agriculture, 53(1), 60-70.
Ariana, D. P. and Lu, R. (2008a). Quality evaluation of pickling cucumbers using
hyperspectral reflectance and transmittance imaging – part I: development of a
prototype. Sensing and Instrumentation for Food Quality and Safety, 2(3), 144151.
Ariana, D. P. and Lu, R. (2008b). Detection of internal defect in pickling cucumbers
using hyperspectral transmittance imaging. Transactions of the ASABE, 51(2),
705-713.
Ariana, D. P. and Lu, R. (2008c). Quality evaluation of pickling cucumbers using
hyperspectral reflectance and transmittance imaging – part II: performance of a
prototype. Sensing and Instrumentation for Food Quality and Safety, 2(3), 152160.
Azzouz, T., Puigdoménech, A., Aragay, M. and Tauler, R. (2003). Comparison between
different data pre-treatment methods in the analysis of forage samples using
near-infrared
diffuse
reflectance
spectroscopy
and
partial
least-squares
multivariate calibration method. Analytica Chimica Acta, 484(1), 121-134.
Ballard, D. D. and Brown, C. M. (1982). Computer vision. Englewood Cliffs, NJ, USA.
Prentice-Hall.
Bandyopadhyay, P., Jha, S. and Imran Ali, S. K. (2009). Picolyl alkyl amines as novel
tyrosinase inhibitors: influence of hydrophobicity and substitution. Journal of
Agricultural and Food Chemistry, 57(20), 9780-9786.
Barnes, R. J., Dhanoa, M. S. and Lister, S. J. (1989). Standard normal variate
transformation and detrending of near infrared diffuse reflectance spectroscopy.
Applied Spectroscopy, 43(5), 772-785.
178
References
Barron, C., Varoquaux, P., Guilbert, S., Gontard, N. and Gouble, B. (2002). Modified
atmosphere packaging of cultivated mushrooms (Agaricus bisporus L.) with
hydrophilic films. Journal of Food Science, 67(1), 251-255.
Basset, O., Buquet, B., Abouelkaram, S., Delachartre, P. and Culioli, J. (2000).
Application of texture image analysis for the classification of bovine meat. Food
Chemistry, 69(4), 437-445.
Beelman, R. B. (1987). Factors influencing postharvest quality and shelf-life of fresh
mushrooms. Mushroom News, 35(7), 12-18.
Bell, A. A. and Wheeler, M. H. (1986). Biosynthesis and functions of fungal melanins.
Annual Review of Phytopathology, 24, 411-451.
Berardinelli, A., Donati, V., Giunchi, A., Guarnieri, A. and Ragni, l. (2003). Effects of
transport vibrations on quality indices of shell eggs. Biosystems Engineering,
86(4), 495-502.
Berendse, H. (1984). Attitudes to mushrooms revealed in bureau survey. The Fruit
Trades Journal, November supplement.
Beyer, D. M. (2003). Basic procedures for Agaricus bisporus mushroom growing.
Pennsylvania State University.
Blasco, J., Aleixos, N. and Moltó, E. (2003). Machion vision system for automatic quality
grading of fruit. Biosystems Engineering, 85(4), 415-423.
Blasco, J., Aleixos, N., Gómez-Sanchís, J. and Moltó, E. (2009). Recognition and
classification of external skin damage in citrus fruits using multispectral data and
morphological features. Biosystems Engineering, 103(2), 137-145.
Bobelyn, E., Hertog, M. L. A. T. M. and Nicolaï, B. M. (2006). Applicability of an
enzymatic time temperature integrator as a quality indicator for mushrooms in the
distribution chain. Postharvest Biology and Technology, 42(1), 104-114.
Bord Bía. (2009). Factsheet on the Irish agriculture and food & drink sector (September
2009).
Retrieved
15
July
http://www.bordbia.ie/industryinfo/agri/pages/default.aspx.
179
2010,
from
References
Brennan, M., Le Port, G. and Gormley, R. (2000). Post-harvest treatment with citric acid
or hydrogen peroxide to extend the shelf life of fresh sliced mushrooms.
Lebensmittel-Wissenschaft und-Technologie, 33, 285-289.
Brennan, M. H. and Gormley, T. R. (1998). Extending the shelf life of fresh sliced
mushrooms. Dublin, Ireland. Teagasc.
Brodey, C. L., Rainey, P. B., Tester, M. and Johnstone, K. (1991). Bacterial blotch
disease of the cultivated mushroom is caused by an ion channel forming
lipodepsipeptide toxin. Molecular Plant-Microbe Interaction, 4, 407-411.
Brosnan, T. and Sun, D.-W. (2004). Improving quality inspection of food products by
computer vision - a review. Journal of Food Engineering, 61(1), 3-16.
Burger, J. and Geladi, P. (2006). Hyperspectral NIR image regression part II: Dataset
preprocessing diagnostics. Journal of Chemometrics, 20(3), 106-119.
Burger, J. and Geladi, P. (2007). Spectral pre-treatments of hyperspectral near infrared
images: analysis of diffuse reflectance scattering. Journal of Near Infrared
Spectroscopy, 15(1), 29-38.
Burton, K., Molloy, S. and Willoughby, N. (2002). Water is the key to bruising.
Horticultural Development Company News, 88, 29-30.
Burton, K. S. (1986). Quality-investigations into mushroom browning. Mushroom
Journal, 158, 68-70.
Burton, K. S., Frost, C. E. and Atkey, P. T. (1987a). Effect of vacuum cooling on
mushroom browning. International Journal of Food Science and Technology,
22(6), 599-606.
Burton, K. S., Frost, C. E. and Nichols, R. (1987b). A combination of plastic permeable
film systems for controlling post-harvest mushroom quality. Biotechnology
Letters, 9, 529-534.
Burton, K. S. (1988). The effects of pre and post-harvest development on mushroom
tyrosinase. Journal of Horticultural Science, 63, 255-260.
180
References
Burton, K. S. (1989). The quality and storage life of Agaricus bisporus. In 12th
International congress on the Science and Cultivation of edible fungi,
Braunschwey, Germany.
Burton, K. S., Love, M. E. and Smith, J. F. (1993). Biochemical changes associated with
mushroom quality in Agaricus spp. Enzyme and Microbial Technology, 15, 736741.
Burton, K. S. and Noble, R. (1993). The influence of flush number, bruising and storage
temperature on mushroom quality. Postharvest Biology and Technology, 3(1),
39-47.
Burton, K. S., Partus, M. D., Wood, D. A. and Thurston, C. F. (1997). Accumulation of
serine protease in senescent sporophores of the cultivated mushroom Agaricus
bisporus. Mycological Research, 101(2), 146-152.
Burton, K. S. and Beecher, T. M. (2000). The effects of calcium chloride on mushroom
whiteness. The Mushroom Journal, 604, 21-22.
Burton, K. S. (2004). Cultural factors affecting mushroom quality - Cause and control of
bruising. In Proceedings of the 16th International Congress on the Science and
Cultivation of Edible and Medicinal Fungi, Miami, USA.
Carey, A. T. and O'Connor, T. P. (1991). Influence of husbandry factors on the quality of
fresh mushrooms (Agaricus bisporus). Mushroom Science, 13, 673-682.
Chang, S. T. and Miles, P. G. (2004). Mushrooms: cultivation, nutritional value,
medicinal effect, and environmental impact. Boca Raton, USA. CRC Press.
Cheng, X., Chen, Y. R., Tao, Y., Wang, C. Y., Kim, M. S. and Lefcourt, A. M. (2004). A
novel integrated PCA and FLD method on hyperspectral image feature extraction
for cucumber chilling damage inspection. Transactions of the ASAE, 47(4), 13131320.
Chevallier, S., Bertrand, D., Kohler, A. and Courcoux, P. (2006). Application of PLS-DA
in multivariate image analysis. Journal of Chemometrics, 20(5), 221-229.
181
References
Chong, I.-G. and Jun, C.-H. (2005). Performance of some variable selection methods
when multicollinearity is present. Chemometrics and Intelligent Laboratory
Systems, 78(1-2), 103-112.
Courtecuisse, R. and Duhem, B. (1995). Mushrooms & Toadstools. London, UK. Harper
Collins.
Cutri, S. S., Macauley, B. J. and Roberts, W. P. (1984). Characteristics of pathogenic
non-fluorescent (smooth) and non-pathogenic fluorescent (rough) forms of
Pseudomonas
tolaasii
and
Pseudomonas
gingeri.
Journal
of
Applied
Bacteriology, 57, 291-298.
Davidson, V. J., Ryks, J. and Chu, T. (2001). Fuzzy models to predict consumer rating
for biscuits based on digital features. IEEE Transactions of Fuzzy Systems, 9(1),
62-67.
Davies, A. M. C. (2005). Back to basics: applications of principal component analysis.
Spectroscopy Europe, 17(2), 30-31.
Davies, A. M. C. (2007). Back to basics: spectral pre-treatments - derivatives.
Spectroscopy Europe, 19(2), 32-33.
Dhanoa, M. S., Lister, S. J., Sanderson, R. and Barnes, R. J. (1994). The link between
Multiplicative Scatter Correction (MSC) and Standard Normal Variate (SNV)
transformations of NIR spectra. Journal of Near Infrared Spectroscopy, 2, 43-47.
Du, C. J. and Sun, D.-W. (2004). Recent developments in the applications of image
processing techniques for food quality evaluation. Trends in Food Science and
Technology, 15, 230-249.
Dubois, J., Neil Lewis, E., Fry, F. S. J. and Calvey, E. M. (2005). Bacterial identification
by near-infrared chemical imaging of food-specific cards. Food Microbiology,
22(6), 577-583.
Eastwood, D. and Burton, K. (2002). Mushrooms - a matter of choice and spoiling
oneself. Microbiology today, 29, 18-19.
182
References
Eissa, H. A. A. (2007). Effect of chitosan coating on shelf life and quality of fresh-cut
mushroom. Journal of Food Quality, 30(5), 623-645.
Elibuyuk, I. O. and Bostan, H. (2010). Detection of a virus disease on Agaricus bisporus
(white button mushroom) in Ankara, Turkey. International Journal of Agriculture &
Biology, 12, 597-600.
ElMasry, G., Wang, N., ElSayed, A. and Ngadi, M. (2007). Hyperspectral imaging for the
nondestructive determination of some quality attributes for strawberry. Journal of
Food Engineering, 81(2), 98-107.
ElMasry, G., Wang, N., Vigneault, C., Qiao, J. and ElSayed, A. (2008). Early detection of
apple bruises on different background colors using hyperspectral imaging. LWT Food Science and Technology, 41(2), 337-345.
ElMasry, G. and Wold, J. P. (2008). High-speed assessment of fat and water content
distribution in fish fillets using online imaging spectroscopy. Journal of
Agricultural and Food Chemistry, 56(17), 7672-7677.
ElMasry, G., Wang, N. and Vigneault, C. (2009). Detecting chilling injury in Red
Delicious apple using hyperspectral imaging and neural networks. Postharvest
Biology and Technology, 52(1), 1-8.
Elmasry, G. and Sun, D.-W. (2010). Principles of hyperspectral imaging technology. In
Sun, D.-W. (Eds), Hyperspectral imaging for food quality analysis and control,
pp. 3-43. Academic Press.
Engelen, S. and Hubert, M. (2005). Fast model selection for robust calibration methods.
Analytica Chimica Acta, 544(1-2), 219-228.
Escoriza, M. F., VanBriesen, J. M., Stewart, S., Maier, J. and Treado, P. J. (2006).
Raman spectroscopy and chemical imaging for quantification of filtered
waterborne bacteria. Journal of Microbiological Methods, 66(1), 63-72.
Espin, J. C., Van Leeuwen, J. and Wichers, H. J. (1999). Kinetic study of the activation
process of a latent mushroom (Agaricus bisporus) tyrosinase by serine
proteinase. Journal of Agricultural and Food Chemistry, 47(9), 3509-3517.
183
References
Esquerre, C., Gowen, A. A., O'Donnell, C. P. and Downey, G. (2009). Initial studies on
the quantitation of bruise damage and freshness in mushrooms using visiblenear-infrared spectroscopy. Journal of Agricultural and Food Chemistry, 57(5),
1903-1907.
Falguera, V., Pagán, J. and Ibarz, A. (2010). A kinetic model describing melanin
formation by means of mushroom tyrosinase. Food Research International,
43(1), 66-69.
Fekedulegn, B. D., Colbert, J. J., Hicks, R. R. J. and Schuckers, M. E. (2002). Coping
with multicollinearity: an example on application of principal components
regression in dendroecology. U.S. Department of Agriculture, F. S., Northeastern
Research Station (Research Paper NE-721), USDA Forest Service Newton
Square PA: 43.
Fermor, T. R., Henry, M. B., Fenlon, J. S., Glenister, M. J., Lincoln, S. P. and Lynch, J.
M. (1991). Development and application of a biocontrol system for bacterial
blotch of the cultivated mushroom. Crop Protection, 10(4), 271-278.
Flegg, P. B., Spencer, D. M. and Wod, D. A. (1985). The biology and technology of the
cultivated mushroom. Chichester, UK. John Wiley and Sons.
Fletcher, J. and Gaze, R. (2008). Mushrooms Pest and Disease Control. London, UK.
Manson Publishing.
Flurkey, W. H. and Ingebrightsen, J. (1989). Polyphenoloxidase activity and enzymatic
browning in mushrooms. In Jen, J. J. (Eds), Quality Factors of Fruits and
Vegetables, pp. 45-54. Washington, USA. American Chemical Society.
Galeazzi, M. A., Sgarbieri, V. C. and Constantinides, S. M. (1981). Isolation, purification
and physicochemical characterization of polyphenoloxidases (PPO) from dwarf
variety of banana (Musa cavendishii). Journal of Food Science, 46(1), 150-155.
Geels, F. P., Van Griesven, L. J. I. D. and Rutjens, A. J. (1991). Chlorine dioxide and the
control of bacterial blotch on mushrooms, caused by Pseudomonas tolaasii.
Mushroom Science, 13, 437-442.
184
References
Geladi, P., MacDougall, D. and Martens, H. (1985). Linearization and scatter-correction
for near-infrared reflectance spectra of meat. Applied Spectroscopy, 39(3), 491500.
Geladi, P. and Kowalski, B. P. (1986a). An example of 2-block predictive partial leastsquares regression with simulated data. Analytica Chimica Acta, 185(1), 19-32.
Geladi, P. and Kowalski, B. P. (1986b). Partial least-squares regression: a tutorial.
Analytica Chimica Acta, 185(1), 1-17.
Goetz, A. F. H., Vane, G., Solomon, T. E. and Rock, B. N. (1985). Imaging spectrometry
for earth remote sensing. Science, 228, 1147-1153.
Gómez-Sanchis, J., Gómez-Chova, L., Aleixos, N., Camps-Valls, G., MontesinosHerrero, C., Moltó, E. and Blasco, J. (2008a). Hyperspectral system for early
detection of rottenness caused by Penicillium digitatum in mandarins. Journal of
Food Engineering, 89(1), 80-86.
Gómez-Sanchis, J., Moltó, E., Camps-Valls, G., Gómez-Chova, L., Aleixos, N. and
Blasco, J. (2008b). Automatic correction of the effects of the light source on
spherical objects. An application to the analysis of hyperspectral images of citrus
fruits. Journal of Food Engineering, 85(2), 191-200.
Gormley, R. (1975). Chill storage of mushrooms. Journal of the Science of Food and
Agriculture, 26, 401-411.
Gormley, T. R. (1967). Prepackaging and shelf life of mushrooms. Irish Journal of
Agricultural Research, 6(2), 255-265.
Gormley, T. R. and O'Sullivan, L. (1975). Use of a simple reflectometer to test
mushroom quality. The Mushroom Journal, 34, 344-346.
Gormley, T. R. (1987). Handling, Packaging and Transportation of Fresh Mushrooms. In
5th National Mushroom Conference, Malahide, Ireland.
Gormley, T. R. (1991). Monitoring mushroom temperature in the chill chain. In
Proceedings of the 13th International Congress on the Science and Cultivation of
Edible and Medicinal Fungi, Dublin, Ireland.
185
References
Gowen, A. A., O'Donnell, C. P., Cullen, P. J., Downey, G. and Frías, J. M. (2007).
Hyperspectral imaging - an emerging process analytical tool for food quality and
safety control. Trends in Food Science and Technology, 18, 590-598.
Gowen, A. A., O'Donnell, C. P., Cullen, P. J. and Bell, S. E. J. (2008a). Recent
applications of chemical imaging to pharmaceutical process monitoring and
quality control. European Journal of Pharmaceutics and Biopharmaceutics, 69,
10-12.
Gowen, A. A., O'Donnell, C. P., Taghizadeh, M., Cullen, P. J. and Downey, G. (2008b).
Hyperspectral imaging combined with principal component analysis for bruise
damage detection on white mushrooms (Agaricus bisporus). Journal of
Chemometrics, 22(3-4), 259-267.
Gowen, A. A., O'Donnell, C. P., Taghizadeh, M., Gaston, E., O'Gorman, A., Cullen, P. J.,
Frías, J. M., Esquerre, C. and Downey, G. (2008c). Hyperspectral imaging for the
investigation of quality deterioration in sliced mushrooms (Agaricus bisporus)
during storage. Sensing and Instrumentation for Food Quality and Safety, 2(3),
133-143.
Gowen, A. A. and O'Donnell, C. P. (2009). Development of algorithms for detection of
mechanical injury on white mushrooms (Agaricus bisporus) using hyperspectral
imaging. In Sensing for Agriculture and Food Quality and Safety. Proceedings of
SPIE, 7315OG, Orlando, USA.
Gowen, A. A., Taghizadeh, M. and O'Donnell, C. P. (2009). Identification of mushrooms
subjected to freeze damage using hyperspectral imaging. Journal of Food
Engineering, 93(1), 7-12.
Gowen, A. A., Taghizadeh, M. and O'Donnell, C. P. (2010). Using hyperspectral imaging
for quality evaluation of mushrooms. In Sun, D.-W. (Eds), Hyperspectral imaging
for food quality analysis and control, pp. 403-430. Academic Press.
Grogan, H. (2008a). Changes and challenges for the Irish mushroom industry.
Mushroom Business, 28, 22-23.
186
References
Grogan, H. (2008b). Communication on mushroom cultivation systems in Ireland.
Gaston, E. Dublin.
Grogan, H. (2008c). Challenges facing mushroom disease control in the 21st century. In
6th International Conference on Mushroom Biology and Mushroom Products,
Bonn, Germany.
Grogan, H. M., Adie, B. A. T., Gaze, R. H., Challen, M. P. and Mills, P. R. (2003).
Double stranded RNS elements associated with the MVX disease of Agaricus
bisporus. Mycological Research, 107(2), 147-154.
Grogan, H. M., Tomprefa, N., Mulcahy, J., Holcroft, S. and Gaze, R. H. (2004).
Transmission of Mushroom Virus X disease in crops. In Proceedings of the XVIth
International Congress on the Science and Cultivation of Edible and Medicinal
Fungi, Miami, USA.
Haaland, D. M. and Thomas, E. V. (1988a). Partial least-squares methods for spectral
analyses. 2. Application to simulated and glass spectral data. Analytical
Chemistry, 60, 1202-1208.
Haaland, D. M. and Thomas, E. V. (1988b). Partial least-squares methods for spectral
analyses. 1. Relation to other quantitative calibration methods and the extraction
of qualitative information. Analytical Chemistry, 60, 1193-1202.
Halaouli, S., Asther, M., Sigoillot, J. C., Hamdi, M. and Lomascolo, A. (2006). Fungal
tyrosinases: new prospects in molecular characteristics, bioengineering and
biotechnological applications. Journal of Applied Microbiology, 100(2), 219-232.
Hayashi, S., Kanuma, T., Ganno, K. and Sakaue, O. (1998). Cabbage head recognition
and size estimation for development of a selective harvester. In ASAE Annual
Meeting, Paper No. 983042, St. Joseph, Michigan, USA.
Hege, E., O'Connell, D., Johnson, W., Basty, S. and Dereniak, E. (2003). Hyperspectral
imaging for astronomy and space surveillance. Proceedings of SPIE, 5159, 380391.
187
References
Heinemann, P. H., Hughes, R., Morrow, C. T., Sommer, H. J., Beelman, R. B. and
Wuest, P. J. (1994). Grading of mushrooms using a machine vision system.
Transactions of the ASAE, 37(5), 1671-1677.
Hertog, M. L. A. T. M., Lammertyn, J., Desmet, M., Scheerlinck, N. and Nicolai, B. M.
(2004). The impact of biological variation on postharvest behaviour of tomato
fruit. Postharvest Biology and Technology, 34(3), 271-284.
Hertog, M. L. A. T. M., Lammertyn, J., De Katelaere, B., Scheerlink, N. and Nicolaï, B.
M. (2007). Managing quality variance in the postharvest food chain. Trends in
Food Science and Technology, 18, 320-332.
Hewett, E. W. (2006). An overview of preharvest factors influencing postharvest quality
of horticultural products. International Journal of Postharvest Technology and
Innovation, 1(1), 4-15.
Howarth, M. S. and Searcy, S. W. (1992). Inspection of fresh carrots by machine vision.
In Food Processing Automation Conference II, St. Joseph, Michigan, USA.
Hughes, D. H. (1958). Discoloration findings noted longer shel-life is the goal. Mushroom
News, 4, 9-13.
Hutchings, J. B. (1999). Food color and appearence. Gaithersburg, USA. Aspen
Publishers, Inc.
Hutchison, M. L. and Johnstone, K. (1993). Evidence for the involvement of the surface
active properties of the extracellular toxin tolaasin in the manifestation of brown
blotch disease symptoms by Pseudomonas tolaasii on Agaricus bisporus.
Physiological and Molecular Plant Pathology, 42, 373-384.
Isaksson, T. and Næs, T. (1988). The effect of multiplicative scatter correction (MSC)
and linearity improvement in NIR spectroscopy. Applied Spectroscopy, 42(7),
1273-1284.
Jolivet, S., Voiland, A., Pellon, G. and Arpin, N. (1995). Main factors involved in the
browning of Agaricus bisporus. Mushroom Science, 14, 695-702.
188
References
Jolivet, S., Arpin, N., Wichers, H. J. and Pellom, G. (1998). Agaricus bisporus browning:
a review. Mycological Research, 102(12), 1459-1483.
Kader, A. A. (1985). Postharvest biology and technology, an overview. In Kader, A. A.
(Eds), Postharvest technology of horticultural crops, pp. 39-48. California, USA.
Division of Agriculture and Natural Resources, University of California.
Kim, K. M., Ko, J. A., Lee, L. S., Park, H. J. and Hanna, M. A. (2006). Effect of modified
atmosphere packaging on the shelf-life of coated, whole and sliced mushrooms.
LWT - Food Science and Technology, 39(4), 365-372.
Kim, M. S., Chen, Y. R. and Mehl, P. M. (2001). Hyperspectral reflectance and
fluorescence imaging system for food quality and safety. Transactions of the
ASAE, 44(3), 721-729.
Kim, M. S., Lefcourt, A. M., Chao, K., Chen, Y. R., Kim, I. and Chan, D. E. (2002).
Multispectral detection of fecal contamination on apples based on hyperspectral
imagery: Part I. Application of visible and near-infrared reflectance imaging.
Transactions of the ASAE, 45(6), 2027-2037.
Knee, M. and Miller, A. R. (2002). Mechanical injury. In Knee, M. (Eds), Fruit Quality and
its Biological Basis, pp. 157-179. Sheffield, UK. Academic Press.
Krutz, G. W., Gibson, H. G., Cassens, D. L. and Zhang, M. (2000). Colour vision in
forest and wood engineering. Landwards, 55, 2-9.
Lee, J. S. (1999). Effects of modified atmosphere packaging on the quality of chitosan
and CaCl2 coated mushroom (Agaricus bisporus). Korean Journal of Food
Science and Technology, 31(5), 1308-1314.
Leemans, V., Magein, H. and Detain, M. F. (1998). Defects segmentation on 'Golden
Delicious' apples bu using colour machine vision. Computers and Electronics in
Agriculture, 20(2), 117-130.
Leemans, V. and Kleynen, O. (2008). Quality evaluation of apples. In Sun, D.-W. (Eds),
Computer Vision Technology for Food Quality Evaluation, pp. Elsevier.
189
References
Lewis, E. N., Schoppelrei, J. W., Lee, E. and Kidder, L. H. (2005). Near-infrared
chemical imaging as a process analytical tool. In Bakeev, K. A. (Eds), Process
Analytical Technology, pp. 187-225. Oxford, UK. Blackwell Publishing.
Liu, L., Ngadi, M. O., Prasher, S. O. and Gariépy, C. (2010). Categorization of pork
quality using Gabor filter-based hyperspectral imaging technology. Journal of
Food Engineering, 99(3), 284-193.
Liu, Y., Chen, Y. R., Wang, C. Y., Chan, D. E. and Kim, M. S. (2005). Development of
simple algorithm for the detection of chilling injury in cucumbers from
visible/near-infrared hyperspectral imaging. Applied Spectroscopy, 59(1), 78-85.
Liu, Y., Chen, Y.-R., Kim, M. S., Chan, D. E. and Lefcourt, A. M. (2007). Development of
simple algorithms for the detection of fecal contaminants on apples from
visible/near infrared hyperspectral reflectance imaging. Journal of Food
Engineering, 81(2), 412-418.
Lukasse, L. J. S. and Polderdijk, J. J. (2003). Predictive modelling of post-harvest quality
evolution in perishables, applied to mushrooms. Journal of Food Engineering,
59(2-3), 191-198.
MacCanna, C. (1984). Commercial Mushroom Production. Dublin, Ireland. An Foras
Talúntais-The Agricultural Institute.
Mahajan, P. V., Rodrigues, F. A. S., Motel, A. and Leonhard, A. (2008). Development of
a moisture absorber for packaging of fresh mushrooms (Agaricus bisporous).
Postharvest Biology and Technology, 48(3), 408-414.
Manickavasagan, A., Sathya, G. and Jayas, D. S. (2008). Comparison of illuminations to
identify wheat classes using monochrome images. Computers and Electronics in
Agriculture, 63(2), 237-244.
Marique, T., Pennincx, S. and Kharoubi, A. (2005). Image segmentation and bruise
identification on potatoes using a Kohonen's self-orgnizing map. Journal of Food
Science, 70, 415-417.
190
References
Martens, H. and Naes, T. (1996). Multivariate Calibration. Chichester, UK. John Wiley &
Sons.
Martens, H. A. and Dardenne, P. (1998). Validation and verification of regression in
small data sets. Chemometrics and Intelligent Laboratory Systems, 44(1-2), 99121.
Mason, H. S. (1955). Comparative biochemistry of the phenolase complex. Advances in
Enzymology, 16, 105-184.
Massart, D. L., Vandeginste, B. G. M., Lewi, L. M. C. and Smeyers-Verbeke, J. (1997).
Handbook of Chemometrics and Qualimetrics: Part A. Amsterdam, The
Netherlands. Elsevier.
Massart, D. L., Vandeginste, B. G. M., Deming, S. M., Michotte, Y. and Kaufmann, L.
(2003). Chemometrics: a textbook. Amsterdam, The Netherlands. Elsevier.
MathWorks. (1994-2010). Statistics Toolbox 7.3: Partial Least Squares Regression and
Principal
Components
Regression.
Retrieved
11th
June
2010,
from
http://www.mathworks.com/products/statistics/demos.html?file=/products/demos/
shipping/stats/plspcrdemo.html.
MathWorks (2007). MATLAB, version R2007b. MATLAB Inc.USA.
Mayer, A. M. and Harel, E. (1979). Polyphenol oxidases in plants. Phytochemistry, 18,
193-215.
Mayer, A. M. (1987). Polyphenol oxidases in plants - recent progress. Phytochemistry,
26, 11-20.
Mayer, A. M. (2006). Polyphenol oxidases in plants and fungi: Going places? A review.
Phytochemistry, 67(21), 2318-2331.
Mehl, P. M., Chen, Y.-R., Kim, M. S. and Chan, D. E. (2004). Development of
hyperspectral imaging technique for the detection of apple surface defects and
contamination. Journal of Food Engineering, 61(1), 67-81.
Mendoza, F. and Aguilera, J. M. (2004). Application of image analysis for classification
of ripening bananas. Journal of Food Science, 69(9), 471-477.
191
References
Mendoza, F., Dejmek, P. and Aguilera, J. M. (2006). Calibrated color measurements of
agricultural foods using image analysis. Postharvest Biology and Technology,
41(3), 285-295.
Menesatti, P., Beni, C., Plagia, G., Marcelli, S. and D'Andrea, S. (1999). Predictive
statistical model for the analysis of drop impact damage on peach. Journal of
Agricultural Engineering Research, 73(3), 275-282.
Menesatti, P. and Paglia, G. (2001). Development of a drop damage index of fruit
resistance to damage. Journal of Agricultural Engineering Research, 80(1), 5364.
Menesatti, P., Paglia, G., Solaini, S., Zanella, A., Stainer, R., Costa, C. and Cecchetti,
M. (2002). Non-linear multiple regression models to estimate the drop damage
index of fruit. Biosystems Engineering, 83(3), 319-326.
Menesatti, P., Zanella, A., D'Andrea, S., Costa, C., Plagia, G. and Pallottino, F. (2008).
Supervised multivariate analysis of hyper-spectral NIR images to evaluate the
starch index of apples. Food and Bioprocess Technology, 2, 308-314.
Menesatti, P., Costa, C. and Aguzzi, J. (2010). Quality evaluation of fish by
hyperspectral imaging. In Sun, D.-W. (Eds), Hyperspectral imaging for food
quality analysis and control, pp. 273-294. Academic Press.
Mohapatra, D., Frías, J. M., Oliveira, F. A. R., Bira, Z. M. and Kerry, J. (2008).
Development and validation of a model to predict enzymatic activity during
storage of cultivated mushrooms (Agaricus bisporus spp). Journal of Food
Engineering, 86(1), 39-48.
Mohapatra, D., Bira, Z. M., Kerry, J. P., Frías, J. M. and Rodrigues, F. A. (2010).
Postharvest hardness and color evolution of white button mushrooms (Agaricus
bisporus). Journal of Food Science, 75(3), 146-152.
Moltó, E. and Blasco, J. (2008). Quality evaluation of citrus fruits. In Sun, D.-W. (Eds),
Computer Vision Technology for Food Quality Evaluation, pp. Elsevier.
192
References
Monteiro, S., Minekawa, Y., Kosugi, Y., Akazawa, T. and Oda, K. (2007). Prediction of
sweetness and amino acid content in soybean crops from hyperspectral imagery.
ISPRS Journal of Photogrammetry and Remote Sensing, 62(1), 2-12.
Moquet, F., Mamoun, M. and Olivier, J. M. (1996). Pseudomonas tolaasii and tolaasin:
comparison of symptom induction on a wide range of Agaricus bisporus strains.
FEMS Microbiology Letters, 142, 99-103.
Naganathana, G. K., Grimesb, L. M., Subbiaha, J., Calkinsb, C. R., Samalc, A. and
Meyera, G. E. (2008). Visible/near-infrared hyperspectral imaging for beef
tenderness prediction. Computers and Electronics in Agriculture, 64(2), 225-233.
Nair, N. G. (1974). Methods of control for bacterial blotch disease of the cultivated
mushroom with special reference to biological control. The Mushroom Journal,
16, 140-144.
Nair, N. G. and Bradley, J. K. (1980). Mushroom blotch bacterium during cultivation. The
Mushroom Journal, 90, 201-203.
Nicolaï, B. M., Scheerlinck, N., Verboven, P. and Maerdemaeker, J. D. (2001).
Stochastic finite element analysis of thermal food precesses. In Irudayaraj, J.
(Eds), Food precessing operations modelling: design and analysis, pp. 265-304.
New York, USA. Marcel Dekker.
Nicolaï, B. M., Lötze, E., Peirs, A., Scheerlinck, N. and Theron, K. I. (2006a). Nondestructive measurement of bitter pit in apple fruit using NIR hyperspectral
imaging. Postharvest Biology and Technology, 40(1), 1-6.
Nicolaï, B. M., Scheerlinck, N. and Hertog, M. L. A. T. M. (2006b). Probabilistic
modeling. In Sablani, S. S., Datta, A. K., Rahman, M. S. and Mujumdar, A. S.
(Eds), Handbook of food bioprocess modeling techniques, pp. 265-290. Boca
Raton, USA. CRC Press.
Novini, A. (1995). The latest in vision technology in today's food and beverage container
manufacturing industry. In Food Processing Automation Conference IV, St.
Joseph, Michigan, USA.
193
References
Nutkins, J. C., Mortishire-Smith, R. J., Packman, L. C., Brodey, C. L., Rainey, P. B.,
Johnstone, K. and Williams, D. H. (1991). Structure determination of Tolaasin, an
extracellular lipodepsipeptide produced by mushroom pathogen Pesudomonas
tolaasii Paine. Journal of American Chemical Society, 113, 2661-2627.
O'Gorman, A., Downey, G., Gowen, A. A., Barry-Ryan, C. and Frias, J. M. (2010). Use
of Fourier Transfor Infrared Spectroscopy and Chemometric Data Analysis to
Evaluate Damage and Age in Mushrooms (Agaricus bisporus) Grown in Ireland.
Journal of Agricultural and Food Chemistry, 58(13), 7770-7776.
Olivier, J. M., Guillaumes, J. and Martin, D. (1978). Study of a bacterial disease of
mushroom caps. In 4th International Conference on Plant Pathogenic Bacteria,
Angers, France.
Osborne, B. G. and Fearn, T. (1986). Near infrared spectroscopy in food analysis.
Hallow, UK. Longman Scientific and Technical.
Osborne, S. D., Jordan, R. B. and Künnemeyera, R. (1997). Method of wavelength
selection for partial least squares. Analyst, 122, 1531-1537.
Paine, S. G. (1919). Studies in bactoeriosis II: a brown blotch disease of cultivated
mushrooms. Annals of Applied Biology, 5, 206-219.
Park, B., Lawrence, K. C., Windham, W. R. and Smith, D. P. (2006). Performance of
hyperspectral imaging system for poultry surface fecal contaminant detection.
Journal of Food Engineering, 75(3), 340-348.
Park, B., Windham, W. R., Lawrence, K. C. and Smith, D. (2007). Contaminant
classification of poultry hyperspectral imagery using a spectral angle mapper
algorithm. Biosystems Engineering, 96(3), 323-333.
Partington, J. C., Smith, C. and Bolwell, G. P. (1999). Changes in location of polyphenol
oxidase in potato (Solanum tuberosum L.) tuber during cell death in response to
impact injury: comparison with wound tissue. Planta, 207, 449-460.
Pinheiro, J. C. and Bates, D. M. (2000). Mixed-effects Models in S and S-PLUS'. New
York, USA. Springer-Verlag.
194
References
Polder, G., Heijden, G. and Young, I. (2002). Spectral image analysis for measuring
ripeness of tomatoes. Transactions of the ASAE, 45, 1155-1161.
Qiao, J., Ngadi, M., Wang, N., Gariépy, C. and Prasher, S. (2007). Pork quality and
marbling level assessment using a hyperspectral imaging system. Journal of
Food Engineering, 83(1), 10-16.
Qin, J., Burks, T. F., Ritenour, M. A. and Bonn, W. G. (2009). Detection of citrus canker
using hyperspectral reflectance umaging with spectral information divergence.
Journal of Food Engineering, 93(2), 183-191.
Qin, J. (2010). Hyperspectral imaging instruments. In Sun, D.-W. (Eds), Hyperspectral
imaging for food quality analysis and control, pp. 129-172. Academic Press.
R Development Core Team (2007). R: a language and environment for statistical
computing, version 2.9.1. Vienna, Austria.
Rainey, P. B., Brodey, C. L. and Johnstone, K. (1992). Biology of Pseudomonas tolaasii,
cause of brown blotch disease of the cultivated mushrooms. In Andrews, J. H.
and Tommerup, I. (Eds), Advances in Plant Pathology, pp. 95-117. Academic
Press.
Reed, J. N., Crook, S. and He, W. (1995). Harvesting mushrooms by robot. In Elliot, T.
J. (Eds), Science and cultivation of edible fungi, pp. 385-391. Rotterdam, The
Netherlands. Balkema.
Rocha, A., Hauagge, D. C., Wainer, J. and Goldenstein, S. (2010). Automatic fruit and
vegetable classification from images. Computers and Electronics in Agriculture,
70(1), 96-104.
Roggo, Y., Chalus, P., Meurer, L., Lema-Partinez, C., Edmond, A. and Jent, N. (2007). A
review of near infrared spectroscopy and chemometrics in pharmaceutical
technologies. Journal of Pharmaceutical and Biomedical Analysis, 44, 683-700.
Roy, S., Ramaswamy, C. A., Shenk, J. S., Westerhaus, O. M. and Beelman, R. B.
(1993).
Determination
of
moisture
content
of
mushrooms
by
Vis-NIR
spectroscopy. Journal of the Science of Food and Agriculture, 63, 355-360.
195
References
Roy, S., Anantheswaran, R. C. and Beelman, R. B. (1995). Fresh mushroom quality
affected by modified atmosphere packaging. Journal of Food Science, 60(2),
334-340.
Royse, D. J. and Wuest, P. J. (1980). Mushroom brown blotch: effects of chlorinated
water on disease intensity and bacterial populations in casing soil and on pilei.
Phytopathology, 70, 902-905.
Russ, J. C. (2005). Analysis of food microestructure. New York, USA. CRC Press.
Sangwine, S. J. (2000). Colour in image processing. Electronics & Communication
Engineering Journal, 12, 211-219.
Sapirstein, H. D. (1995). Quality control in commercial baking: machine vision inspection
of crumb grain in bread and cake products. In Food Processing Automation
Conference IV, St. Joseph, Michigan, USA.
Schisler, L. C., Sinden, J. W. and Sigel, E. M. (1967). Etiology, symptomology and
epidemiology of a virus disease of cultivated mushrooms. Phytopathology, 57,
519-526.
Schouten, R. E. and Van Kooten, O. (1998). Keeping quality of cucumber batches: Is it
predictable? Acta Horticulturae, 476, 349-355.
Schouten, R. E., Jongbloed, G., Tijskens, L. M. M. and Van Kooten, O. (2004). Batch
variability and cultivar keeping of cucumber. Postharvest Biology and
Technology, 32(3), 299-310.
Seida, S. B. and Frenke, E. A. (1995). Unique applications of machine vision for
container inspection and sorting. In IV Proceedings of the Food Processing
Automation Conference, St. Joseph, Michigan, USA.
Shrestha, B. P., Nagata, M. and Cao, Q. (2001). Study on image processing for quality
estimation of strawberries (part 1): detection of bruises on fruit by color image
processing. Journal of Science and High Technology in Agriculture, 13(2), 115122.
196
References
Sinden, J. W. and Hauser, E. (1950). Report of two new mushroom diseases. Mushroom
Science, 1, 96-100.
Siyami, S., Brown, G. K., Burgess, G. J., Gerrish, J. B., Tennes, B. R., Burton, C. L. and
Zapp, R. H. (1988). Apple impact bruise prediction models. Transactions of the
ASAE, 31(4), 1038-1046.
Skoog, D. A., Holler, F. J. and Nieman, T. A. (1998). Principles of instrumental analysis.
Saunders College Pub.
Soler-Rivas, C., Arpin, N., Olivier, J. M. and Wichers, H. J. (1997). Activation of
tyrosinase in Agaricus bisporus strains following infection by Pseudomonas
tolaasii or treatment with a tolaasin-containing preparation. Mycological
Research, 101(3), 375-382.
Soler-Rivas, C., Jolivet, S., Arpin, N., Olivier, J. M. and Wichers, H. J. (1999).
Biochemical and physiological aspects of brown blotch disease of Agaricus
bisporus. FEMS Microbiology Reviews, 23, 591-614.
Soler-Rivas, C., Moller, A. C., Arpin, N., Olivier, J. M. and Wichers, H. J. (2001).
Induction of a tyrosinase mRNA in Agaricus bisporus upon treatment with
tolaasin preparation from Pseudomonas tolaasii. Physiological and Molecular
Plant Pathology, 58(2), 95-99.
Solomon, J. M., Beelman, R. B. and Bartley, C. E. (1991). Addition of calcium chloride
and stabilized chlorine dioxide to irrigation water to improve quality and shelf life
of Agaricus bisporus. Mushroom Science, 13, 695-701.
Sonka, M., Hlavac, V. and Boyle, R. (1999). Image processing, analysis and machine
vision. California, USA. PWS Publishing.
SPSS (2006). SPSS for Windows, version 15.0.0. SPSS Inc.Chicago, USA.
Storbeck, F. and Daan, B. (2001). Fish species recognition using computer vision and a
neural network. Fisheries Research, 51, 11-15.
197
References
Strachan, N. J. C. and Kell, L. (1995). A potential method for differentiation between
haddock fish stocks by computer vision using canonical discriminant analysis.
ICES Journal of Marine Science, 52, 145-149.
Sun, D.-W. (2000). Inspecting pizza topping percentage and distribution by a computer
vision method. Journal of Food Engineering, 44(4), 245-249.
Sun, D.-W. and Zheng, L. (2006). Vacuum cooling technology for the agri-food industry:
Past, present and future. Journal of Food Engineering, 77(2), 203-214.
Sun, D.-W., Ed. (2008). Computer vision technology for food quality evaluation. Food
Science and Technology, International Series. Elsevier.
Taghizadeh, M., Gowen, A. and O'Donnell, C. P. (2009). Prediction of white button
mushroom (Agaricus bisporus) moisture content using hyperspectral imaging.
Sensing and Instrumentation for Food Quality and Safety, 4, 219-226.
Taghizadeh, M., Gowen, A. and O'Donnell, C. P. (2010). Use of hyperspectral imaging
for evaluation of the shelf-life of fresh white button mushrooms (Agaricus
bisporus) stored in different packaging films. Innovative Food Science and
Emerging Technologies, 11(3), 423-431.
Tan, B. K. and Harris, N. D. (1995). Maillard products inhibit apple polyphenoloxidase.
Food Chemistry, 53(3), 267-273.
Tan, F. J., Morgan, M. T., Ludas, L. I., Forrest, J. C. and Gerrard, D. C. (2000).
Assessment of fresh pork colour with colour machine vision. Journal of Animal
Science, 78(12), 3078-3085.
Tao, F., Zhang, M. and Yu, H. (2007). Effect of vacuum cooling on physiological
changes in the antioxidant system of mushroom under different storage
conditions. Journal of Food Engineering, 79(4), 1302-1309.
Teagasc (1994). Introduction to Mushroom Growing in Bags. Dublin, Ireland. Teagasc.
Teagasc (2006). The Irish Agriculture and Food Development Authority. Teagasc
Mushroom Newsletter, 27.
198
References
Teagasc (2007). The Irish Agriculture and Food Development Authority. Teagasc
Mushroom Newsletter, 29.
Thybo, A. K., Jespersen, S. N., Lærke, P. E. and Stødkilde-Jørgensen, H. J. (2004).
Nondestructive detection of internal bruise and spraing disease symptoms in
potatoes using magnetic resonance imaging. Magnetic Resonance Imaging,
22(9), 1311-1317.
Tijskens, L. M. M. and Konopacki, P. (2003). Biological variance in agricultural products.
Acta Horticulturae, 600, 661-670.
Tijskens, L. M. M., Konopacki, P. and Simcic, M. (2003). Biological variance, burden or
benefit?. Postharvest Biology and Technology, 27(1), 15-25.
Tijskens, L. M. M., Konopacki, P. J., Schouten, R. E., J.Hribar and M.Simcic (2008).
Biological variance in the colour of Granny Smith apples. Modelling the effect of
senescence and chilling injury. Postharvest Biology and Technology, 50(2-3),
153-163.
Timmermans, A. J. M. (1998). Computer vision system for online sorting of pot plants
based on learning techniques. Acta Horticulturae, 421, 91-98.
Turner, E. M. (1974). Phenoloxidase activity in relation to substrate and development
stage in the mushroom, Agaricus bisporus. Transactions of the British
Mycological Society, 63, 541-547.
Vamos-Vigyazo, L. (1981). Polyphenol oxidases and peroxidases in fruits and
vegetables. CRC Critical Review of Food Science and Nutrition, 15, 49-127.
van den Vooren, J. G., Polder, G. and van der Heijden, G. W. A. M. (1992). Identification
of mushrooms cultivars using image analyses. Transactions of the ASAE, 37(5),
1671-1677.
Van Horen, L. (2008). Economic developments in the mushroom industry - Grasping the
opportunities. In The 6th International Conference on Mushroom Biology and
Mushroom Products, Bonn, Germany.
199
References
Van Loon, P. C. C. (1996). Het bepalen van het ontwikkelingsstadium bij dechampignon
met computer beeldanalyse. Champignoncultuur, 40(9), 347-353.
Van Zeebroeck, M., Tijskens, E., Van Liedekerke, P., Deli, V., De Baerdemaeker, J. and
Ramon, H. (2003). Determination of the dynamical behaviour of biological
materials during impact using a pendulum device. Journal of Sound and
Vibration, 266, 465-480.
Van Zeebroeck, M., Van linden, V., Ramon, H., De Baerdemaeker, J., Nicolaï, B. M. and
Tijskens, E. (2007). Impact damage of apples during transport and handling.
Postharvest Biology and Technology, 45, 157-167.
Viscarra-Rossel, R. A., Taylor, H. J. and McBratney, A. B. (2007). Multivariate calibration
of hyperspectral gamma-ray energy spectra for proximal soil sensing. European
Journal of Soil Science, 58, 343-353.
Vízhányó, T. and Felföldi, J. (2000). Enhancing colour differences in images of diseased
mushrooms. Computers and Electronics in Agriculture, 26(2), 187-198.
Wells, J. M., Sapers, G. M., Fett, W. F., Butterfield, J. E., Jones, J. B., Bouzar, H. and
Miller, F. C. (1996). Postharvest discoloration of the cultivated mushroom
Agaricus bisporus caused by Pseudomonas tolaasii, P. "reactans" and P.
"gingeri". Phytopathology, 86(1098-1104)
Williams, P. C. (1987). Variables affecting near-infrared reflectance spectroscopic
analysis. In Williams, P. and Norris, K. (Eds), Near-Infrared Technology in the
Agricultural and Food Industries, pp. 143-166.
Wong, W. C. and Preece, T. F. (1979). Identification of Pseudomonas tolaasii : the
White Line in Agar and Mushroom Tissue Block Rapid Pitting Tests. Journal of
Applied Bacteriology, 47, 401-407.
Wong, W. C. and Preece, T. F. (1985). Pseudomonas tolaasii on mushroom (Agaricus
bisporus) crops: bacterial effects of six disinfectants and their toxicity to
mushrooms. Journal of Applied Bacteriology, 58, 269-273.
200
References
Xing, J., Bravo, C., Jancsók, P., Ramon, H. and De Baerdemaeker, J. (2005). Detecting
bruises on 'Golden Delicious' apples using hyperspectral imaging with multiple
wavebands. Biosystems Engineering, 90(1), 27-36.
Xing, J. and De Baerdemaeker, J. (2005). Bruise detection on 'Jonagold' apples using
hyperspectral imaging. Postharvest Biology and Technology, 37(2), 152-162.
Xing, J., Jancsók, P. and De Baerdemaeker, J. (2006). Stem-end/calyx identification on
apples using contour analysis in multispectral images. Biosystems Engineening,
96(2), 231-237.
Xing, J. and De Baerdemaeker, D. (2007). Fresh bruise detection by predicting softening
index of apple tissue using VIS/NIR spectroscopy. Postharvest Biology and
Technology, 45(2), 176-183.
Xing, J., Saeys, W. and De Baerdemaeker, D. (2007). Combination of chemometric tools
and image processing for bruise detection on apples. Computers and Electronics
in Agriculture, 56(1), 1-13.
Xing, J., Hung, P. V., Symons, S., Shahin, M. and Hatcher, D. (2009). Using a Short
Wavelength Infrared (SWIR) hyperspectral imaging system to predict alpha
amylase activity in individual Canadian western wheat kernels. Sensing and
Instrumentation for Food Quality and Safety, 3, 211-218.
Zarkower, P. A., Wuest, P. J., Royse, D. J. and Mayers, B. (1984). Phenotypic traits of
fluorescent pseudomonas causing bacterial blotch of Agaricus bisporus
mushrooms and other mushroom-derived fluorescent pseudomonas. Canadian
Journal of Microbiology, 30, 360-367.
Zeaiter, M., Roger, J. M. and Bellon-Maurel, V. (2005). Robustness of models developed
by multivariate calibration. Part II: The influence of pre-processing methods.
Trac-Trends in Analytical Chemistry, 24(5), 437-445.
Zheng, G., Chen, Y., Intes, X., Chance, B. and Glickson, J. D. (2004). Contrastenhanced near-infrared (NIR) optical imaging for subsurface cancer detection.
Journal of Porphyrins and Phthalocyanines, 8(9), 1106-1117.
201
References
Zhu, J., Wang, Z. and Xu, Y. (2006). Effects of postharvest storage temperature and its
fluctuations on the keeping quality of Agaricus bisporus. International Journal of
Food Engineering, 2(1), art. 3.
Zion, B., Shklyar, A. and Karplus, I. (1999). Sorting fish by computer vision. Computers
and Electronics in Agriculture, 23(3), 175-187.
Zwiggelaar, R., Yang, Q., Garcia-Pardo, E. and Bull, C. R. R. (1996). Use of spectral
information and machine vision for bruise detection on peaches and apricots.
Journal of Agricultural Engineering Research, 63(4), 323-331.
202
LIST OF PUBLICATIONS
Journal publications
•
Gaston, E., Frias, J.M., Cullen, P.J., O’Donnell, C.P. and Gowen, A.A (2010).
Visible
near-infrared
hyperspectral
imaging
for
the
identification
and
discrimination of brown blotch disease on mushroom (Agaricus bisporus) caps.
Accepted for publication in Journal of Near Infrared Spectroscopy on 20.09.2010.
•
Gaston, E., Frias, J.M., Cullen, P.J., O’Donnell, C.P. and Gowen, A.A (2010).
Prediction of polyphenol oxidase activity using visible and near-infrared
hyperspectral imaging on mushroom (Agaricus bisporus) caps. Journal of
Agricultural and Food Chemistry 58 (10), 6226-6233.
•
Gowen, A.A., O’Donnell, C.P., Taghizadeh, M., Gaston, E., O’Gorman, A.,
Cullen, P.J., Frias, J.M., Esquerre, C. and Downey, G. (2008). Hyperspectral
imaging for the investigation of quality deterioration in sliced mushrooms
(Agaricus bisporus) during storage. Sensing and Instrumentation for Food
Quality and Safety 2 (3), 133-143.
Conference abstracts
•
Gaston, E., Frias, J.M., Cullen, P.J., O’Donnell, C.P. and Gowen, A.A (2009).
Hyperspectral imaging for the detection and discrimination of brown blotch
disease on Agaricus bisporus mushroom caps. IFT 2009 Annual Meeting and
Food Expo. Anaheim-USA. 6th -9th June 2009.
•
Gowen, A., Gaston, E., O’Gorman, A., Esquerre, C., Taghizadeh, M., Cullen,
P.J., Barry-Ryan, C., O’Donnell, C.P., Downey, G. and Frias, J.M. (2009). Postharvest
mushroom
quality
assessment
by
hyperspectral
imaging
and
metabolomics. All Ireland Mushroom Conference and Trade Show. MonaghanIreland. 21st May 2009.
•
Gaston, E., Frias, J.M., Cullen, J.M., Sullivan, C., O’Gorman, A., Barry-Ryan,C.
and
Aguirre, L. (2008). Efficacy of a simple computer imaging system to
204
differentiate and monitor the browning of mushrooms. 6th International
Conference on Mushroom Biology and Mushroom Products. Bonn-Germany. 30th
September-3rd October 2008.
•
Aguirre, L., Gaston, E., Frias, J.M., Barry-Ryan, C. and Grogan, H. (2008). Image
analysis of mushrooms stored in controlled environmental conditions. 6th
International Conference on Mushroom Biology and Mushroom Products. BonnGermany. 30th September-3rd October 2008.
•
Aguirre, L., Gaston, E., Frias, J.M., Barry-Ryan, C. and Grogan, H. (2008).
Respiration rate (RR) & transpiration rate (TR) in mushroom (Agaricus bisporus)
under environmental storage conditions. 6th International Conference on
Mushroom Biology and Mushroom Products. Bonn-Germany. 30th September-3rd
October 2008.
•
O’Gorman, A., Frias, J.M., Barry-Ryan, C., Gaston, E. and Downey, G. (2008).
The use of infra-red spectroscopy coupled with chemometric tools to evaluate
between damaged and undamaged mushrooms. 6th International Conference on
Mushroom Biology and Mushroom Products. Bonn-Germany. 30th September-3rd
October 2008.
•
Gaston, E., Frias, J.M., Cullen, P.J., Gowen, A.A., Taghizadeh, M. and
O’Donnell, C.P. (2008). The effect of mechanical damage on the kinetics of PPO
activity and correlation with hyperspectral imaging. 38th Annual Research
Conference on Food, Nutrition & Consumer Sciences. Cork-Ireland. 4th
September 2008.
•
Gowen, A.A., O'Donnell, C., Taghizadeh, M., Gaston, E., O' Gorman, A., Cullen,
P.J., Frias, J.M., Esquerre, C. and Downey, G. (2008). Partial least squares
regression for prediction of mushroom (Agaricus bisporus) slice quality during
storage. CAC2008, 11th conference on Chemometrics in Analytical Chemistry.
Montpellier-France. 30th June-4th July 2008.
205