Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test

Page created by Timothy Oconnor
 
CONTINUE READING
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
Journal of Agricultural Science; Vol. 11, No. 13; 2019
                                                                               ISSN 1916-9752 E-ISSN 1916-9760
                                                                 Published by Canadian Center of Science and Education

      Classification and Pixel-Based Segmentation to Evaluate Soybean
                     Seeds Submitted to Tetrazolium Test
                 Davi Marcondes Rocha1, Lúcia Helena Pereira Nóbrega2, Maria de Fátima Zorato3,
                            Vitor Alex Alves de Marchi1 & Arlete Teresinha Beuren1
1
    Computer Science Department, Federal University of Technology-Paraná, Santa Helena, Paraná, Brazil
2
    Tec. and Exact Sciences Center, Western Paraná State University, Cascavel, Paraná, Brazil
3
    Biologist, Seed Science & Technology, Londrina, Paraná, Brazil
Correspondence: Davi Marcondes Rocha, Computer Science Department, Federal                              University    of
Technology-Paraná, Santa Helena, Paraná, Brazil. E-mail: davimarcondesrocha@gmail.com

Received: April 25, 2019          Accepted: June 3, 2019          Online Published: August 15, 2019
doi:10.5539/jas.v11n13p186              URL: https://doi.org/10.5539/jas.v11n13p186

Abstract
Production and use of high quality seeds are essential for the soybean crop. Thus, the quality control system in
seeds industry must be reliable, precise, and fast. Tetrazolium test evaluates not only seeds viability but also their
vigor, as well as provides information concerning agents that cause their quality reduction. Although this test
does not use expensive devices and reagents, it requires a well-trained analyst. Its precision depends on
knowledge of all techniques and required procedures. Besides, also necessary is the observer’s subjectivity. So,
this trial aimed at developing a computational tool that could minimize the implicit subjectivity in carrying out
this test. It also contributes to generate a greater credibility of information and to guarantee precise answers.
Algorithms of supervised classification were applied based on extraction of digital images characterization of
tetrazolium test. This procedure aimed at producing pixel-based segmentation of those images, to produce a
digital segmented image of tetrazolium test according to damage classes. This tool allows, based on image of
tetrazolium test, to identify damage on soybean embryos, as well as its site and extension on tissues, so that the
interpretation is less subjective. The applied method allowed identifying damage on images of tetrazolium tests
in a straightforward way, as well as extracting safer information about those damages and carrying out
management control of tetrazolium test according to a seed data file.
Keywords: machine learning, patterns recognition, seeds vigor
1. Introduction
Soybean crop has much contributed to the Brazilian commercial agriculture. It helped on crop mechanization
speed, as well as it modernized transportation, expanded agricultural boundaries, and has taken part in the
technological evolution and in another crops production. It also helped the development of Brazilian poultry and
swine farming (Dall’Gnol, 2000). Its high protein content (40%) is considered a raw material to feed domestic
animals, and although it does not have a high oil content (about 20%), soybean and palm are the greatest oil
producers (Roessing, Dall’Agnol, Lazzarotto, Hirakuri, & Oliveira, 2007).
According to Marcos-Filho (2015), seed technology has as main objective the development of efficient
mechanisms for the productive chain, using high-quality batches. Quality seeds are those that contain high
germination potential and vigor, with high physical and genetic purity, characteristics that can be analyzed by
multiple tests, following rigorous norms (Marcos-Filho, 2015; FAO, 2013).
The tetrazolium test has stood out among the quality tests adopted by the seed industry, especially for soybeans.
This occurs not only because of its accuracy and speed when compared to other tests, but also because of the
substantial number of information provided as diagnostic of potential causes for quality reduction (mechanical
damage, deterioration due to humidity and bed bug damage) and the possibility of evaluating seed batches’
quality and vigor (França Neto, Krzyzanowski, & Costa, 1998). This test enables to distinguish between
vigorous (bright red-carmine colored) and deteriorating (deep red) tissues, and dead or heavily deteriorated
tissues (which retain their whitish color) (Deswal & Chand, 1997), and it has been shown that each type of
damage is associated with typical patterns of injury (Figure 1).

                                                         186
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                               Journal of A
                                                         Agricultural Sciience                         V
                                                                                                       Vol. 11, No. 13; 2019

         (a) Mechanicall damage          (b) Deteriorration by humiddity           (c) Bed bug ddamage
  Figure 11. Damage pattterns found in soybean embrryos submittedd to the tetrazollium test. Sourrce: França Ne
                                                                                                              eto,
                                         Krzyzanowsski and Costa (1998)

The staininng tonality off embryos tissuues submitted to tetrazolium
                                                                   m test can be uused to classify
                                                                                                fy seeds into viable
                                                                                                              v
(high and low vigor) annd non-viable.. Observing suuch color nuaances, combineed with the knnowledge of seeds  s
characterisstics, enables the determinaation of presennce, location, and nature off the damagess that can occur in
embryonicc tissues (Moorre, 1973; Frannça Neto, Krzyzanowski, & C   Costa, 1998).
Seeds are considered non-viable
                        n           wheen contain meechanical dam
                                                                mage, sting byy bed bug, annd deterioratio   on by
humidity iin regions of apical
                        a                           mbryonic axis, along the cenntral cylinder, and in the vascular
                               and roott meristem, em
region (IST
          TA, 1985).
Although tthe establishedd criteria are ffollowed (ISTA A, 1985; Françça Neto, Krzyyzanowski, & C      Costa, 1998), color
perceptionn (tetrazolium test,
                           t   fast greenn) varies from oone individuall to another, ass well as the crriteria of normality/
abnormalitty seedling annd vigor, resulting in greateer or less classsification rigoor. Thus, toolss that minimiz     ze or
eliminate tthe implied subbjectivity in soome tests woulld contribute too the best inforrmation credibbility and guaraantee
results, alsso in addition to
                           t reduce the tiime spent for cclassification.
Techniquees to analyze images are preesented as a prromising technnology due to the improvem       ments in systemms of
capture, prrocessing, and interpretationn of images. Thhe agility in processing plus the reduction oof acquisition costs
have madee the image annalysis systemss that perform classification and evaluationn of agriculturral products a more
attractive aautomated wayy. These technniques allow grreater precision in the analyssis, intrinsic too the computattional
processingg, besides the extraction
                         e          of innformation, whhich is impossiible to be obseerved (Carvalhho, 2010).
Image seggmentation is defined as a decompositioon process intto non-interseecting regions,, where a lab        bel is
assigned too these regions (pixel set), w
                                        which share soome visual chaaracteristics. M
                                                                                   Most traditionall targeting metthods
are based only on pixxel intensity information, bbut human beeings, naturallly, use other information when        w
performingg segmentatioon. For this rreason, recenttly, trainable segmentation methods havve emerged as          a an
important alternative too improve the accuracy of tthe areas labeeling process (Alekseev, Orrlova, Rozalie      ev, &
Zaboleevaa-Zatova, 20144; Arganda-Carrreras, Kaynig, Schindelin, C  Cardona, & Seung, 2014).
As stated by Cícero annd Banzatto-Juunior (2003) w when using im  mage analysis techniques too characterize seed
damage, thhis promising procedure
                        p         is possible becausse each seed iss analyzed indiividually, sugggesting in detaiil not
only the ddamaged areass’ analysis buut also the loccation and exttent of their ddamage, with a good degre      ee of
accuracy.
                                                          which digital prrocessing and image analysiss were used to help
In this reseearch field, theere are scientiffic papers in w
improvingg information and/or
                           a       image eextraction inforrmation in seedd quality tests..
Corrêa, Cíícero, and Abbud (2017) havve evaluated tthe cottonseedd vigor througgh computer image analysiss and
concludedd the system is a viable alternnative to assesss the seed vigoor. Pinto, Cicerro, França Netto, and Forti (2
                                                                                                                    2009)
Evaluated mechanical and  a stink bug damage in sooybean seed uusing x-ray annd concluded the technique was
effective inn detecting thee damages studdied in soybeann seed.
Pereira, Saito, and Buggatti (2016) sttudied an imagge analysis frramework classsification of seed damagess and
showed thhat the proposeed frameworkk presented nootable contribuutions to the soybean vigor definition pro
                                                                                                          ocess,
allowing too automate andd accelerate it the process.
Silva, Barrros, Albuquerque, Marcos-F    Filho, Gomes Junior, and Nascimento (20017) evaluatedd squash seed vigor
                                                                                                            v
using commputerized imaage analysis annd concluded that system prrovided data oon indexes of seedling vigorr and
uniformityy of seedling development and on seeddling growth indicating thaat automated process a relliable
alternativee for obtaining consistent infformation on thhe physiological potential off seeds.

                                                            187
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                             Journal of Agricultural Science                       Vol. 11, No. 13; 2019

Given this context, it is possible to use computational systems based on digital processing and image analysis to
help obtaining information from quality tests carried out on the studied seeds.
2. Methodology
Tool development and experimental tests were carried out in the Seed and Plant Evaluation Laboratory from
UNIOESTE (Western Paraná State University)-Campus Cascavel-PR and in the Teaching Laboratories of
Biology and Chemistry from UTFPR (Federal Technological University of Paraná), Campus Santa Helena-PR.
The proposed tool, named SATTz (System of Support for Tetrazolium Test), consists on a base system to
classify seeds submitted to the tetrazolium test due to the need for an automation source for this test, to assist the
analyst in the decision-making process to classify damages present in soybean embryos samples, and to carry out
the test control during the batches’ evaluation.
Thus, specific tools were used for each step and, to reduce development cost, open and/or free source
technologies were chosen since an open source software is ideal in scientific development, as it can be freely
reviewed, modified, and redistributed (Schindelin, Rueden, Riner & Eliceiri, 2015).
Java object-oriented programming language version 1.8.0_91, in the Integrated Development Environment
(Eclipse) version 4.5.2, was applied for this system development. This language was chosen because it is a
multiplatform, which allows using and integrating libraries and frameworks for system development and digital
image analysis.
The maintenance of data was carried out in the MySQL data file managing system, version 6.3. This system uses
a relational database, thus, it was necessary to draw a mapping system of an oriented model to the developed
application object for the relational model of MySQL data file. This mapping used Hibernate version 4.3.5, open
source framework for object relational mapping, which allows keeping an object-oriented application. Also, that
changes on the database do not imply in profound consequences on their application (O’Neil, 2008).
The ImageJ platform is a software distributed freely under GPL (General Public License) and has a dynamic
users’ community, made up of researchers from several areas of knowledge. Its use allows several applications,
ranging from data visualization to advanced image processing and statistical analysis. Due to its extensibility, it
attracts biologists and computer scientists who efficiently implement specific image processing algorithms
(Schindelin et al., 2015).
Libraries of the FIJI software was used in the process of image analysis, an ImageJ software distribution that
adds several functionalities, which make easier scientific image analysis. This software was proposed as a
productive collaboration platform between computer science researchers and the research communities on
biology (Schindelin et al., 2012).
To perform the image segmentation process, the TWS (Trainable Weka Segmentation) plug-in was used
(Arganda-Carreras et al., 2016), which is an integrating part of FIJI software. This plug-in act as a link between
the machine learning fields and the digital image processing, providing the framework needed to use and
compare the classifiers that make the image segmentation. This combines a collection of machine learning
algorithms with a set of image characteristics to produce pixel-based segmentations, by using a set of methods to
extract statistical properties from an image based on pixel samples (Arganda-Carreras et al., 2011;
Arganda-Carreras et al., 2014). The WEKA (Waikato Environment for Knowledge Analysis) is an open source
software consisting of a range of machine learning algorithms destined for data mining, which includes tools for
data preprocessing, classification, regression, clustering, rules of association and visualization (Hall et al., 2009).
All WEKA classification, regression, and clustering algorithms can be used by TWS (Arganda-Carreras et al.,
2014).
In this trial, the classification algorithm based on Random Forest decision trees was used, a classification and
regression technique created by Breiman (2001) and composed of a set of decision trees, in which the class
prediction for new values is based on a voting system. After the generation of many trees (forest), the class is
chosen, based on most tree votes (Breiman, 2001). According to the same author, a Random Forest is a classifier
consisting of a set of tree-based classifiers, formally described as h (x, Θk), where h is the decision tree, x is the
input to be sorted and Θk is the k-th identically distributed random vector. Thus, each tree votes in the most
popular class for x input to be classified.
Random Forests are obtained by means of a multi-variant version of a predictor, known as bagging
(bootstrapping aggregating) (Breiman, 1996), where the final forecast is performed by the average B predictions
(Equation 2) or by taking the majority vote (Equation 1) (Goldstein, Polley, & Briggs, 2011).

                                                         188
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                                    Journal of A
                                                              Agricultural Sciience                      V
                                                                                                         Vol. 11, No. 13; 2019

                                                                     T1         t
                                                            m (x) = ∑t=1 f ((x
                                                      fmedium
                                                                      T
                                                                                                                          (1)
          t
where,   f (xx):   function witth the characteeristic to be stuudied; T: Numbber of trainingg samples.
                                                                  1         *b
                                                       fbag x =       ∑Bb=1 f ((x)                                        (2)
                                                                  B
          *b
where,   f ((x):   function with the charactteristic to be sttudied; B: Num
                                                                          mber of bootstrrap samples.
The basic algorithm of trree constructioon is the Classsification and R
                                                                      Regression Treees (CART) (B  Breiman, Freaddman,
Olshen, & Stone, 1984)), which has aas particularityy the ability too generate higgh-performancce trees, with great
capacity oof generalizatioon, using as ccriteria for thee best feature selection the IImpurity (Equuation 3) and GINI
                                                                                                                 G
(Equation 4) indices (HHan, Pei, & Kaamber, 2003). In the forestss, the CARTs are expandedd to the maxim        mum,
without prruning proceduure.
                                                                           ωi                  ωi
                                              Impurity N = - ∑M
                                                              i=1 P                 llog 2 P                              (3)
                                                                           t                   t
where, N: set examples;; M-number of classes; ωi: nnumber of exaamples at ith ppartition; t: totaal number of N set
examples.
                                                                             2
                                                        Gini(N) = 1 – ∑M
                                                                       i=11 Pi                                            (4)
where, Pi : frequency of each class in eeach knot; M: number of claasses.
The TWS plugin allow     ws the extractioon of image characteristicss by edge deteectors, texturee filters, and noise
                                                                                                                   n
reduction. SATTz extraacts the imagges characterisstics using TW       WS/FIJI core,, by HSB coolor attributes, and
delimiting a radius (1 too 8 pixels) arouund each pixel of the imagee to obtain the histogram of that circle, div vided
into four pparts (32, 64, 128, 256 intervvals in the histoogram). Then, the system callculates the Enntropy (Equatio
                                                                                                                   on 5)
of image ddata. This proccedure is parammount in the classifier learninng process.
                                                           E = ∑ – plog2 (pp)                                             (5)
where, p: pprobability of each histogram
                                       m part.
The propoosed system (F   Figure 2) is ccomposed of a graphical usser interface, w     which allows, in addition to
                                                                                                                  o the
training off a classifier through
                          t       the TW WS interface and the printinng of the tetraazolium test evaluation card d, the
registrationn and manipullation of the daata on batches,, samples, repllications, and sseeds.

                                      Figure 22. Analytical syystem manageement framewoork
The tetrazolium test wass carried out oon soybean seeeds of cultivars from agriculltural propertiees, as well as seeds
                                                                                                                   s
known as carriers of speecific damagess. They were pprovided by ann official seed analysis laboraatory in the region.
The stockk solution at 1.0%
                          1     was preppared by mixiing 10.0 g off tetrazolium ssalt in 1.0 L distilled waterr. As
suggested by Associatioon of Official Seed Analystss [AOSA] (1983) 100 seedss of each sampple were used (two
subsamplees of 50 seedss each). The sseeds were paacked in moissture germinattion paper andd kept under these
conditionss for 16 hours in a BOD (B    Biochemical O Oxygen Demannd) type oven at 25 °C. Aft      fter this period
                                                                                                                  d, the
seeds weree put in plasticc bags to receivve 0.075% (Eqquation 6) tetraazolium solutioon to stay com
                                                                                                   mpletely subme ersed.
Afterwardds, the temperature was betweeen 35 and 40 °C for 150-1880 minutes (2.55 to 3 hours) uuntil the stainin  ng.
                                                    TS0.075% = SS75ml + D
                                                                        DW925ml                                           (6)
where, TS: Tetrazolium solution; SS: S
                                     Stock solutionn; DW: Distilleed water.
After the iideal staining, the seeds weree removed from m the solutionn, washed with tap water, andd cut longitudinally
through thhe embryonic axis,
                          a     with a sccalpel. The cottyledons were arranged on a blue and squuared sheet of EVA

                                                                 189
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                                 Journal of A
                                                           Agricultural Sciience                     V
                                                                                                     Vol. 11, No. 13; 2019

(Ethyl Vinnyl Acetate) (F
                        Figure 3). Eachh cotyledon paair was placed in a rectanglee (2.5 × 2.8 cm
                                                                                                m) of pattern. Itt was
necessary to previously train a classiffication model as the TWS uuses superviseed classificatioon technique. Thus,
                                                                                                                T
embryos ssamples (in reeplication imaages) were selected, represeenting the maajor information about dam      mages
(bedbug, mmechanic, andd humidity) annd healthy seeeds that take ppart on the tesst to create a new image, called
                                                                                                               c
training im
          mage (Figure 3c). It was obtaained from the composition oof 16 embryoss image, selecteed from the immages
of repetition I (Figure 3a)
                        3 and II (Fiigure 3b) that satisfactorily described thee pattern in thhe tetrazolium test,
covering mmechanic, hummidity, and beddbugs’ damages, and healthy seeds.

                  (a) Repettition I                    (b) Repetition II                  (c) Traininng
                                      Figure 3. Arrangemennt of embryos in EVA patterrn

The imagees were obtaineed using a Sonny HX200V 188.2 Mega Pixeels camera withh Charge-Couppled Device (C CMD)
Exmor R  R™ CMOS Sensor,  S       withoout flash. Foor the traininng stage of the classifierr, TWS interface
(Arganda-Carreras, 2016) was used, allowing chaaracteristics exxtraction and classifier traiining for attrib
                                                                                                           butes
related to tthe tetrazolium
                          m test.
The characcteristics of the analyzed imaages are color in HSB patterrn and Entropyy with rmax = 8 and rmin = 1. The
used classsifier was the FastRandomF  Forest beginninng with 200 trrees and two random charaacteristics per knot.
FastRandoomForest is a multitaskingg reimplementtation of Ranndom Forest, created by F           Fran Supek, which
                                                                                                                w
optimizes speed and mem   mory usage (AArganda-Carreeras, 2016). Thhis process aim
                                                                                 ms to extract thhe characteristiics of
images useed in the systeem and the buuilding of deciision trees, bassed on the chaaracteristic vecctor constructe
                                                                                                                 ed by
the previouus process.
3. Results and Discussion
The set used has severral damages thhat, in some cases, were fo found in the ssame embryoss. The regionss that
presented background patterns
                       p        were identified in the training iimage, bed buug damage, deeterioration du
                                                                                                         ue to
humidity ddamage, and tissue without lesions.
During thiis process, 39 regions
                           r        from tthe bottom class of image (668.236 pixels),, 141 regions oof the class without
damage (448.005 pixels),, two regions iin mechanic ddamage class ((470 pixels), 559 regions in ddeterioration due to
humidity ccategory (30.3354 pixels) andd eight regionns related to thhe class of dam   mage per bedbbug (2.775 pixxels),
creating a training data set
                          s with 143,2991 instances too describe the cclasses. The trraining was perrformed to obttain a
classifier ppattern and, as a result, the cllassified trainiing image was obtained (Figuure 4).

                                                              190
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                           Journal of A
                                                     Agricultural Sciience                   V
                                                                                             Vol. 11, No. 13; 2019

                       ( MD
                       (a)                                                        (b) HD

                      (cc) BBD                                                    (d) BBD

                       ( HD
                       (e)                                                        (f) BBD

                       ( HD
                       (g)                                                        (h) WD

                      (ii) BBD                                                    (j) HD

                       ( HD
                       (k)                                                        (l) HD

                      (m
                       m) HD                                                      (n) HD

                       ( MD
                       (o)                                                        (p) HD
 Figure 4.. Training imagges with mechhanic damage ((MD), humiditty deteriorationn (HD), bedbuug damage (BBBD),
   and emmbryos withouut damage (WD    D), with respecctive segmenteed images: Greey-backgroundd; Red-humidity
                deteriorattion; Green-beedbug damage; Blue-mechannic damage; Yeellow-vigor tisssue

                                                        191
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                               Journal of A
                                                         Agricultural Sciience                        V
                                                                                                      Vol. 11, No. 13; 2019

The registtration of batcch samples annd their replications (consissting of two im       mages of tetraazolium test) were
carried outt through the batch
                         b      creationn interface, afteer obtaining thhe classificatioon pattern that serves as basiis for
inference in any soybeaan cotyledons sample. Oncce the batch w         was registered with its resppective sample   e, the
informatioon was extraccted from cotyyledons contaained in their replications. The replicatioons I and II were
analyzed aas from the classification
                         c               interface (Figgure 5), usingg the system individual sellection tool in      n the
boundariess where there was
                          w a seed cottyledons pair, hence obtaininng informationn about the em       mbryos displayed in
         mation extractioon interface to help on the classification prrocess of the teetrazolium test.
the inform

                                  Figure 5. Soybean em
                                                     mbryos classifiication interfacce

The informmation extracttion interface of embryos (F        Figure 6) alloows the decisioon-making (viigor (Note 1) class
definition to the seed) from the cotyyledons imagee segmented iinto classes (hhealthy seed, mechanic dam           mage,
bedbug daamage, and hum   midity deteriooration) as well as the percenntages of pixells classified inn these classes. It is
also possibble to make a visual comparrison between segmented annd original embbryos (Figure 6b) to improv         ve the
analyst’s innference proceess or to verifyy if the classifiication has beeen correctly performed.

                                  (a) Embbryos informatioon                                  (b) Originnal image
                                  Figure 6. Embryos innformation extrraction interfacce

                                                              192
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                                  Journal of A
                                                            Agricultural Sciience                         V
                                                                                                          Vol. 11, No. 13; 2019

To improvve the predictivve properties oof the classifierr pattern, durinng the classificcation process of the 100 seeeds in
the test, thhe characteristiics extraction was carried ouut in some immages to feed thhe training sett. At the end of o the
classificatiion process, thhe data set recoorded 230.388 instances (addding 87.097 insstances to the iinitial tests sett).
The imagees labeled by the pattern arre shown in F  Figures 7a andd 7b, whose percentage of ppixels classifie  ed as
healthy seed was higherr than 99.33% and representeed seeds of viggor class 1. Thheir external ccolor characteristics
are uniform
          m and superficcial, indicatingg slow penetraation of the tettrazolium salt (França Neto, Krzyzanowsk ki, &
Costa, 19998).

                  (a) Healtthy seeds (99.333%)                                      (b) Heaalthy seeds (99.662%)
                              Figure 7. S
                                        Segmented imaages classifiedd as without damage

It can be oobserved that in Figure 7a, the soybean eembryos recordded 99.33% of pixels labeleed as healthy seeds
                                                                                                              s
and 0.67%% with humiditty deteriorationn. When comppared with the original imagee, there were ttwo small stria
                                                                                                             ations
of a more intense red coolor in one off the cotyledonns, originatingg from humiditty deterioratioon. The embryos of
Figure 7b did not show   w any deteriorration. Howevver, some pixxels were classified as beloonging to hum   midity
deterioration class (0.35%
                         %).
The interppretation of huumidity deteriioration in tettrazolium test can be descrribed by somee patterns, suc  ch as
striations ((evidenced by the test in darrk red carminee), caused by compression oof integument on cotyledonss as a
function of expansion (hhydration) and contraction (ddehydration) cyycles, whose results are wrinnkles in integument
and lesion in cotyledons. This kind of pattern may bee found in embbryonic axis annd cotyledons..
Areas of mmore intense red
                        r or milky w    white staining are also charaacteristic patterrns of this detterioration. Mo
                                                                                                                   ostly,
they are a result of smalll striations inccrease and cauuse a greater ddamage (Françça Neto, Krzyzzanowski, & Costa,
                                                                                                                   C
1998).
It can alsoo be seen in Figures
                          F         8a andd 8b images w whose most appparent charactteristics, labeleed by the classifier
pattern, aree those that givve to pixel hummidity deteriorration patterns.

                                                               193
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                                 Journal of A
                                                           Agricultural Sciience                         V
                                                                                                         Vol. 11, No. 13; 2019

           (a) Humidity deeterioration on thhe opposite
                  region to embryonic axxis                              (bb) Humidity detterioration on em
                                                                                                          mbryonic axis

                       Figure 8. Segmennted images wiith humidity deeterioration chharacteristics

In Figuress 8a, it is possible to observve the number of pixels classsified as humiidity deterioraation in cotyled
                                                                                                                 dons,
clustered iin the oppositee region to embbryonic axis. In the non-segm
                                                                     mented image,, striations andd bands can be seen
in these regions, characteristics of hummidity deterioraation.
In Figures 8b, it is obserrved that the em
                                         mbryonic axis also presentedd pixels classiffied in this waay. When comp pared
with the ooriginal image (without damaage), it is posssible to observve, both in thee cotyledons annd embryonic axis,
striations aand bands in darker
                          d      red carmmine color annd whitish partts (dead tissuee), which are aaspects of hummidity
deterioration. When looking at imaages set that presented hum          midity deteriooration, it is observed that the
segmentatiion generated by the system corroborates w   with the real loocation and exxtension of dammage in the original
images.
The mechaanic damage result
                        r       from immpacts, cuts, annd abrasions arrising from the harvesting, ttrail, transport, and
processingg process identtified in the tettrazolium test presents charaacteristic patterrns such as craacks and abrassions.
When receent, they are coolored in darkk red carmine, or white in caase they have nnot been recenntly identified, both
in cotyleddons as in emmbryonic axis (Bewley & B       Black, 1994; F França Neto, K   Krzyzanowskii, & Costa, 1998).
Characteriistic embryos with
                        w this kind oof damage are presented in F      Figures 9a to 99b.
Pixels classsified as mecchanic damage can be obseerved in one cotyledon (Figgure 9a), in thhe vascular re           egion
adjacent too the attachmennt point. Theree has been recoorded humidityy deteriorationn in cotyledonss in embryonic   c axis
                          o the other haand, it is not distributed throughout its lenggth. It can be sseen, in the original
and its oppposite region; on
image, thaat the mechanicc damage came from an abraasion in seed.
In Figure 99b some pixelss classified as mechanic dam mage can be seeen in the regioon just above hhalf of the embbryos
(vascular rregion), plus a substantial exxtent of deteriioration due too humidity disstributed betweeen the cotyled
                                                                                                                  dons.
And whenn analyzed thee original imagge, it was recorded that meechanic damagge is due to seeeds kneading        g that
happened during some of the producction processees, and that thhe seed also hhas deterioratioon patterns du       ue to
humidity iin its cotyledonns.
In general,, pixels with thhis pattern of ddamage were w
                                                       well identified and classified.

                                                              194
Classification and Pixel-Based Segmentation to Evaluate Soybean Seeds Submitted to Tetrazolium Test
jas.ccsenet.org                                     Journal of A
                                                               Agricultural Sciience                          V
                                                                                                              Vol. 11, No. 13; 2019

           (a) Mechanic daamage in vascullar region                            (b) Mechanic ddamage in centraal portion
                             F
                             Figure 9. Segm
                                          mented images with mechanicc damage charracteristics

The damagge caused by bedbug
                        b       sting iis easily identiified by typicaal lesions, withh circular patteerns and whitish or
yellowish color, as well as an intense red carmine rring around it (França Neto,, Krzyzanowskki, & Costa, 1998).
The embryyos that were segmented
                       s           withh this characteeristic pattern aare shown in F
                                                                                    Figures 10a andd 10b.

                  (a) Sting beddbug in vascularr region                          (b) Sting bedbbug below vascuular region
                               Figure 10. Seegmented imagges with bedbuug sting characcteristics

Pixels classsified as bedbbug damage arre found in cottyledons, as obbserved in Figuure 10a. They are in the sup  perior
part of thee soybean embbryo in a cotyleedon and a smmall part in thee central positiion of the otheer cotyledon. Pixels
                                                                                                                  P
observed aaround this daamage are alsoo classified as humidity deteerioration. Pixxels classified as bedbug dam   mage
reach a noon-vascularizedd region of cootyledons, as ccan be observeed in Figure 10b, in the region opposite to the
embryonicc axis. It is poossible to obseerve that pixels were correcctly classified by the visuall analysis from  m the
original immages of thesee segmentationns, both for bbedbug damage and humiditty deterioratioon. Even thoug       gh in
some cases the red ring around
                          a      the dam
                                       mage was idenntified as a deteerioration patteern by humiditty.
After the embryos weree classified, thhe results interrface of the teetrazolium testt was generateed by the repo    orting
interface. T
           This data sheeet is based on tthe tetrazoliumm test data file (França Neto, Krzyzanowskki, & Costa, 1998).
The systemm generates thhis data file auutomatically, coonsidering thee previous classsification, in w
                                                                                                     which viabilityy and
vigor perccentages are reegistered. It alsso identifies annd quantifies tthe kinds of daamages, as weell as highlightts the
most apparrent ones.
After seleccting the batchh and sample, the system gennerates a Portaable Documennt Format (PDF
                                                                                              F) file with the
                                                                                                             e test
results (Figgure 11).

                                                                  195
jas.ccsenet.org                              Journal of A
                                                        Agricultural Sciience                      V
                                                                                                   Vol. 11, No. 13; 2019

                                          Figure 11. Tetrazolium test report

The reportt determines the
                         t percentagee of seeds, claassified in eacch viability cllass, and the ssoftware calcu   ulates
germinatioon potential upp from the sum
                                      m of estimated vvalues in the cclasses from 1 to 5, and the vvigor is obtaine
                                                                                                                  ed by
the sum oof the levels from 1 to 3, following w     what França N    Neto, Krzyzannowski, and C    Costa (1998) have
establishedd.
4. Conclussion
Concerninng the developeed software, itt can be concluuded it presentts an alternativve to performinng a less subje
                                                                                                                  ective
seeds classification for tetrazolium
                           t           tesst, since perceentages extracttion of classifi
                                                                                      fied pixels for damage classe
                                                                                                                  es by
the segmented images provides
                         p         valuabble informatioon in the decission making. M   Moreover, Auttomated generration
of tetrazolium test reporrt from the classsified samples saves time inn the extracting process of teest information
                                                                                                                  n and
shows itseelf as a potentiaal automation ttool in researcch and commerrcial environm ments.
Acknowleedgements
Authors aare thankful to t Western PParaná State UUniversity, Feederal University of Technnology-Paraná and
Coordinatiion for the Im  mprovement oof Higher Edducation Persoonnel/Araucáriia Foundationn (CAPES/FA) for
supportingg this research..
Referencees
Alekseev, A. V., Orlovaa, Y. A., Rozaaliev, V. L., & Zaboleeva-Z Zotova, A. V. ((2014). Two-SStage Segmentation
    Methood for Conteext-Sensitive IImage Analysis (pp. 331-3340). Paper ppresented at tthe eleventh Joint
    Confe
        ference JCKBS SE, Volgograd,, Russia. https://doi.org/10.1007/978-3-3199-11854-3_28
Arganda-C
        Carreras, I., Cardona,
                       C          A., Kaynig, V., Rueden, C., & Schindelinn, J. (2011). Trainable WEKA
                                                                                                  W
                                                                  Weka_Segmentation
    segmeentation. Retriieved from httpps://imagej.nett/Trainable_W
Arganda-CCarreras, I., Kaynig,
                       K        V., Rueden, C., Schindelin, J., Cardona, A., & Seunng, H. S. (2         2016)
    Trainnable_Segmenttation: Releasee v3.1.2 Retrieeved from httpss://zenodo.org/record/59290#.XO6-cohKiU
                                                                                                         Uk

                                                           196
jas.ccsenet.org                             Journal of Agricultural Science                      Vol. 11, No. 13; 2019

Arganda-Carreras, I., Kaynig, V., Schindelin, J., Cardona A. H., & Seung, S. (2014), Trainable Weka
    Segmentation: A machine learning tool for microscopy image segmentation. Retrieved from
    http://bit.do/eTHEL
Association of Official Seed Analysts. (2008). Seed vigor testing handbook. East Lansing: AOSA.
Bewley J. D., & Black M. (1994). Seeds. Physiology of development and germination (2nd ed). New York, NY:
    Plenum Press. https://doi.org/10.1007/978-1-4899-1002-8
Breiman, L. (1996). Bagging predictors. Machine Learning, 24(2), 123-140. https://doi.org/10.1023/A:101805
    4314350
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32. https://doi.org/10.1023/A:1010933404324
Breiman, L., Freadman, J., Olshen, R., & Stone, C. (1984). Classification and regression trees. Grove, CA:
    Wadsworth International.
Carvalho, M. L. M. (2010). Utilização da análise de imagem-conceitos, metodologias e usos. Informativo
    ABRATES, 20(3), 45-47.
Cícero, S. M., Banzatto-Junior, H. L. (2003). Avaliação do relacionamento entre danos mecânicos e vigor, em
    sementes de milho, por meio da análise de imagens. Revista Brasileira de Sementes, 25(1), 29-36.
    https://doi.org/10.1590/S0101-31222003000100006
Corrêa, P. D., Cicero, S. M., & Abud, H. F. (2019). Evaluation cottonseed vigor through computer image
    analysis of the seedlings. Journal of Seed Science, 41(1), 54-59. https://doi.org/10.1590/2317-1545
    v41n1200649
Dall’Gnol, A. (2000). The impact of soybeans on the Brazilian economy. Technical information for agriculture.
     São Paulo: Máquinas Agrícolas Jacto.
Deswal, D. P., & Chand, U. (1997). Standardization of thetetrazolium test for viability estimation in rice bean
    (Vigna umbellata (Thunb.) Ohwi & Ohashi) seeds. Seed Science and Technology, 25(3), 409-417.
FAO (Food and Agriculture Organization of the United Nations). (2010). Seeds in emergencies: A technical
   handbook. FAO Plant Production and Protection Paper 202. FAO, Rome, Italy.
França Neto, J. B., Krzyzanowski, F. C., & Costa, N. P. da. (1998). The tetrazolium test for soybean seeds.
    Londrina, PR: EMBRAPA-CNPSo.
Goldstein, B. A., Polley, E. C., & Briggs, F. B. (2011). Random forests for genetic association studies. Statistical
    Applications in Genetics and Molecular Biology, 10(1), 32. https://doi.org/10.2202/1544-6115.1691
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The WEKA data mining
     software: An update. ACM SIGKDD Explorations Newsletter, 11(1), 10-18. https://doi.org/10.1145/
     1656274.1656278
Han, J., Pei, J., & Kamber, M. (2011). Data mining: Concepts and techniques. Waltham, MA: Elsevier.
ISTA (International Seed Testing Association). (1985). International rules for seed testing. Seed Science and
    Technology, 13, 229-526.
Marcos-Filho, J. (2015). Seed vigor testing: an overview of the past, present and future perspective. Scientia
    Agricola, 72(4), 363-374. https://dx.doi.org/10.1590/0103-9016-2015-0007
Moore, R. P. (1973). Tetrazolium staining for assessing seed quality. In W. Heydecker (Ed.), Seed ecology (pp.
   347-366). London, Butterworth.
O’Neil, E. J. (2008). Object/relational mapping 2008: Hibernate and the entity data model. Paper presented at
   ACM SIGMOD international conference on management of data (pp. 1351-1356). ACM, New York.
   https://doi.org/10.1145/1376616.1376773
Pereira, D. F., Saito, P. T. M., & Bugatti, P. H. (2016). An image analysis framework for effective classification of
     seed damages (pp. 61-66). Paper presented at the 31st Annual ACM Symposium on Applied Computing.
     ACM, New York, NY. https://doi.org/10.1145/2851613.2851637
Pinto, T. L. F., Cicero, S. M., França Neto, J. B., & Forti, V. A. (2009). An assessment of mechanical and stink
     bug damage in soybean seed using X-ray analysis test. Seed Science and Technology, 37(1), 110-220.
     https://doi.org/10.15258/sst.2009.37.1.13

                                                         197
jas.ccsenet.org                              Journal of Agricultural Science                        Vol. 11, No. 13; 2019

Roessing, A. C., Dall’Agnol, A., Lazzarotto, J. J., Hirakuri, M. H., & Oliveira, A. B. (2007). O complexo
    agroindustrial da soja brasileira. Circular Técnica, 43.
Schindelin J., Arganda-Carreras I., Frise, E., Kaynig, V., Longair, M., Pietzsch, T., … Cardona A. (2012). Fiji:
    An open-source platform for biological-image analysis. Nature Methods, 9(7), 676-682. https://doi.org/
    10.1038/nmeth.2019
Schindelin, J., Rueden, C. T., Hiner, M. C., Eliceiri, K. W. (2015). The ImageJ ecosystem: An open platform for
    biomedical image analysis. Molecular Reproduction and Development, 82(7), 518-529. https://doi.org/
    10.1002/mrd.22489
Silva, P. P., Barros, A. C. S. A., Marcos-Filho, J., Gomes Junior, F. G., & Nascimento, W. M. (2017).
     Assessment of squash seed vigor using computerized image analysis. Journal of Seed Science, 39(2),
     159-165. https://dx.doi.org/10.1590/2317-1545v39n2171177

Notes
Note 1. The assignments of seed vigor classifications based on the information extracted by the system were
according to the methodology of tetrazolium test for soybean seeds (França Neto, Krzyzanowski, & Costa,
1998).

Copyrights
Copyright for this article is retained by the author(s), with first publication rights granted to the journal.
This is an open-access article distributed under the terms and conditions of the Creative Commons Attribution
license (http://creativecommons.org/licenses/by/4.0/).

                                                          198
You can also read