MEALINESS DETECTION IN APPLES USING TIME RESOLVED REFLECTANTE SPECTROSCOPY

C. VALERO . P. BARRhIRO, M. RU1Z-ALTISHNT , R. CUBEDDÜ , A. PlhFERI , P. TARONI , A. TOKRJCELU , CJ. VALENTJN] , D. JOHWSON

aodC. DOVER

Department Ingeniería Rural, E.T.S.f. Agramases Universidad Politécnica Madrid

Avenida Complutense, 2H040 Madrid Spaín

INFM-Dipartimento di Física and ÍFN-CNR Politécnico tfi Milano, Piapía

Leonardo da Vinri 32, 20133 Milán Italy

East Mailing Research East Mailing, Kent. MEJ90H.I

U.K.

ABSTRACT

Mealiness is a textura! (turibule related to infernal fruil disorder ihal is characterized by íhe combination of abnormal softmss of ihe fruít and absence of frec jukiness in ¡he moulh when eaten by the consumer. Time-resolved láser refieclance spectroscopy was used as a tool lo determine mea-liness. This ncw techniaue in agrojbod research may providc physical and rhemical informaron independen;ly and simuUaneousiy. which is relevan! lo charaeíerize mealiness. Using visible and netir infrared tasers as ligtit sources, time-resolved láser reflectaría? spectmscopy was applled to Golden Delicious and Cox appies (r\ = 90), to charaeíerize batches of untreated síimpk's and samples ihai v/ere stored under conditions that promole ihe development of meaiiness (20C & 95% RH). The collected datábase was clustered into dijferent groups according to iheir instrumental lesi valúes. The óptica! copffic.ients were used as expUmatory variables lo bidld discriininam functions for mealiness. The performance ofthe classification modeís creuted rangedfrom 47 to i 00% of corree tly identified mealy versas nonmealy appies.

KKYWOKDS

Apple, láser reHeclance spectroscopy. mealiness, near infrared spectros­copy. time-resolved spectroscopy, visible spectroscopy

INTRODUCTION

Consumers' decisión at purchase is affected bolh hy exlerna] aspect and internal quality of fruits. Among the qualily parameters that a consumer can flnd in apples, mealiness is a inain issue. Ii has been defined as a negative attribute of sensory íexturc thal combines the feeling oi' disaggregaled tissue with lack oí juiciness. II is a consequence oí a biochemical process resulling in pectin degradalion in the middle lamellae (Gross and Sams 1984; Von Mollendorf et ai. 1993), and Ibeir consequences are visible by microscopic observaron (De Smedl el al. 1998). Mealiness is associaled with late harvesl and long-lerm slorage (Fisher 1943; Harkerand Haücll 1992). Mealiness onsel may also be accelerated by Lenvperaiure treatmenls (Von Mollendorf et al. 1992) combiried with extremely high relalivc humidity (De Smedl 2000).

The characterization of mealiness has been done tradilionally by means of sensory panels defining and grading sensorial descriptora (Bignami a al. 2003). Howcvcr, relations belween physiology changes, frait lexlure and sensory measuremeiHs of tissue slrength and juiciness have been eslahlished successfully (Paoletti et ai. 1993; Harker el al 1997; Verlinden eí ai 1997). Correlations of sensory aüributes (crispness and juiciness. a! first hite and during chewing) to a eomhinatian of parameters measured using instrumental tests (coníinedcompression of probes) have been establislied. Wiihin the same work (Barreiro et al. 1998) a redermitiort of mealiness in apples and peaches in terms of rheological properties was proposed lollowing sensory descriptors in mealiness perception. A new instrumental mealiness scale was created, hased on a combination of i n si rumen tal parameters like loss of crispness, of hardness and oí juiciness, which is ahle to characlerize objeclively Ihe meali­ness slate of a given sample and correlates well with sensory mealiness (Barreiro et al. 1998). More sludies can be round also relaling to peaches (Sonego el at. 1995; Orlk et al. 1999) and lomatoes (Ahrens and Huber 19H7).

Several nondeslruelivc rnethods for instrumental mealiness measurement of peaches, apples and lomatoes have been alLempled using difieren! lech-niques: near infrared spectroscopy (NJR) spectroscopy combined with low mass ¡ñipad (Ortiz et al. 2001), acoustic impulse response (De Smedl 2000), ullrasonic wave propagation íhrough fruit tissues (Bechar et al. 2005) and nuclear magneíic resonance (NMR) (Uarreiro ef at. 1999). Nevertheless, Uie

developmenl of stand-alone nondesirueiive tech ñiques is siül needed, espe-cially if Ihey are Tast and iaf thcy could be engincered inlo an aulnmatic on-line classiii catión system (Abbott 1999V

Time-resol ved reflectance spearnscopy (TRS) is a nonconventional spec-troscopic technique dial has been developed for use in medicine (Cubeddu ei al. 1994a) to characierize the optical properties of tissues, and lo lócale discrtntinuilies and alTecled áreas like liuman and animal tumors. Among its advaniages compared with more Iradilional speclroscopic lechniques, diere is the (easihility lo derive simultaneólas Iy lwo (in principie) independen! oplical parameters: the absorption oí the lighl inside the irradiated body, and ihe scattering of the photons aeross the tissue, For each waveiength. TRS gener­ales two coefficients (¿ja, absorption coefficient; and u'5í transpon scattering eoefíicient). Botfi aje dependent on waveiength, By measnring at eonsecutive waveiengths two arrays of valúes, an absorption spectrum and a "scattering spectrum," can be obtained. The working hypothesís of Ihe present work was ihat these coefficients are related, respectively, to chemical components (j.U) and to physical properties (/J'S) of the sample. Therefore, TRS can be applied to the quantiíication of cheinicals and the measurement of the rheological propenies (or the combinaron of hoth: soft lexturc and dryness, i.e., meali­ness) at Üie same lime. TRS has been appüed salisfactorily Tor ihe nondeslmc-1ive evaluation of fniil internal qualily (Valero et al. 2004a,h). Using 490 apples from different varieties, links were found between the optical properties of the samples and their qualily allribuies obtained with standard procedures: statístical models were developed for the quantiticatíon of several aspeets of l'ruil qualily (firniness estima (ion, sugar contení and acidily). In the present work, ihe ohjective was to apply TRS as a nondeslruclive inspeclion method for mealiness.

MATERIALS AND METHODS

Fruit Material

Two groups of samples were prepared:

(1) Apples with expecled "natural mealiness." In order lo enhanced ihe risk of mealiness or abnormal lexlure degradaron, apples were harvested laíe in the seasom 50 Uolden Delicious apples were picked from lwo orchards (La Alinunia de Doña Godina, Zaragoza, Spain) during ihe lasi week of October 1998 (late harvest), packed and sent lo Milano (1NFM) in November. Both firm and sofl fruits were picked, in order to obtain all possible texnjral .siafes. A bou I half of ¡he samples were ejípecíed lo be mealy as a consequence of developing mealiness on the Iree.

(2) Apples wilh mealiness induced inchamberslorage. In thiscasecolleagues at the Catholic Universily ofLeuven (Jielgium) prepared al ong theaulumn sámples froin Ihe Cox variety, harvesled early ín the season (end of Septemher) and stured iint.il Novemher 1998 under s pee i fie conditions: 20 of ihem were kept inside an ULO chamber (ultra iow oxygen) lo preserve their freshness at máximum levéis; and 20 apples were kept during 16 days in an atmospliere of 95% relativo humidity and 20C, to promota the developmenl of mealiness (Andani el al. 1999). Nol all of Ihem were expected to be finally mealy in terms of instrumental mealiness. as indi-cated by previous siudies (Barreiro e¡ ai. 1998).

The reference tests (firmness and juiciness) and TRS mensurements that were carried oui on the samples are presenled wiüiin in the same sequence as for the tests.

TUS Me&surcments

Time-resol ved reflectance speetroscopy is based on the measurement of the broadening of a short lighl pulse, transmitted across a turbid médium (fruil tissues) (Cubeddu eí ai. 1994b). The TRS equipmeni used in this work is described in delail in the study by Cubeddu. ei ni. (2001a,b). A simplified scheme of il is shown in Fig I. The lighl source is a láser heam, iherefore monoch.romatic, huí tunable al several wavelengths. The light is injected in the fruit through the intact skin by means of fiber optíes positioned ortbogonal to the equator of the fruir, Ihe light flux crosses the tissues and parí of il tinds ¡ls way out of the samplc ata particular región adjacenl to the injeciion point, This portion of refieeted light was rccovcTed with another fiber placed at 20 mm distan ce (Fig. 2). The optical paths of the phototts wilh larger probabilily of being recovered after suffering interna) refleclion form a three-dimensional lighl región wilh a semiloroidal shape inside the applc, constructed from the incoming fiber contact point to the recovering fiber contact point on the skin. If an adequate theoretical model is used for Ihe experimenta! analysis oí" dala and several hypulheses are eslablished, il is possible lo calcúlate ai the same time the absorption coefficient (pa) and the transpon scattering cocffkient (¿i',) al each wavclenglh (Cubeddu el al. 1994a). In our case, the dilíusion theory (Eq, 1) was considerad to process the TRS data poinls, and the curve wilh Ihe besi lit wasobtainedhy iteration:p;,and/Aarederived l'rum ihe curve (Fig. 3). Nolaiion in Eq. (I) is: 0(p, /), phornn fluence rale; Si.p, t), photon source: p, radial posilion; /, time; n, Índex of refraclion; e, speed of light in vacuum, Detailed descriptiou is given in the study by Cubeddu eí al. (1994u).

"^-^#Jf)-í3CMíl+íí;)r'-v^(p,/)+^fp,/) = 5(p,/) ( ] )

Ar+ (CW.ML) ÍV

Ti:Sapphire Q )

CD - Dye

^57

¡ÜT ¿z¿

<HM-

5> M

TCSPC

v~u QJ Sample II

O Scanning MCP1

MC PMT >-

FLG. 1. SCHEME 0P THE SYSTEM SBT UP KOR TRS MEASUREMENTS AT INEM LAB0RAT0KY: LINES INDÍCATE OPTICAL PATHS (HBER OPTtCS) FROM THE LASERS

(ARGÓN. DYEANDTTTANlUMiSAPPllIRE) TIIROL'Gll Tllfc SAMPLE AND TOWARDS THE DETECTOR (MULTICHANNEL PLATE PHOTDMULT1PLIER, MCP PMT), OTHER

CÜMPONENTS: TIME-CORRELATED SINGLE PI10T0N COUNTING ÜNIT (TCSPC), SCANNING MOriOCROMATOk (MC) (CUBEQDU ETAL L994A)

Fnr tJiis stiidy. the absorption and transpon scattering coefficients of footh sides of each sample were regislered at severa] wavelenglhs: far-visible (672, 750 and 8l8nm usjng diode lasers as líght sources) and NiR (f'rom 900 [o 1000 nm, al steps óf 10 nm using a tunahle láser). The nolalion íor these variables isas Ibllows: "#¿.«2" forsMarpÚ&a coefficienlat 672 nm, " ^ 7 2 " J'or fransporl scattering cocfficient at 672 nm, and so on.

Mechanical Measurements

A corrfined comprensión test was carried our ror instrumental mealíness asscssmcnt. Using a lexlure analyzer {TA.XT2, Stable Micro Systems Ltd. Surrey, UK) a máximum de formal ion oí' 2.5 mm was applicd ai 20 mm/min deíbrmation rate on cyllndrical probes of 1.7 cm heighl and diameler. Deíbr-mation was immedialeW removed al the same rale; Lwo me asúreme nls were

DifíusB reífectert ñght rjairpíSHl ítaoss tlie tample | r *» V «ad

B

Lost photürts by doep transfruESJon

'hr; l l rTI^ •"ll^ffH' lighi región (máximum prnbabiliTy palh}

FIO. 2. SCUEMbOM'Hb SAMPLE l'KbSHM'ATION POR TUS MEASU REMEN CS (Al Fruii holdtr wiüi iiianiiiiiy litar ;tiid outcomíng líber (white am>ws m pboto) in L'ouiacL witil

íiuil skiii. (tí) A i:ulre Oí i>liL>Um> was {iijuipcü iJmxiyli llic iucoraiúg lint! and Lhe [llxilims Ltil suffering absorption nr detp transmissiun were recovered.

scatlering

I 0.1

0.01

0,001

i or\ ^ \ ¡A %

,' " iV

" ¡ * ; «i ||

j P a l

- • / 1 " i i t y. i i^-f i t * »

absorption

•

V te.

1000 2000

Time (ns) 3000 4000

KIG..V TTPICAL LJAI-A ÓBTAINED WITHTRS TECHN1QUB: QVJECTED PULSE I N A R R Q W PEAK) AND RKMirCEJJ PULSE (BKOADüR CURVE, NORMAL I* KD TO THE PEAK

INTKNSLTY) ARE LATEK BEST-HTJLD TO DIEFLSION THEOKYTO OBTA1N SCATTERINO AND ABSORPTION COEFFJCIENTS fCUBEDÜU ETAL. 1<394A)

made per fruil (orte per side) using ihe average for thc subsequent analyses for this loadAinload test. Cylinders were confined in a disk that had a hole of ihe probé sizc. Thc rod used for the dompression íesí was 15.3 tnm diameler lo avoid any contact wi(h the disk during testing. A filter paper f Albet n° 1305 o)' 77.84 g/m2) ahout the si/e of the disk was placed beneath the disk in order to recover Üie juice extracled during the compression lest. The área of the juice spot was measured later on using convenlional image analysis equipmenl (Hitachi CCD black and white camera, Tokyo, Japan). The outline of the juice spot on tlie Bíter papel was cstracted (Image Pro Plus software. Media Cyber-netics Silver Spring, MD) and rhe área inside it was nsed as indicator of juiciness. The variables registered with this test and used in the statistical analyses (abhrevialions of the variables wilhin hruckets) are: hardness of the sample ("S" N/mm) expressed as the slope of the force-deformation curve during confined loading. and the juice área 07 , " mnr) recovered in the filter paper placed nnderneath the probé during the test.

Using the variables extracted from confined comprensión, the samples were labeled according to their insirumental mealiness stage, as a comhination of two possible hardness states (firm, not firm) with two possible juiciness states (juicy, nonjuicy). A sampie was considered "not firm" when S < 20 N/mm, and "firm" otherwise. Besides ti was labeled "nonjuicy" when J < 4 c m \ and "juicy" otherwise (Barreiro er al. 1998), Instrumenta] definition of tneal ¡ness nequires a sample to be "not firm" and "nonjuicy" simultaneously (Tahle I).

Statistical Analysis

As a first step, linear coirctations hetween TRS variables, hardness and juiciness were analyzed. Beeause of the poor regresston results (see Results section) and ihe Tact that the aim was malching a categórica! instrumental mealiness scale using TRS ¡nformation, Ihe statistical approach was switched to a classification algorithm.

Following an industry-oriented application for the TRS lechnique, a pass/faiJ classification was searched íoi: Discíiimnanl analysis teehjjique was

TABLE t. NOTATION OF t'RL'lT CATECiORtZATIONS ACGOROING TO THE DESTOUCTÍVE

INSTRUMENTAL MEALINBSS STATE: "EIK.M" SAMPLE WHEN 5 > 20 N/mm; "JUICY" SAMPLE II" J > 4 cm!. SAME GRAY TONE INDICATES MEMBERS1IIPOF SAME GROUP

Two elasses "Büte elasses J our classcs

Nonjuicy Juicy Nonjuicy Juicy Nonjuicy Juicy

Mol firtí» "Mealy" "Nonmealy" "Mealy" "Nunmealy" "Meuly" Firm "Nomnealy" "Nonmealy" "Nonmealy" "Frcsh" "Diy"

• SuÍL'

1 rol

used as the stattstical too! used to créate the classification models. Classiríca-lion runclions were buill wilh a stcpwi«: approach, seiecting or removingeach variable by Ihe analysis of the unique contribuüon ol'the respective variable to lhe discriminalory power of the model. The discriminalory abiliíy oí the models was evaluated either comparing the percentaje of well-classífied samples obtamed with every model, or calculaiing the similariiy índex, K {Eq. 1) used by Steinmetz el al. (1996) as a measuremenl oí the agreemenl between two instrumental classificatíons. K calculation stans (rom a elassifi-cation malrix thal is converLed into a probability inatrix where pr¡ is the probability of a sample to be elassilied by metliod A aiid B into die same caiegory i. The probabililies p-,} úhat mcihod A classifies ati objecl inlo cal-egory í whereas method B classifies Üie saine object into category j) are computed as staled in Eq. 2. Thus, 8] (Eq. 5) is the sum oT coinciden! prob-abiüties (along the main diagonal ofthe probability malrix: p„> and di (Eq. 6) is lhe sum oí lhe produets ofthe accumulated probability ol'being in each row or column (p;+ is the accumulated probability of being in row i, whereas p.n is the accumulated probability of being in colman j , Eqs. 3 and 4).

na Tí

(3) ;=i

0j = 2 > « (5)

lsEtt) (6) 6, - A ,

K 1-ff, (7)

ln this work lhe K index is used to compare the destructive classification procedure with die nondestruclive TRS method of classificalion. The advati-tage of using lhis simílarily índex is [nal, (rom iheabsolule proporlion of cases where the melliods agree (0[), il subtraets the proportiun of cases where the methods agree purely by chance (&).

ClassificatioH models weit; creaied firsi for the pooled datábase. joining bolh apple varietics. Varietal models were also created in a second phase. In al] ihe cases, calibration of Ule classilication models was performed using hall of Ibe samples for thc discriminaní analysis. Validation of lhe models was carried out widí Üie resl of Lhe dalabase.

For the creation of the classification models, in a firsi approach only iwo fruit statcs were used in order to créate two-group classifications: "mealy" and "nonmealy." A "mealyvv sample was nol ti raí and nonjuicy al lhe same lime, whereas "nonmealy" samples galhered all lhe oiher pnssibie combínations, in lilis first analysis.

Funher approaches included classificalion modela with samples classified ínlo: (1) three sTaies ("mealy" [nonjuicy and nol firm], "nonmealy" [nonjuicy or nol firm], "fresh" [juicy and firm]), and (2) four texlural stales ("mealy" [nonjuicy and nni firm], "dry" bul firm, "sofr" hnt juicy, "f'resh" [juicy and finnj). Table I synlhesizes the notation used for each comhination of hardness and juiciness slates, when crealing two-group, ihree-group or four-group models.

RESULTS

Linear Relalion between Destructiva and Nondestructive Variables

Linear eorrelalion between the variables extraeted froni Üie desiructive lest (hardness and juiciness) and lhe nondestruclive mcasurements (TRS ahsorption and seattering coefficienis) was analy/ed. In Fig, 4, a selection of TRS variables is presented in the fonn of a graphical eorrelalion matrix. Conelation coefftciems were lew (/ < 0,4) in any case, lnterna.1 eorrelalion in the TRS dala (in.sideabsorplion spectra, scailering specira. or belween ahsorp­tion and scatlering coefficients) was high (r = 0.9) in most cases. Conelalion was expecled between ahsorption variables (as is normal within N1R speetta); however, eorrelalion was noi expecled to be so high helween scattering and absorption parameters, as they refer to dilTerent physical phenomena.

Destructive Classificalion of Samples according to Instrumental Mealiness

Figure 5 shows lhe result of fruit classification into firm, soft, dry and mealy. A higher numberof mealy samples was obiained for Goiden (22 out of 50) lhan for Cox (7 ouL of 40; Fig, 5). A wíder range of lexluje varialion was found for Goiden Lhan for Cox, in ihis case. Cox is a typical exaniple of mealiness susceptible variety with very erralic mealiness onset, which is

Hardness aruj luciness

Absorpíion

S iNrmml

riftfU u%°

M». ?;u

T U

2 a.

Mi. sis

"ISK^ • * ***

M*. oon "=«" & . ü •w** * * *

n. .,-,, • Con O - >' •''

Harciness and

..;".-. i i

í [Nftrtmj

Jffiík^ CÍI 1 r . i r n r i -

MfcffJ

dtOK

á& u. «

8* &

Í^T 8 ! ** I k 590

2jfe* * ' . Mt. 950

••^v _J • COK a Gciflen

Frn, 4. IIISTOORAMS AIM11 CORHELATIONS RETWKEN SELECTED TRS VARIABLES Horizontal ases: thc lirsi column nf scalterplots, ttie ülopc [S. N/inm) of thc curve diiring ihe

mcchanical lest ís presented; Ihe sccond column, Ule j ni ce arca rccoverEd (V, mni"). Vertical axes: absorptitio and iranspurt scaltcring TRS coefíidcnrs. ai severa! wavcíeiigUis.

«c IJ

¥ 5

•

'SDfi*

O

í 2 o

1 Ó° 3

"mea/y*

natfifm s'

o

o

0 a

o 0 o

? a

O "dry

lirt-i

"soff" '/resíj"

o 8 ° « Ü C 11°

0 o ** 0

0 CD O í

"(Tipa/y' o & a " f í r y -

o nal film finn

(2 16 20 24 23 32 12 (6 20 2A 28 32

VARIETY Cent VARIETY Onlden

Hardness {'$", NAMl)

FIG. 5. HARDNESS VERSUS JUICINESS ÜF TESTED APPLES: LINES INDÍCATE D1SCRIMINATING LIMITS J 0R "DRY" TEXTURA! CATF.G0RY (JUIC1NESS < 4 cra!),

"SOFT {HARDNESS < 20 INVmm) AND THE COMBINARON OF B0T1I CMIÍALY")

eónfirmed in this case. Textura! disorder ¡n balches being dry and sofi for Cox lacks from a comparable ce!I disaggregation status.

Models to Classily Nondestructiveiy Samples Using TRS

Discrimina!» analysis functions were buill using selccied TRS coeffi-cienls (variables /i;,/^ lo ¿ia.tmo, and ¡l^tm to A'N.HKKI)

a!> expJanatory variables [or the classificadon yf samples inio (Tablel): (I) two textural categories ("mealy" versus "nonmealy"), (2) three textura! categories ("meaiy," "non-mealy," tcfresh"), or (3) Ibur textura! categories ("mealy," "dry," "son," "fresh"). The niimber of individuáis in each siage was not homogeneous and pmliiil classificalion scores oblained for each category were uneven; tlmb ilie global score of well-classifled samples in every model is a weighed average, proportional lo group sizes.

A first model was calibrated using hall'of llie pooled data of both vaiieties l'or the discrimination beíween "mealy" and "nonmealy" siates. Fifleen TRS variables were usedin the ntodel achieving apereentageof eoneclly ciassitied individual fruits oí 98%. Validation wilh the rest of the datábase (Tab!e 2) resulted ir* 80% of the samples corred!y classified. More ínisclassifications

TABEE-2. VALIDATION CLASS1FICATION MATRLX FOR APPLES (rOX AND GOLDEN) AS "MEAEY" OR "NONMEALY"

L'SING 15 TRS VARIABLES AND HALE Oh THE DATABASE. GRAYED ÁREAS DENOTE CORREC1'CLASSIHCATIONS. KOWS: qBSBRVED BY MEC'IIANICAL MBASUEBMHNTS.

COLÜMNS: PREDICTED BY TRS

Non mealy Mmity

1 lililí

Nunmuaty

26

29

Mealy

6 10 16

% Conect

82 77

'I A BLE 3. VALIDATIÜN CLASS1HCATIÓN MAI'BIX FOR AJ»PLE5 (CÜX AND GOLDEN) AS "MEALY/" "NONMEALY" OR

"mEStV GRAYEU ÁREAS DENOTE CORRECT CLASSirtCATIONS. KOWS: OHSEKVED BY MBCHANICA1

MEASCREMENTS. COLUMNS: PREDICTED BY TRS

Fresh Non mealy Mealy TOLII

Fresh

9 2 2

13

Nonmcaly

5 7 4

Mealy

3 6 1

16

% Correct

53 47 54 51

were proportionally obtained for mealy samples incorreclly predicted as Eresh (1^15 in thecalihraiion and 3/10 in Ihe validation, lack of sensiliviiy), Lhan I he opposiie case f0/29 and 6/26, Jack of speciíicity).

The 15 variables in the classification fiinclions were mostly flhsotptinn coefficients and some scattering coefficienls. al wavelengths around the ehJo-rnphyll absorption peak (670 nm) and a wide range of NIR wavelengths (818, 900. 930, 940. 960 and 980 nm}.

When irying to eslirnate three texture stages ("fresh," "nonmealy" and "mealy") wilh TRS. the performance of Ihe new models decreased consider-ably bota 98 to 717o. Moren ver, validation oniy achieved 51% (Table 3). The estimaltnn of four textura! slages ("fresh," "dry," "soff" and "mealy") achicved 71% in Ihe calibraüon model. bul only 47% in ihe validation (Table 4). In holh cases i t was noticed that ihe "mealy" category was beslpredicted than iheolher groups ("fresh," "dry." "soft" or the combination of ¡hem, Le., "nonmealy").

In order to reduce Ihe number of variables in the models to enhance robustness, new anaiyses were performed with more restrictive conditions in the number of wavelengths and the tolerance level of Üie stepwise melhod. i!

TABLE 4. VALIDACIÓN CLASSIFICAT10N MATRIX FOR APPLES (COX AND GOLDEK) AS "MEALY," "SOFT," "DRY" OR

"FRESil." GRAYED ÁREAS DENOTE CORRECT CLASSIFICATÍONS. ROWS: OBSERVAD BY MECUAMCAL

MbASUKEMENTS. COLUMNS: PREDICTBD BY I RS

Fresh Son Dry Mcaly Tolal

hresh

1 l 2 1

Ll

Sofl

1 1 0 1 3

Dry

5 3 3 1

12

MtaJj1

4 0 5

10 19

% Crirrccl

63 50 64 86 47

TABLE 5. VALIDATION CLASS1F1CAT10N MATRIX FOR APPLbS (COX AND COLDEN) AS "MEALY" OR "•NONMEALY"

USIKGHVETRS VARIABLES. GRAYED ÁREAS DENOTE CORRECT CLASSIEICATIONS ROWS: OllSERVED BY

MECIJAN1CAL MEASUREMENTS. COLUMNS: PREDICTED BYTRS

Nunrnealy Mealj Touil

NonmL'uly

22 3

25

Mealv

0 10 20

% Corrvcl

69 77 71

was seen Ihat all Ihe remaining variables in the modeis were absorption coefficients, and all the scattering ones were reinoved in the stepwise algo-rithm. The rernaining wavelenglhs were around Ihe 67Ü nm peak also coni-fiined with (he 960-980 nm N1R región. Overa!) segrcgafion ahilily in the estimafion of two textura! stages with five TRS variables and with tlnee TRS variables were, respect i vel y, 89% and 82%, with validalions achievíng 71% (Table5)and88%(Table6).

Sepárale modeis were also buill Cor each varieLy, to assess mealiness onsel ("nonmealy" versus "mealy"). In the case of Cox applos the classification score reached 100% whereas Ibr (Jolden apples the seore dropped lo 92%. Víilidatíoits showed 80% of well-elassified fruits forCox and 68% forGoJden (Table 7).

Table 8 shows ihe evalualion of modeis in terms of similarily index (**JC %). The sublraction oi" random agreement belween. procedures ieads to poorer scores with respect to the perecntage of well-classified individuáis. The hest

TABEES CLASSIFICATION PERFORMANCE OF THE MODELS. COWPARING THE IMUMBER OF CORRECTLY CLASSiFLED FRUITS (CC) WITH THE

SIMILARITY INDEX (K). AND ITS COMPONEMTS: tíi (ABSOLUTF PROPORTION OF CASES WHERE THE DESTRUCTIVE AND TRS CLASSIFICATIONS AGREE) AND ft (PROPORTION OF CASES WHERE THEY AGREE PURELY BY CHANCE)

1

2

3

4

5

6

7

Model

90 apples (Cox and Golden) as "namriealy'

"mealy" or uíicg 15 TRS variables

!J0 apples (Cox and Golden) as "nünmealy"

"mealy" or using five TRS variables

91) apples (Cox and Golden) as "ndiunealv"'

"mealy" or using Ihree TRS variables

50 Golden apples as "mealy" or L non me al y"

40 Cux apples using seven

90 apples (Ctu 'nonmcalv" variables

90 apples {C<» 'inft." "dry' \ariables

using íeven TRS variable*, áb ' mealy" or "nonmealy" TRS variables ; and Golden) M or "fresh'' using

Í and Golden) as ' or ''fresh'' using

''mealy." seven TRS

'mealy," «ven TRS

CC (calibration) (%)

98

¡49

H2

92

IDO

71

71

CC (validation) (#)

80

71

88

SS

80

51

+7

K(%)

54

39

73

33

54

27

25

^ - C C ( % )

-26

-32

-15

-35

-26

-24

-22

9]

0.8000

0.7111

0.KSÜ8

0.6800

0.8000

mttí

0.4666

&

0.5609

0.5234

0.5797

0,5200

0.5609

0.3303

0,2809

TABLE6. CLASS1F[CATIÓN MATRIX Í'OR APPLHS (COX AND

GOLDEN) AS "MEALY" OK "NONMEALY" UK1NGTMREE IJíS VAKIABLES. GKAYED AKfcAS DENOTE CORUtCT

CLASSIHCATIONS. RQWSí ÜBSERVEU BY MECUANICAL MEASUREMENTS. COLUMNS: PREDI CTED BY TRS

Nonmtíalj Mealy % Corren

Noamealy 2<J 3 ?í Mealy 1 11 es Total 31 14 88

TABLE7. CLASSIHCATION MATKICbS EOK GOLDEN APELES CIIFPER PART) A^Ü COX APPLES [LOWEK PART) AS

"MEALY" OR "NONMEALY" l'StNG SEVEN TRS VARIABLES C1KAYED ÁREAS DENOTECORRECT

CLASSIH1CATIONS. ROWS: OBSERVED. COLUMNS: PREDI Cl'ED

Gokleti EíflBmeaíy Mtulv Total

Cm Nmimt'aly Mealy Tiste)

Nonmualy

11 4

15

14 1

15

Mcalj

4 6

10

3 2 5

% Correct

73 60 68

82 67 80

result corresponds lo model 3 wilh a decrease of 15%. This model uses only ihree TRS parameters, which is a premise for less overfitting and better robust ness.

DISCUSSION

As it lias been observed ¡n other sludies (Barreiro et al. 1998; Andani el al. 1999), Ihc process for obtaining mealy sampies to carry oul research is noi afways a straighlfonvard rotiiine. The COA sampies slored urider sirici relalive humidity and lemperaiure conditions to promole inealincss develop-nienl, not always show al the end oí* me Irealmenl a mealy stage. In fací, a low perceniage of them were found mealy using the insirumenial desiructive lesl.

This may indícate thal there are more factOTS affeding mealiness than just the humidity and temperaiure during slorage, harvesi dale and variely. On lhe olher hand, ti seems clear that late harvested Golden apples can develop lextural disordcrs hy ihemselves and already "in ihe lree,'n without shelf life. This is of high importante for apple growers when, because oí climalic condilions or labor problems, it is impossible to pick. the whole produelion on lime, leaving part o i" il on the orchard,

The predictive models that estímate two instrumental mealiness slales C'nonmealy," "mealy") using absorption and scattering TRS coefíicients show hjgh discrimination performance when classilying samóles from both apple varieiies (80% of weH-classified frnits in rhe validation; Í5 variables nscd in the model). The robustness in the validation process should be enhanced, both using more samples and reducing ihe number oí variables.

Models eslimaling more than lwo textura! slales offer lower segregalion abilities. The prediction of three and four stales shows, in both cases, a score of 71% of well-classified frnits. with very poor validation residís (_i1% and 47% of well-classified frnits, respectively). The fact that the highest miscías-sificalion stores were found in the groups other than "mealy" suggcsts Ihat the use of the TRS technique itself is not adequate to delecl Lhe individual quality paramelers involved in the developmeni of mealiness (apparent droughl oí' lissues, üoflening). Itcan be also a problein oí lhe syslem set uporjusl a matter of detection resotutton.

The altempl to reduce the number of variables in lhe models was satis-fací ory. Performance in lhe detection of mealiness onset C'nonmealy,'' "mealy") was enhanced Trom 80 to 88%' using three variables, when lhe validation scores are compared. This factconfirms the mercase in rohuslness for lower number of variables and lhe risk thal is commonly assumed when using global models.

Given the differences in mealiness treaiments among lhe varielies, lhe resuils of the classification models applied to the pooled data from both of them, show that a real cause-effect may be detecíed by Üiis technique, related to Míe "loss of tightness in lhe tissues" and L1dryness" associated lo mealiness. The development of sepárate models Tor lhe varielies is a lógica! step ahead in the process, and rhe first attempte show different results for the two of them. In lhe case ofCox lhe classii ¡catión is lhe besUin tlie case of Golden the modei is unable lo obtain a belter score than lhe pooted data. These varietal models were created with a teduced number of Iruits, and at lhe same lime many variables are used (seven), which decreases the confidence on their robustness.

The similarity index (K ) is useful lo estáñate lhe robustness of lhe models in terms of rea! segregation capacity, as it eliminates the eíTecl of "agreement hy chance," In lhe models presenled in this work, K index punishes lhe performance of the models by nearly 20-30% in many cases. This indícales

that: (1) the destructive í.]uantification of instrumental mealiness is not mea-sunng exactly the same properly as TRS. because the agreement bcíween bullí methods is only 70%; and (2) the model may be unstable or nol repeatablc enough for industrial usage. The applicability of the prese ni models orí an industria] basis (i.e., a sensor on a pauking line) has to be improved, eilber wiib belter models or with a (jompiemenlary sensor logeiher with sensor fusión decisión algorithms.

CONCLUSIONS

Time-rcsolved reflectance spectroscopy has been proven to be a nseful technique to identify mealiness in apples nondestructively. Error rates in classincation models discriminating mcaly samples from nonmealy ones are low, when considering ihc iraditionai indicator of the scgregation ahility (pcr-contagc of correctly classifictl fhiits). The scgregation hetween more than two textora] stages of mealiness {other than "nonmealy" and "mealy") cannot be achíeved lo date. A TRS prolotype lo delect mealiness should be devcloped for on-line measurements. and sufficient data should be. gathered to tune the models. The technique. new in the field of food sensors, shows interesting poieniial for internal parameier detection of quaiity ailributes and disorders,

ACKNOWLEDGIVIENTS

The aulhors acknowledge Dr Veerle De Smedl, from the Katholieke LInjversiteil Leuven (Belgium), for the Cax apples, picked ¡ind lreated ai their íacilities. Also we Ihartk Carlos Gil, from Agro21 S.A. (La Almunia de D" Godina, Spain), for the Gol den apples and his expertise knowledge. Final i y, we acknowledge ihe EC for FATR CT96-1060 project funding and the Comu­nidad de Madrid for the PhD (FP1) grant to the first author.

REFERENCES

ABBOTT, J.A. 1999. Quaiity measuremení of fruits and vegetables. Poslhar-vest Biol. Technol. 15, 207-225.

AHRENS, M. and HUBER, D. 1987, A Methud jar Measurtng Mealiness in Tómalo Fruí! Vegetable Crops Department, Universily of Florida, Gainesvílle, FL326IÍ.

ANDANI, Z , DF SMEDT, V. and NICOLAÍ, B.M. 1999. Developmenl of mealiness in apples under shelf-iife condilions. Food Sci. Technol. Today 4(13), 203-204,

BARRE1R0, P., ORTÍZ, C, RUIZ ALTISENT, M., DE SMEDT, V., SCHOTTE, 5., BHANJf, Z., WAKELING, I. and BEYTS, P.K. 1998. Cnmparison bctwccn sensorial and instrumental mcasurcments íbr mea­liness assessmenl in apples. A Collaborative Test. J. Textura Siudies 29, 509-525.

BARREiRO, P-, RUÉ-CABELLO, L, FERNÁNDEZ-VALLE, M.E.,ORTIZ, C. and RUIZ-ALTISENT, M. 1999. Mealiness assessment in apples using MR1 techniques. Magn. Reson. Imag. 17, 275-281.

BECHAR, A„ MIZRACH, A., BARRH1RO, P. and LANDAHL, S. 2005. Determination of mealiness in apples using uHrasonic measurenienLs. Biusyst. Eng. 91, 329-334.

BIGNAMÍ. C , SCOSSA, A. and VAGNONl, G. 2003. Evahialion of oíd Italian apple cultivars by means oí* sensory analysis. Acta Hori. (ISHS) 598, 85-90.

CUBEDDU, R., MUSOL1NO, M., P1FFERI, A., TARONI, P. and VALEN-TINI. G. 1994a. Time-resolved reflectance: A systematic study fov appli-cation to ihe opüeat eharacteriziation oí' tlssues. IEEE. J. Quantum Elect. 30, 2421-3430.

CUBEDDU, R., MUSOLINO, M„ P1FFERL, A„ TARONI, P., VALENTÍN!, G. and CANTI, G. 1994b. Absorpliun specLmm of hematoporphirin derivative in vivo in a murine tumor model. Pnolochem, Pholobiol. 60, 582-585.

CUBEDDU, R., PIFFER1, A„ TARONI, P., TORR1CELLI, A., VALENTÍNl, G., DOVER, C, JOHNSON, D.. RUIZ-ALTISENT, M. and VALERO, C. 2001a. Non-destructive quami fication of cliernical and physical prop-erlies of fruils by lime-resolved refleelance spectroscopy in the wave-lenglh range 650-1000 nm. Appl. Opl. 40, 538-543.

CUBEDDU, R., PIFFER1, A., TARONI, P., TORRICELLI, A., VALENTÍN!, G., RUIZ ALTISENT, M., VALERO, C , ORTIZ, C, DOVER, C. and JOHNSON. D. 2001b. Time-resolved reflectance spectroscopy applied ro ihe non-desiructive monitoríng ofibe interna! optical properties in apples. Appl. Spectrosc. 55, 1368-1374.

DE SMEDT, V. 2000. Mcasurement and madelling of mealiness in appíes. PhD Dissertalion Thesis. Katholieke Universileil Leu ven, Leu ven, Beigium.

DE SMEDT. V., PAUWELS, E., DE BAFRDEMAEKER, J. and N1COLAÍ, B.M, 1998. Microscopic observaron of mealiness in apples: A quantita-tive approach. Poslharvest Biol. Technol. 14. 151-158.

F1SHER, D,V, 1943. Mealiness and quality of delicíous apples affected by growing conditions, maturily and storage techniques. Sci. Agrie. 23, 569-588.

GROSS, K.C. and SAMS, C E . 1984. Changes in cell wall neutral sugar composilion during í'ruk ripening; A species survey. PhyLoehemislry 23, 2457-2461.

HARKER, F.R. and HALLETT, 1,C. 1992. Physiologieal changes associated with development oí' mealiness oí apple during Storage, HortScience 27. 129Í-I294.

HARKER, F.R., STEC, M.G.H., HALLET, I.C. and BENETT, C L . 1997. Texture oí" pare ncb yin alous planl tissue; A compartson between tensile and uther instrumental and sensory measurenienls oí tissue slrength and juieiness. Poslharvest Biol, Technol. U, 63-72.

ORTIZ, C , BARRE1RÜ, P„ RUIZ-ALTISENT, M. and R1QLELME, F. 1999. An identiíieation procedure of woolly soft-flesh peaches (cv_ Maycrest) by instrumental assessment. J. Agrie. Eng. Res. 76, 355-362.

ORTIZ, C., BARRE1RO, P., CORREA, E , RUIZ-ALTISENT, M. and RTQUBLME, F. 200L Non-desrructive iden fifi catión of woolly peaches using mechanical impacl response and N1R spectroscopy. J. Agrie. Eng. Res. 73,281-289.

PAOLETTI, F., MONETA, E. and SIN ESTO, F. 1993. Mechanical properties and sensory evaluation of selected apple cutlivars. Lebensm.-Wiss. Technol. 26, 264-270.

SONEGO, L., BEN-ARIE, R„ RAYNAL, .1. and PECH, J.C. 1995. Biochemi-cal and physical evaluation of textura! characteristics of neclarines exhib-iiing woolly breakdown: [VIVIR imaging, X-ray computed tomography and peclin composition. Poslharvest Biol. Tecbnol. 5, 187-198.

STEINMETZ, V„ CROCHON, M., BELLONMAUREL, V.. GARCÍA FERNÁNDEZ, J.L., BARREiRO, P. and VERSTREKEN, L. 1996. Sensors for fruit firmness assessment: Comparison and fusión. J. Agrie. Eng. Res. 64. 15-18.

VALERO, C , RUIZ-ALTISENT, M„ CUBEDDU, R„ PIFFERI, A., TARONI, P., TORR1CELL1, A., VALENTÍN!, G., JOHNSON, D. and DOVER, C. 2004a. Selection models for Üie interaal quality of fruit, based on time domain láser refíectance speciroscopy. Biosyst. Eng. 88, 313-323.

VALERO, C , RUIZ-ALTISENT, M., CUBEDDU, R„ PIFFERI, A., TARONI, P., TORR1CELL1, A., VALENTÍN!, C , JOHNSON, D. and DOVER, C. 2004b. Deteclion of internal quality in kiwi with time-doniain diffusc reflectauce spectroscopy. Appl. Eng. Agrie. 20, 223-230.

VERLINDEN, BE. , NICOLA1, B.M. and DE BAERDEMAEKER, J. 1997. Modelling the relalion between macrn.scnp'tc vegetable lissue sirength and the sirength of cell walls and midále lamellae: A stachastic approarh. ASAE Paper 976025.

VON MOLLENDORF, L.J., JACOBS, C. and VILLIERS, OT. 1992. Post-harvesl Taclors involved in the development ofchilling injuries in peaches and neclarines. J. S. Afr. Soc. Hortic. Sci. 5, 58-66.

VON MOLLENDORF, L.J., VILLIERS, O.T., JACOBS, G. and WESTRAAD, I. 1993. Molecular characteristics oípectic consliluenls in relation to firmness, extraclable juice, and woolliness in nectarines. J. Ani. Soc. Hortic. Sci. 118, 77-80.