rapid nondestructive determination of edible meat content in crabs (cancer pagurus) by near-infrared...

9
submitted papers Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy JENS PETTER WOLD,* MARTIN KERMIT, and ASTRID WOLL Nofima Mat AS, Osloveien 1, 1430 A ˚ s, Norway (J.P.W.); QVision AS, Drengsrudhagen 2, 1372 Asker, Norway (M.K.); and Møreforskning Marin, Gangstøvika, 6009 A ˚ lesund, Norway (A.W.) This article presents a method by which noncontact near-infrared (NIR) interactance imaging spectroscopy can be applied to determine the amount of edible meat in single live crabs (Cancer pagurus) on a conveyor belt at high speed. The physiology and optical properties of the crabs are presented and discussed in order to explain the requirements for representative spectroscopic sampling. Two different sampling and calibration strategies are discussed. One strategy is based on the extraction of one average NIR spectrum from certain locations in each crab. The other strategy relies on first making a model based on average spectra from a certain location, and then using this model for pixel-wise prediction of the meat content within the crabs. A measure of the predicted distribution of meat is then used for calibration. Reference measurements of meat content were based on an objective quantitative metric of the meat content. The results show that NIR imaging enables on- line grading of the crabs with a correlation of 0.96 with the measured meat content. Due to seasonal variations in the crabs, a piece-wise regression strategy performs slightly better than a global model. Pixel- wise predictions of meat content provide informative images showing the distribution and amount of meat within each crab. Index Headings: On-line; Near-infrared imaging; NIR imaging; Quality classification; Crabs. INTRODUCTION The brown crab (Cancer pagurus) (Fig. 1) is an important commercial species distributed along the west coast of Europe. It is fished throughout the year with a peak season from August to November. Total annual landings were 43 000 tons in 2006 (EUROSTAT, 2007) and the main fishing nations are the UK, Ireland, France, and Norway. The meat yield of individual crabs can vary from completely filled high-quality crabs to empty water-filled crabs. Only a portion of the catches, depending on season and geographical area, are of good quality. Crabs with high meat content have a higher value and can be sold live or cooked whole; they are also an excellent raw material for the processing industry. Medium quality crabs are generally limited to being used as raw material by the processing industry. During processing, the crabs are cut open before cooking, with the claws, body, and carapace (the dark brown shell) cooked separately due to differences in cooking time. A correct quality sorting of live crabs is essential to optimize these logistics. Today, crabs are quality sorted by manual inspection. Such grading has to go very fast, so the result of the procedure strongly depends on the skill and experience of the personnel. During the season, the water in the carapace is gradually replaced by tissue growth. Due to this, crabs with a low meat content may have a high water content, and this makes it difficult, even for skilled and experienced graders, to determine the meat content based on the crab’s exterior and weight. Several attempts have been made to find technical solutions for quality grading of live crabs. A technology based on visible light transmitted through the crab shell has been used in a Norwegian processing plant since the late 1990s. The transmitted light is registered by a sensitive camera, and a picture is shown on a monitor. 1 By evaluating the pictures, a skilled grader can screen up to 1000 crabs per hour. A similar technology is used on board some crab boats in the Swedish crab fishing industry. Refractometry is another method to gauge the quality of live crustaceans. 2 This is a simple nondestructive field technique for assessing the blood protein concentration. The method can reliably assess muscle mass, especially in lobsters; however, the method is not suited for total volume screening, since a small amount of liquid has to be extracted from each individual. To enable quantification of the interior parts of the crab, measurements must be taken through the carapace. It is well known that near-infrared (NIR) spectroscopy measured in the so-called interactance or transmission mode can be used to determine chemical properties of the interior of various products, for instance, estimation of sugar content in the fruit meat of melons and mandarin oranges in spite of the rather thick layer of peel. 3,4 Another example is fat determination of live salmon based on measurements through the dark skin. 5 Transcranial NIR measurements on humans to monitor oxygenation and blood flow in the brain have also been reported. 6 In many cases in which visible light cannot penetrate apparently nontransparent layers, NIR light in the region 850– 1050 nm can be used. NIR spectroscopy can be combined with imaging to obtain multispectral images. Multispectral imaging is a rapidly evolving methodology and is presently being used for both macro and micro imaging in combination with many different spectroscopic techniques (e.g., fluorescence, Fourier transform infrared (FT-IR), and Raman). On-line NIR imaging is successfully used for industrial waste sorting, 7 and, for instance, real-time detection of apple fruit firmness. 8 Recently, Received 8 September 2009; accepted 19 April 2010. * Author to whom correspondence should be sent. E-mail: jens.petter. wold@nofima.no. Volume 64, Number 7, 2010 APPLIED SPECTROSCOPY 691 0003-7028/10/6407-0691$2.00/0 Ó 2010 Society for Applied Spectroscopy

Upload: astrid

Post on 03-Oct-2016

221 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

submitted papers

Rapid Nondestructive Determination of Edible Meat Content inCrabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

JENS PETTER WOLD,* MARTIN KERMIT, and ASTRID WOLLNofima Mat AS, Osloveien 1, 1430 As, Norway (J.P.W.); QVision AS, Drengsrudhagen 2, 1372 Asker, Norway (M.K.); and Møreforskning Marin,Gangstøvika, 6009 Alesund, Norway (A.W.)

This article presents a method by which noncontact near-infrared (NIR)

interactance imaging spectroscopy can be applied to determine the

amount of edible meat in single live crabs (Cancer pagurus) on a conveyor

belt at high speed. The physiology and optical properties of the crabs are

presented and discussed in order to explain the requirements for

representative spectroscopic sampling. Two different sampling and

calibration strategies are discussed. One strategy is based on the

extraction of one average NIR spectrum from certain locations in each

crab. The other strategy relies on first making a model based on average

spectra from a certain location, and then using this model for pixel-wise

prediction of the meat content within the crabs. A measure of the

predicted distribution of meat is then used for calibration. Reference

measurements of meat content were based on an objective quantitative

metric of the meat content. The results show that NIR imaging enables on-

line grading of the crabs with a correlation of 0.96 with the measured

meat content. Due to seasonal variations in the crabs, a piece-wise

regression strategy performs slightly better than a global model. Pixel-

wise predictions of meat content provide informative images showing the

distribution and amount of meat within each crab.

Index Headings: On-line; Near-infrared imaging; NIR imaging; Quality

classification; Crabs.

INTRODUCTION

The brown crab (Cancer pagurus) (Fig. 1) is an importantcommercial species distributed along the west coast of Europe. Itis fished throughout the year with a peak season from August toNovember. Total annual landings were 43 000 tons in 2006(EUROSTAT, 2007) and the main fishing nations are the UK,Ireland, France, and Norway. The meat yield of individual crabscan vary from completely filled high-quality crabs to emptywater-filled crabs. Only a portion of the catches, depending onseason and geographical area, are of good quality. Crabs withhigh meat content have a higher value and can be sold live orcooked whole; they are also an excellent raw material for theprocessing industry. Medium quality crabs are generally limitedto being used as raw material by the processing industry. Duringprocessing, the crabs are cut open before cooking, with the claws,body, and carapace (the dark brown shell) cooked separately dueto differences in cooking time. A correct quality sorting of livecrabs is essential to optimize these logistics.

Today, crabs are quality sorted by manual inspection. Suchgrading has to go very fast, so the result of the procedurestrongly depends on the skill and experience of the personnel.During the season, the water in the carapace is graduallyreplaced by tissue growth. Due to this, crabs with a low meatcontent may have a high water content, and this makes itdifficult, even for skilled and experienced graders, to determinethe meat content based on the crab’s exterior and weight.

Several attempts have been made to find technical solutionsfor quality grading of live crabs. A technology based on visiblelight transmitted through the crab shell has been used in aNorwegian processing plant since the late 1990s. Thetransmitted light is registered by a sensitive camera, and apicture is shown on a monitor.1 By evaluating the pictures, askilled grader can screen up to 1000 crabs per hour. A similartechnology is used on board some crab boats in the Swedishcrab fishing industry.

Refractometry is another method to gauge the quality of livecrustaceans.2 This is a simple nondestructive field technique forassessing the blood protein concentration. The method canreliably assess muscle mass, especially in lobsters; however, themethod is not suited for total volume screening, since a smallamount of liquid has to be extracted from each individual.

To enable quantification of the interior parts of the crab,measurements must be taken through the carapace. It is wellknown that near-infrared (NIR) spectroscopy measured in theso-called interactance or transmission mode can be used todetermine chemical properties of the interior of variousproducts, for instance, estimation of sugar content in the fruitmeat of melons and mandarin oranges in spite of the ratherthick layer of peel.3,4 Another example is fat determination oflive salmon based on measurements through the dark skin.5

Transcranial NIR measurements on humans to monitoroxygenation and blood flow in the brain have also beenreported.6 In many cases in which visible light cannot penetrateapparently nontransparent layers, NIR light in the region 850–1050 nm can be used. NIR spectroscopy can be combined withimaging to obtain multispectral images. Multispectral imagingis a rapidly evolving methodology and is presently being usedfor both macro and micro imaging in combination with manydifferent spectroscopic techniques (e.g., fluorescence, Fouriertransform infrared (FT-IR), and Raman). On-line NIR imagingis successfully used for industrial waste sorting,7 and, forinstance, real-time detection of apple fruit firmness.8 Recently,

Received 8 September 2009; accepted 19 April 2010.* Author to whom correspondence should be sent. E-mail: [email protected].

Volume 64, Number 7, 2010 APPLIED SPECTROSCOPY 6910003-7028/10/6407-0691$2.00/0

� 2010 Society for Applied Spectroscopy

Page 2: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

rapid on-line and noncontact NIR multispectral imaging hasbeen reported to function well for the estimation of average fat,water, and protein in fish fillets, as well as the spatialdistribution of these crude components within the fillets.9,10

Distribution of ice fraction in super-chilled salmon fillets canbe imaged in the same way.11 In the three latter studies,interactance measurements were used in order to measuredeeper into the samples than just the surface.

In this paper we demonstrate how noncontact NIRinteractance imaging spectroscopy can be applied to determinethe amount of meat in single live crabs on a conveyor belt athigh speed. The physiology and optical properties of the crabsare presented and discussed in order to explain the require-ments for representative spectroscopic sampling. Two differentsampling and calibration strategies are discussed. One is basedon the extraction of average NIR spectra from different parts ofthe crab. The other relies on pixel-wise prediction of the meatcontent within the crab shell. Calibration strategies for optimalhandling of seasonal variations are also evaluated. Thereference measurement of meat content was a metric basedon weight of the meat and size of the crab.

MATERIALS AND METHODS

Crab Physiology of Relevance to Optical Properties. Theexoskeleton of the crab (the carapace) and the meat contentvary through the molt cycle. The white meat (muscle) andbrown meat, mainly consisting of liver (hepatopancreas),gradually increases during the longest lasting period in themolt cycle. Organic reserves are mainly stored in these organs.In the carapace, the liver dominates, symmetrically distributedin two lobes on each side in the crab shell with the cardiacstomach and the urine bladder located anterior between thelobes (Fig. 1). Muscles are found in the cage (thorax) where theappendages are attached. The gonads are distributed in twolobes, similar to the liver but with two additional funnelsleading backwards to the genital openings. The onset of femalegonad development (the roe) commences in summer, withspawning in late autumn and early winter.12 During this periodthe roe increases in both spread and thickness. Seen from thedorsal side of the crab, it will eventually cover the liver. In theposterior, on each side of the crab, there are two water-filledchambers where the gills are located.

The chemical composition of the liver, gonads, and musclesdiffers and probably varies through the molt cycle anddepending upon feed intake.13 In general, the lipid content ofthe muscle is low (0.2–0.9%) compared to that of the liver (9–16%) and the roe (9–10%), as reported for crabs of mediumquality caught in autumn in Norwegian waters.14,15 The proteincontent from the same tests varied from 21–23%, 13–16%, and22–27% for white claw meat, liver, and gonads, respectively,and dry matter from 24–26%, 33–36%, 40–42%, respectively.

Experimental Procedure. A total of 110 female crabs werecollected for calibration and testing during two experimentaltrials from the production line at Hitramat AS, Hitra, Norway.Sixty crabs were sampled in June 2009 and 50 in September2009 in order to include seasonal variation in the data set. Thecrabs within the two trials were selected with the aim ofobtaining a relevant span of the amount of edible meat in thecrab shell. The crabs also spanned a relevant range in weightand size. Female crabs were chosen because they normallycontain more meat compared to males, and because they alsocontain roe, which is a highly desired quality attribute.

The crabs were scanned live by the NIR system for thepurpose of calibration. Before scanning, the crabs wereelectrically stunned on-line—an established method also usedfor fish before processing—to avoid movement during scan-ning. All crabs were placed on a black conveyor belt (speed: 0.4m/s) and then passed under the NIR system with the legs downand carapace up. Each crab was scanned three times with thecrab facing backwards, and then a fourth time arbitrarilyoriented. Optimal presentation of the crabs to the scanner waswith the front half of the crab shell horizontal. However, sincethe crabs tended to move/walk, this was not always obtained.

Directly after scanning, the crabs were stunned, legs andthorax were removed from the carapace, and the amount of meatin the carapace was determined for each crab. In addition to thewhole crabs, extracted liver and roe from ten crabs weremeasured by the NIR instrument to compare with the crabspectra.

Near-Infrared Measurements. A commercial NIR imagingscanner (QMonitor, QVision AS, Oslo, Norway) was used toscan the crabs (Fig. 2). The scanner was installed in theproduction line at Hitramat AS, above the conveyor belt. Eachcrab was scanned when it passed under the scanner, anddepending on the size of the crab, the scanning time wasbetween about 0.5 and 1 s.

The instrument was based on interactance measurements inwhich the light was transmitted into the crab and then back-scattered to the surface. The system measured the lighttraversing the interior of the crab, but not the direct reflectedlight from the crab shell (Fig. 2). Optical sampling depth in thecrabs was not determined; however, for live salmon and driedsalted cod a typical sampling depth of about 10–15 mm wasobtained with a similar optical setup.5,10 The scanner wasplaced 12 cm above the conveyor belt and the illuminating fieldwas focused on to the belt along a line perpendicular to thedirection of movement. The field of spectral collection wasparallel to the illuminating field about 1 cm further down theconveyor. It was focused on the detector using a cylindricallens. To minimize collection of direct reflected light from thesample surface, a metal plate shield blocked the light betweenthe field of illumination and the field of detection. The distancebetween the crabs and the lower edge of the light shield variedwith the height of the crabs and was approximately 1–5 cm. Thelight source consisted of eight halogen lamps of 50 W each. Thescanner collected spectral images of 15 wavelengths/channelsbetween 760 and 1040 nm with a spectral resolution of 20 nm.The output was an image of the sample 60 pixels in the directionperpendicular to belt movement and around 200 pixels(depending on crab size) in the direction of belt movement.Each pixel represented a spatial area of about 10 mm 3 5 mmacross and along the conveyor direction, respectively. Furtherinstrument descriptions are given by Wold et al.10

Reference Measurements. Directly after scanning, thecrabs were stunned and the carapace was separated from thethorax (the cage). The cardiac stomach was removed and thecarapace with the remaining meat content was drained for 30minutes. The meat content, separated into liver and roe, wasdissected from each crab and placed in separate containers.Excess fluid was drained with blotting paper before weighingthe fractions. Several morphological parameters were measuredin order to correlate the weight to the size of the crab. Theseparameters included carapace width and length, live weight,body weight without legs and claws, weight of empty shell, and

692 Volume 64, Number 7, 2010

Page 3: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

weight of internal organs, including the gonad, liver, andepidermis.

The index for the meat content, hereafter called MQI(measured quality index), was found by dividing the weight ofthe liver and roe by the square of the carapace width (CW):

MQI ¼ 100 3ðraw weight of liver þ roeÞ=CW2 ð1Þ

The inner skin (epidermis) laying close to the carapace andsurrounding the liver was not included in the MQI even thoughthe industry uses this product in a paste called brown meat, amixture mainly consisting of the liver and the epidermis.

Sampling and Calibration Strategies. Multispectral imag-es give rich opportunities for optimized spectral sampling andin this paper two different strategies are presented anddiscussed. Both strategies were based on operation on thecrab shell only. The crab shell therefore had to be correctlydetected and distinguished from the conveyor belt, as well as

from the claws and legs. This segmentation was done in realtime based on spectral differences between shell and legs,claws and conveyor. The exact technical segmentation criterionused is not included here. Figure 3 illustrates how the crabswere detected and segmented.

The first strategy was based on extraction of one average

FIG. 2. (Left) Principal schematic of the on-line NIR noncontact interactance imaging system. (Right) Profile sketch of the position of the illumination and thedetection region on the crab, as well as proximate sampling volume. ‘‘IS’’ indicates imaging spectrograph.

FIG. 3. (A) Raw scanner image of crab at 840 nm. (B) Image A segmentedinto shell, claws and legs, and conveyor. (C) Image indicating the 50% frontpart of the shell used for extraction of the average NIR spectrum.

FIG. 1. Female brown crab as seen from the dorsal side. (A) Sketch of the internal anatomy as seen directly under the carapace and the inner skin (epidermic layer).(B) Female crab in the same position as the sketch. Right side of carapace removed and the inner skin turned down to show the anatomy.

APPLIED SPECTROSCOPY 693

Page 4: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

NIR spectrum from a certain region of each crab. Thisspectrum was then used as X in a multivariate regression modelto estimate y, the meat content (MQI). The other approach wasbased on a pixel-wise prediction of MQI within each crab, andthe distribution (histogram) of these predicted MQI values, orsimply the average predicted MQI, was used as X. The twostrategies are described in detail below.

Extraction of the Average Spectrum from a CertainLocation. Since the meat is located in the front part of thecrab shell, it was logical to extract spectral information fromthis region. When the crab shell was detected, a certain amountof the front part could be selected for spectral extraction. Toevaluate different sizes of this area, different percentages of thewhole shell were selected; 10, 20, 30, 40, 50, 60, 70, 80, 90,and 100%. Figure 3C shows the case in which 50% of the frontpart of the shell is selected.

From the selected region, the mean raw intensity spectrumwas extracted (raw). This was transformed to absorbance (abs)by log (raw�1), which was then used for calibration against thereference values (MQI). Normalized spectra were also evalu-ated for calibration, and the standard normal variate (SNV)15

was used as a pretreatment. For each spectrum, the mean of thespectrum was subtracted and the resulting spectrum was thendivided by the standard deviation (SD) of the spectrum. Theintention of using this transform is to remove spectral variationthat is not connected to chemical features, such as effects ofvarying distance between detector and sample, variation in lightscattering, etc. The measurement situation with the crabs wasrather unstable, and removal of unwanted spectral effects couldpotentially be important for proper modeling.

A possible drawback with this sampling strategy, from apractical point of view, was that the orientation of the crab had tobe defined in order to know where the front part was. In thisstudy, the orientation of the crabs was known, since they allentered the scanner facing backwards. In an industrial process, itis more convenient if the crabs do not have to be positioned onthe belt with a particular orientation. However, when the crabsare arbitrarily oriented, it is not trivial to automatically determinethe orientation of the crabs based on the spectral images.

Pixel-Wise Predictions. A calibration approach invariant tothe crab orientation would therefore be an advantage to avoidthe difficulties described in the previous strategy. One way todo this is to apply the following procedure:

(1) Make a calibration for MQI based on a location in the crabwhere the content of meat varied over the whole range andalso according to the overall MQI of the crab. This calibrationwas in this study based on spectral data obtained from the40% front part of the crab shell (as found by the approachdescribed above).

(2) This model is then applied on every pixel within the crabshell to obtain a MQI distribution image for each crab.

(3) A histogram of the predicted MQI values can then be madefor each crab/crab scan. The histogram values should bedivided by the number of pixels to normalize them withrespect to crab shell size.

(4) One histogram is obtained for each crab scan, and these wereused as X in a regression model to make a new predictionmodel for MQI. In addition to using a histogram per crab forcalibration, a simple average value of the predicted MQIvalues within each crab can also be applied.

Split Models. During the study we found that due to seasonal

and compositional variation in the crabs, it was relevant toconsider prediction models consisting of two separate calibra-tions. One approach was based on piece-wise regression.17 Twopartial least squares regression (PLSR) sub-models were made,one for crabs with low MQI values (11–30), and one for crabswith high MQI values (25–42). A global model covering thewhole range was used to determine the correct sub-model foreach spectrum. A simple criterion based on the global modelwas used to classify new samples; prediction values below orabove 27 resulted in the use of the low range or high range sub-models, respectively.

The other split model approach was based on creatingseparate models for roe and liver. MQI values were calculatedseparately for roe and liver (MQI roe and MQI liver) (Eq. 1),and these values were used as y for two separate models. Thesemodels were then used to predict the amount of roe and liver inthe test set crabs. Total MQI was calculated as the sum of MQIroe and MQI liver. These two split model approaches wereevaluated only for one dataset: the average absorption spectraextracted from the 40% front part of the carapace.

All image and data processing prior to multivariate calibrationwas performed with Matlab version 7.7.0.471 (R2008b) (TheMathworks Inc., Natick, MA).

Multivariate Calibration. The PLSR technique16 was usedto create calibration models for MQI based on extracted spectraldata from the NIR scanner or histograms from predictionimages. The collected data was divided in two: two-thirds of thesamples (74 crabs) were used for calibration, and one-third (36crabs) were used for the test set. Division into the two sets wasdone randomly. Cross-validation was used to determine theoptimal number of PLS factors in the different calibrationmodels. In the case of replicate scans, all scans from one crabwere left out in the same validation segment (leave-one-crab-out). The correlation coefficient (R) and the prediction errorexpressed as root mean square error of cross-validation(RMSECV) were used to describe the calibration models. Themodels were evaluated on the test sets. In the case for whichthree scans per crab were available, a predicted value wasobtained for each scan in order to study repeatability. Thecorrelation coefficient (R) and the prediction error expressed asroot mean square error of prediction (RMSEP) were used toevaluate the performance of the different models. Data analyseswere performed with the software The Unscrambler ver. 9.2(CAMO PROCESS AS, Oslo, Norway).

FIG. 4. Percentage of the crab meat made up of roe as a function of MQI andtime of year.

694 Volume 64, Number 7, 2010

Page 5: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

RESULTS AND DISCUSSION

Reference Measurements. The MQI varied from 11.4 to42.3, with a mean value of 24.9 and a standard deviation of 8.1,which is a rather broad range of values spanning a relevantquality range for mid and high season crabs. There was a cleartrend that crabs caught in June had on average less meat (meanMQI¼20.1) compared to those caught in September (mean MQI¼ 29.2). There was also a greater variation in MQI for theSeptember data (SD¼8.1) compared to the June data (SD¼4.6).As described above, the amount of roe generally increases duringthe season. Liver and roe together make up 100% of the meat.Figure 4 shows that the content of roe in June was rather low,making up 0–30% of the MQI, while in September the roesometimes constituted more than 60% of the meat for somecrabs. There was also a trend, in particular for September,indicating that the amount of roe increased with increasing MQI.With these seasonal variations it was reasonable to expectcorresponding changes in the NIR spectra. A NIR model basedon June crabs could not necessarily be expected to work well onSeptember crabs and vice versa. For instance, a June crab had anMQI of 28 and 1% roe, while a September crab had the same MQIvalue but with 43% roe. Because liver generally contains morefat than roe, the spectral signature from the two crabs would bequite different.

Finally it should be noted that the color of the liver variedsignificantly from light brown to dark brown. Some crabs alsolost much more water from the liver during draining thanothers, indicating large differences in water-holding capacity.Based on the reference values, Table I summarizes the mainstatistics for the calibration and the test sets.

Spectral Analysis. Typical NIR spectra from crabs withvarying meat content and from liver, roe, and water are shownin Fig. 5. Even though the spectral resolution was low, wecould distinguish the prominent spectral features typical forfoods in the covered wavelength region. The second overtoneof the OH stretching band (water) could be observed as a broadpeak around 980 nm.18 The third overtone of the CH stretchingband expected around 930 nm due to fat18 was clearly seen as ashoulder in the spectrum obtained from the liver. This shoulderwas not easy to discern in the crab spectra but contributed to abroader water peak. Empty crabs gave spectra with a morenarrow pure water peak. The large offset variation in the crabspectra was probably due to light scattering properties andoverall absorption in the crabs. Note that full crabs tended togive higher average absorption compared to empty crabs, and asimilar effect could be seen for liver compared to roe andwater.

A principal component analysis (PCA) of absorbance spectraobtained from the front 40% of the crab shells (calibration set)indicated that the spectra were composed of typical variationcomponents usually observed in NIR spectra from biologicaltissues such as meat and fish (Fig. 6). The first principal

component (88% of the spectral variation) was related to thevariation in spectral offset, which could be related to bothscattering properties and overall absorbance. The secondcomponent was probably related to some scattering effect(9%), while the third component described variation in thewater peak (about 2%). Component 4 reflected variation in fat

TABLE I. Statistics for calibration and test sets.

Data setNo. ofcrabs

No. ofscansa

MinMQI

MaxMQI

MeanMQI STD

Calibration 74 222 (74) 12.9 42.0 25.3 7.9Test 36 108 (36) 11.4 42.3 23.7 8.4

a No. of scans in parentheses indicates number of crabs scanned in an arbitraryorientation.

FIG. 5. (Top) NIR absorbance spectra from liver (–––), roe (– – –), and water(- - -). (Bottom) NIR absorbance spectra from full (- - -), medium full (–––),and scarce crabs (– – –).

FIG. 6. Spectral loadings from a principal component analysis of crab spectra.PC1 (––––), PC2 (– – –), PC3 (- - -), and PC4 (�).

APPLIED SPECTROSCOPY 695

Page 6: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

with a clear peak at about 930 nm (0.5%). It should be notedthat a PCA on the spectra from the extracted liver and roesamples resulted in very similar spectral loadings, whichindicated that we were probing inside the crabs and not only onthe surface of the crab shell.

Calibration Results. The calibration results for the strategyof extraction of one average NIR spectrum from each crab aresummarized in Table II. It is clear that the NIR data containedsufficient quantitative information to establish sound calibra-tions, and satisfactory predictability was obtained. Results forthe SNV normalized spectra are not shown, but those modelsrequired six PLS factors and gave systematically slightlyhigher prediction errors than for non-preprocessed spectra. Bestmodels for both types of spectra were obtained when 30–50%of the front shell was used for spectral extraction. This isreasonable, since the meat is located in this part of the shell andis seldom distributed further back. When a larger part of theshell was used, the prediction error increased systematically,with the poorest accuracy obtained for the whole shell. Theback half of the crab contains mainly water. Average spectrafrom the whole shell would then be more dominated by water,the differences caused by varying meat content would be less

distinct, and the models would be less accurate. Figure 9B(below) shows the predicted versus measured MQI values forthe test set using the model based on the 40% front part. Threemeasurements per crab are shown in the figure to give animpression of the repeatability. Mean standard deviation for theprediction values of replicates was 0.62, which is much lessthan the prediction error. The three scans per crab wereperformed within about one minute; however, since the crabswere alive the scans could look quite different due to crabmovement, slightly different position and orientation on thebelt, etc. The shown repeatability can therefore be regarded asrealistic.

The SNV transform involves two steps, first subtraction ofthe mean spectrum, and then division by the standard deviationof the spectrum. It should be noted that if only the first step ofthe transform was performed, results very similar to thoseobtained with non-preprocessed spectra were obtained. Divi-sion by the SD removed differences in spectral contrast, andthis reduced the information in the spectra. It can be mentionedthat multiple scattering correction (MSC) was also evaluatedfor preprocessing of the spectra, but the performance was notsuperior to SNV.

The PLS regression models were based on factors verysimilar to the principal component loadings shown in Fig. 6(for the model based on non-preprocessed spectra). Thisindicates that the main spectral variation in the crabs wasclosely related to the measured food content. To explain thevariation in MQI, factor 1 contributed 27%, factor 2contributed 19%, and factor 3 contributed 18%, while factor4, connected to fat, contributed 17% of the variation.

The results in Table II are useful only when the orientationof the crabs is known. For the crabs with arbitrary orientation,pixel-wise prediction of meat content was one strategy of usinginformation from the whole crab shell. Figure 7 shows scarce,medium, and full crabs with predicted meat distribution. Theimages indicate the amount of meat and how it was distributedin the crabs. In spite of low image resolution, a distributionsimilar to the physiology seen in Fig. 1 could be observed.Note that meat was also detected in the back part of the crabshells, especially for the full crabs. This corresponds to anactual occurrence of liver and roe, as can be seen in Fig. 1.

TABLE II. Overview of calibration results for MQI based on averageabsorption spectra from different percentages of the front part of the crabshell. (# lv) Number of latent variables used in the regression models, (R)correlation coefficient, (RMSECV) root mean square error of cross-validation, and (RMSEP) root mean square error of prediction.

% shell

Calibration Test

# lv R RMSECV R RMSEP

10 6 0.78 4.90 0.81 4.8220 7 0.89 3.63 0.89 3.8130 7 0.91 3.27 0.93 2.9840 7 0.91 3.20 0.94 2.8850 7 0.91 3.28 0.93 3.0260 7 0.90 3.47 0.89 3.9170 7 0.88 3.80 0.86 4.2280 7 0.87 3.85 0.86 4.2490 7 0.88 3.75 0.84 4.45

100 7 0.87 3.8 0.83 4.61

FIG. 7. Result of pixel-wise prediction of meat content in crabs. Top row: crabs with low meat content, middle row: medium full crabs, and bottom row: full crabs.

696 Volume 64, Number 7, 2010

Page 7: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

Note also that the meat was not always symmetricallydistributed, which is an argument for not basing measurementson a very limited area of the shell, risking collection of non-representative spectra.

Figure 8 shows MQI histograms for three of the crabs in Fig.7. The histograms were divided into 15 bins, and crabs withlittle meat had a high number of pixels with low predicted MQIvalues. Fuller crabs resulted in histogram distributions skewed

FIG. 8. Histograms of predicted MQI values at pixel level for low meat (gray bars), medium full (white bars), and full crab (black bars).

FIG. 9. Predicted versus measured MQI based on (A) average NIR spectrum extracted from front 40% of the crab shell (piece-wise regression), (B) average NIRspectrum extracted from front 40% of the crab shell (global regression model), (C) average pixel-wise prediction of MQI, and (D) average NIR spectrum extractedfrom 100% of the crab shell. Results in (A) and (B) are for crabs with known orientation, three scans per crab. Results in (C) and (D) are for arbitrarily oriented crabs,one scan per crab.

APPLIED SPECTROSCOPY 697

Page 8: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

toward higher MQI values. When these histograms were usedas the basis for calibration, we obtained rather simplemultivariate models (three components) (Table III) andprediction errors a little larger than those for the best modelbased on average spectra (Table II). Histograms were used withthe assumption that the distribution of MQI values carriedrelevant information. It turned out that using simply theaverage of the pixel predictions from each crab gave quitesimilar results to those obtained using the histograms. Actually,the correlation between the average of predicted MQI valuesper crab had a correlation of 0.87 with the measured MQI forthe calibration set. The single PLS component predictionmodel presented in Table III involves mainly an offset andslope correction of these data. Compared with the averagespectrum approach applied on the whole crab shell (100%)(Table III, Fig. 9D), the pixel-wise prediction modelsperformed considerably better (Fig. 9C).

As pointed out above, the average NIR spectrum from thewhole crab shell was not very representative for the measuredMQI (Tables II and III). The prediction model used for pixel-wise predictions, based on the 40% front part of the crabs(Table II), represented the variation in MQI much better. It wastherefore reasonable that this model would perform well on allparts of the shell, and the prediction images in Fig. 7 confirmedthis. The strategy of creating a calibration model based on alimited region in the multispectral image where there is a goodmatch between spectra and reference values, and then applyingthis model to the rest of the image, worked well. This is apractical approach that could be used in other types ofmultispectral imaging, also at the microscopic level. The

additional image information on the meat distribution might beused to extract more information about the quality of the crab,as well as being an important instrument for marine biologistswho study the life cycle of crabs.

In an industrial setting, the choice between an accuratemodel, which requires fixed orientation, and a less accuratemodel with no requirements for orientation, would be a tradeoffbetween accuracy and practical processing.

The result in Fig. 4 indicated that the seasonal variationscould affect our calibration models. The varying amount of roeand liver would most likely produce changes in the NIRspectra, which would not necessarily correspond to thevariation in MQI. Table IV shows the results for the twodifferent split models used in this article. A piece-wiseregression in which the calibration range was split into a lowand a high MQI range gave a better result than using a globalmodel. The two sub-models needed one and two componentsless than the global model, and the RMSECV for both thesemodels was lower than for the global model. The predictionerror for the test set was also notably lower for the piece-wisemodel compared to the global model (Fig. 9A). The lowernumber of components in the sub-models indicated that thesamples within each group were more equal. One could alsoconsider using separate models for different times of the year,but this would be less practical.

The methodology of making separate calibrations for roe andliver gave results comparable with the best models reportedabove, but not better. However, from a quality perspective, itwould be valuable to quantify the amount of roe in each crab.Crabs with high shares of roe are more valuable than thosecontaining mainly liver.

CONCLUSION

The results illustrate that NIR interactance imaging is highlysuitable for high-speed noncontact determination of edible foodcontent in live crabs. The described equipment has a capacityof typically 120 crabs per minute. The obtained accuracy issufficient for sorting the crabs into three to four quality classes,which is also the industrial need. The reference methoddenoted MQI is well defined, but it should be clarified whetherthis measure is representative of the actual end quality of thecrab, that is, the amount of meat after boiling of the crabs.Before the meat content was weighed, it was drained for 30minutes. It was noted that meat from some crabs lost a lot ofwater during draining, while meat from others had good water-holding capacity. This quality difference could also affect theresulting meat content after cooking. The NIR scanner

TABLE III. Calibration results for MQI for arbitrarily oriented crabs. Calibrations are based on three types of data: histograms of predicted MQIvalues from each crab (Histograms), the mean value of the predicted MQI values from each crab (MeanMQI), and average absorption spectrum from thewhole crab shell of each crab (Average100%). The calibration approaches are tested on crabs scanned with known orientation (scanned with crab facingbackwards) and arbitrarily oriented crabs. (# lv) Number of latent variables used in the regression models, (R) correlation coefficient, (RMSECV) rootmean square error of cross-validation, and (RMSEP) root mean square error of prediction.

Data

Known orientation (backward) Arbitrary orientation

Calibration Test Calibration Test

#lv R RMSECV R RMSEP #lv R RMSECV R RMSEP

Histograms 3 0.92 3.15 0.90 3.57 3 0.90 3.34 0.91 3.35MeanMQI 1 0.89 3.55 0.90 3.58 1 0.90 3.44 0.92 3.20Average100% 7 0.87 3.8 0.83 4.61 8 0.87 3.95 0.86 4.10

TABLE IV. Calibration results for MQI for two different split models.One is based on piece-wise regression in which the calibration range isdivided in two, low and high. The other is based on separate models forliver and roe, in which the predicted MQI is the sum of predicted values ofliver and roe. All models are based on average absorption spectra fromthe front 40% of the crab shell and performed on crabs with knownorientation. (# lv) Number of latent variables used in the regressionmodels, (R) correlation coefficient, (RMSECV) root mean square error ofcross-validation, and (RMSEP) root mean square error of prediction.

% shell

Calibration Test

# lv R RMSECV R RMSEP

Split low 6 0.78 2.320.96 2.58

Split high 5 0.87 2.52

Liver 6 0.72 2.450.93 3.05

Roe 6 0.76 3.22

698 Volume 64, Number 7, 2010

Page 9: Rapid Nondestructive Determination of Edible Meat Content in Crabs (Cancer pagurus) by Near-Infrared Imaging Spectroscopy

measures the distribution as well as the chemical compositionof the meat. This might also be a good basis for predicting themeat content after boiling, and the feasibility of this should beinvestigated.

The best models from this study have been industriallyvalidated, and the system is in daily use in a Norwegian crabprocessing facility. According to the automatic quality grading,the crabs are sorted to different processing lines, and the overallprocess yield has increased.

Since it is possible to determine meat content in crabs on aconveyor belt, it is also possible to measure such qualityfeatures onboard crab fishing boats. The fishermen can thenkeep only crabs of high quality while returning low qualitycrabs. This would facilitate not only high quality crabs to theconsumers and increased profit for the fisherman, but alsoenable a much more sustainable harvesting regime than what isthe case today. A boat system can be made to be much simplerand cheaper than the scanner presented in this paper. It can bebased on low-cost NIR instrumentation that measures ininteractance mode. The distribution of meat in the shell varies;however, two to four different measurement spots on the frontpart of the shell should be sufficient to obtain a useful estimateof meat content. The technique presented here might also besuitable for determination of muscle/meat content in othercrustaceans, such as lobster and king crab.

ACKNOWLEDGMENTS

This work was funded by the EU-project CrustaSea COLL-CT-2006-0340421 and The Fund for the Research Levy on Agricultural Products.We would like to thank Trond Edvardsen for the initial idea, Hitramat AS for

technical assistance and enthusiasm, and Jon Tschudi (Sintef ICT, Norway) andMartin Høy (Nofima Mat, Norway) for fruitful discussions.

1. Marintek, SINTEF group, report no.: MT40 F95-0311.402565.00.01(1995).

2. G. Ozbay and J. G. Riley, Aquacult. Res. 33, 557 (2002).3. C. V. Greensill and K. B. Walsh, Meas. Sci. Technol. 11, 1674 (2000).4. S. Kawano, T. Fujiwara, and M. Iwamoto, J. Jpn. Soc. Hort. Sci. 62, 465

(1993).5. A. Folkestad, J. P. Wold, K. A. Rørvik, J. Tschudi, K. H. Haugholt, K.

Kolstad, and T. Mørkøre, Aquaculture 280, 129 (2008).6. D. A. Benaron, C. D. Kurth, J. M. Steven, M. Delivoria-Papadopoulos, and

B. Chance, J. Clin. Monitor. Comput. 11, 109 (1995).7. G. El Masry and J. P. Wold, J. Agric. Food Chem. 56, 7672 (2008).8. A. Kulcke, C. Gurschler, G. Spock, R. Leitner, and M. Kraft, J. Near

Infrared Spectrosc. 11, 71 (2003).9. R. Lu and Y. Peng, Opt. Eng. 46, 123201 (2007).

10. J. P. Wold, I. R. Johansen, K. H. Haugholt, J. Tschudi, J. Thielemann, V.H. Segtnan, B. Narum, and E. Wold, J. Near Infrared Spectrosc. 14, 59(2006).

11. S. Ottestad, M. Høy, A. Stevik, and J. P. Wold, J. Near Infrared Spectrosc.17, 77 (2009).

12. H. Charniaux-Cotton and G. Payen, Endocrin. Select. Invertebrate Types25, 279 (1988).

13. G. M. Berge and A. K. Woll, Aquaculture 258, 496 (2006).14. S. O. Østvik, L. Grimsmo, M. Schei, and E. D. Halvorsen, SINTEF, Report

no. SFH 80A055068 (2006).15. R. J. Barnes, M. S. Dhanoa, and S. J. Lister, Appl. Spectrosc. 43, 5 (1989).16. H. Martens and T. Næs, Multivariate Calibration (John Wiley and Sons,

Chichester, UK, 1989).17. T. Næs, J. Chemom. 5, 487 (1991).18. L. Weyer and S. C. Lo, ‘‘Spectral structure correlations in the near

infrared’’, in Handbook of Vibrational Spectroscopy, J. M. Chalmers and P.R. Griffiths, Eds. (John Wiley and Sons, Chichester, UK, 2002), pp. 1817–1837.

APPLIED SPECTROSCOPY 699