Hyperspectral Imaging

NIR hyperspectral imaging to predict beef quality, including sub-projects
Potential of Raman Spectroscopy
Consumer Assessment of ‘NIR Hyperspectral Imaging

Project number:                    72506

Lead contractor:                   Campden BRI, Agriculture Food and Biosciences Institute

Start & end date:                   01 February 2009 – 31 August 2009

Actual end date:                    June 2010

 

The Problem:

Despite careful process control of pH/temperature decline and correct implementation of electrical stimulation and other post slaughter processes (aitch bone hanging, aging times) there is a still a considerable range in beef tenderness.  In order to guarantee tenderness for specific markets it is necessary to have an additional means of selecting tender carcases.  Recent work in the USA described the potential for the on-line application of near infrared reflectance (NIR) to predict tenderness.  Quality grades for beef in the USA are based on marbling score and there is a much greater range of marbling in the USA than in the UK.  Similar NIR equipment has been evaluated in GB in an earlier trial funded by EBLEX to evaluate for prediction of tenderness in carcases with a lower range of marbling.  The results of this trial show that although NIR can predict tenderness as measured by shear force, further evaluation is required.  It was decided that hyperspectral imaging should be evaluated as it provides more information than standard NIR.

The opportunity was also taken to evaluate RAMAN spectroscopy.

 

Project Aims:

  1. To evaluate the potential of hyperspectral imaging for prediction of aspects of beef eating quality and nutritional quality.
  2. To evaluate the potential of Raman spectroscopy for prediction of aspects of beef eating quality
  3. To evaluate the relationship between Warner-Bratzler shear force (wbsf) consumer evaluation of beef longissimus dorsi.

 

Approach:

Hyperspectral imaging involves scanning the cut surface of meat.  This gives a large amount of information, and forms a compositional map of the subcutaneous fat, intermuscular fat, intramuscular fat (marbling) and connective tissue.  The technology allows the different components to be separated.  This means it may be possible to develop a rapid on-line method to predict fatty acid profiles which may benefit human health. Information on fat cover and eye muscle area/lean tissue area obtained from the hyperspectral image could be used to predict carcase yield and lean meat percentage.

Raman spectroscopy involves shining a laser of specific wavelength at meat and measuring the light scattered due to changes in the vibration of molecules.  Modern technological advances both in optic design and computing/software have lead to the development of both bench top and portable Raman spectrometers.  Raman spectroscopy has the advantage that, unlike near-infrared spectroscopy, it is not influenced by water content of the material and may be better suited to use on high water content foods such as meat.  This project will use the same carcases used in the evaluation of NIR hyperspectral imaging, allowing a comparison of the two techniques.

Samples from the same animals were subjected to Hyperspectral imaging, RAMAN Spectroscopy and laboratory assessments of quality as well as consumer evaluation.

 

Results:

NIR hyperspectral imaging showed that, by using an absorbance band at a specific range of wavelengths, the areas corresponding to lean and fat (marbling fat, intermuscular fat and subcutaneous fat) could be clearly identified. The distribution of fat and lean in an image of the forerib produced using spectral characteristics was almost identical to that seen in a colour image taken with a digital camera (Nikon D70).

Prediction models for meat quality were developed based on the NIR characteristics of the total longissimus dorsi (LD) muscle surface, including areas of marbling fat, and also on the LD lean regions alone, excluding the marbling fat. The spectral characteristics of the marbling fat were used to predict fatty acid profiles in the muscle. Prediction models were developed by using data sets of both unbloomed and bloomed samples independently, and combined spectra of both bloomed and unbloomed samples.

Prediction of Warner-Bratzler Shear Force (WBSF) using the total LD area (including marbling) was marginally better when the bloomed spectra were used (R2 = 0.61 WBSF at 14 day aging: R2 = 0.49 WBSF at 21 day aging) than when the unbloomed or combined spectra were used. Using the total LD spectra, prediction of eating quality attributes at 14 days aging ranged from R2 = 0.44 for overall liking (bloomed spectra) to R2 = 0.53 for juiciness (bloomed spectra). The coefficients of determination (R2) for prediction models developed using the spectra of the LD lean (excluding marbling fat) were generally of a similar order to the models developed using the spectra of the total LD.

The spectra of the total LD (including marbling) resulted in good prediction of intramuscular fat (IMF), with the best prediction model obtained using the unbloomed samples (R2 = 0.74). Using the total LD spectra the best prediction model for total saturated fatty acids (SFA) (R2 = 0.68) was obtained from the combined (unbloomed and bloomed) spectra model, whilst the best prediction of total monounsaturated fatty acids (MUFA) (R2 = 0.80) was obtained from unbloomed samples. Prediction of total polyunsaturated fatty acids (PUFA) and conjugated linoleic acid (CLA) using the total LD spectra was generally poorer than for SFA and MUFA.

 

The prediction models for most fatty acid groups such as SFA, MUFA and CLA obtained from the spectra of the marbling fat were improved by using the spectra with identified marbling fat with an area more than 50 pixels (R2 = 0.66 SFA, R2 = 0.58 MUFA and R2 = 0.65 CLA). However, this did not improve the R2 value in the prediction model for PUFA (R2 = 0.48) for which the best result was obtained from the unbloomed spectra (R2 = 0.53). The use of identified marbling fat with an area more than 50 pixels generally eliminated the factors such as noise, misclassified pixels and also the fat adhering to the meat surface.

Using Raman spectroscopy, prediction of WBSF values and consumer assessed eating quality was poor. The only statistically significant model was the prediction of juiciness, which only explained 9% of the variation in the data (R2 = 0.09). Raman spectroscopy showed greater potential for the prediction of intramuscular fat content (R2 = 0.23), marbling score (R2 = 0.41), MUFA (R2 = 0.26), SFA (R2 = 0.25) and CLA (R2 = 0.19).

The low prediction ability of the Raman spectra to predict quality aspects may be due to the spectrometer not operating optimally.

A statistically significant logistic regression model was developed which showed that consumers in the trial rated LD muscle with a WBSF value of 5.4 kg force or less as 70% acceptable.

 

Conclusion:

NIR hyperspectral imaging offers potential for improved prediction of shear force, fatty acid composition and sensory traits of beef.

Further discussions are underway with the researchers to establish the next steps in making this NIR technology suitable for online application.