Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

4.1 Introduction

Infrared spectra of food products can help to reveal information pertaining to molecular bonds present and hence provide details of their molecular structures. This ultimately can be related to various quality indices. Infrared spectroscopy is an ideal process analytical technology (PAT) tool that can rapidly, accurately and usually non-destructively assess the quality and functional properties of raw, in-process and final product materials. In addition to the need for efficiency, there is an emerging need in food processing for all major compositional and quality parameters to be determined, on-line and in real time. In addition to this, there is a need for food manufacturers to be able to demonstrate the authenticity of their products (Woodcock 2008) .

Spectroscopic techniques, other than infrared spectroscopy, have been investigated as potential PAT technologies in the food industry. These include Raman spectroscopy (Chap. 5), fluorescence spectroscopy (Chap. 12) and UV–Vis spectroscopy. UV–Vis has been employed to detect adulterated and authentic spirits (Contreras et al. 2010) ,discriminate between brands (Barbosa-García et al. 2007) , classify coffee (Souto et al. 2010) and quantify β-carotene (Biswas et al 2011). However, the focus of this chapter is infrared spectroscopy, and it will provide an overview of its theory, its instrumentation and its applicability as a PAT tool. Finally, it will review applications of infrared spectroscopy to food products .

4.2 Theory of Near- and Mid-infrared Spectroscopy

Infrared spectroscopy results from the interaction of infrared radiation and matter. The energy provided by the infrared radiations results in transitions between quantized vibrational energy states of molecules, i.e. resulting in molecular vibration . Atoms in a molecule can have a number of vibrational modes. Each mode (i) involves approximately harmonic displacements of the atoms from their equilibrium positions (Griffiths 2010) . When atoms vibrate as a simple harmonic oscillator, i.e. according to Hooke’s law (Eq. 4.1) where x is the displacement away from equilibrium, k is the proportionality (or force) constant and F is the force in newtons, the vibrational energy states (V iv ) can be described according to Eq. 4.2, where h is Planck’s constant, ν i is the fundamental frequency of the particular mode and v i is the vibrational quantum number of the ith mode (0, 1, 2, etc.):

$$F=kx$$
(4.1)
$${{V}_{iv}}=h{{v}_{i}}\left( {{v}_{i}}+\frac{1}{2} \right)$$
(4.2)

While the energy difference between v i  = 0 and v i  = 1 of most vibrational modes corresponds to the energy of radiation in the mid-infrared (MIR) range, overtone bands which relate to the transition between v i  = 0 and states higher than v i  = 1 are located in the near-infrared (NIR) region. Combination bands in the NIR region occur when there is a simultaneous promotion of two modes (Griffiths 2010).

A number of studies have assigned various food constituents (lipids, amides, moisture, sugars) to specific bands in MIR and NIR spectra. A selection of these regions and their associated mode of vibration of some food constituents are given in Tables 4.1 and 4.2. The characteristic broad peaks, resulting from overtone and combination bands, observed in the NIR spectra of a food product are shown in Fig. 4.1a; a corresponding MIR spectrum is shown in Fig. 4.1b. Such infrared spectra (Fig. 4.1) can contain a wealth of information on the molecular make-up of a food product. However, the spectral response of a molecular group can be influenced by neighbouring molecular groups (Reh 2001) . The complexity of food substances enhances these difficulties as the presence of various substances can result in peak shifts (Fagan and O’Donnell 2007) . Therefore, powerful statistical techniques, for example, principal component analysis (PCA) and partial least squares (PLS) regression, can be used for data compression and model development (Chap. 2).

Fig. 4.1
figure 1

Characteristic a NIR spectra and b MIR spectra of cheese

Full size image
Table 4.1 Selected molecular group absorption frequencies in the MIR region
Full size table
Table 4.2 Selected chemical assignments of absorption frequencies in the NIR region
Full size table

4.3 Instrumentation

There have been significant developments in the field of infrared instrumentation over the past decades . Initially, equipment focused on the use of monochromator or filter (fixed or tunable)-based systems. However, developments in instrumentation, such as Fourier transform infrared (FTIR) spectrometers and polychromators with InGaAs detectors , substantially improved the instrumentation performance, the range of applications and therefore the popularity of such equipment. The principles of such MIR and NIR instrumentation have been reviewed previously and will not be discussed further (Fagan and O’Donnell 2007; Griffiths 2010) . However, technical developments in infrared spectroscopy instrumentation will facilitate the transfer of this technology from laboratory to on-line application, thereby enhancing its potential as a PAT tool. Equipment manufacturers have moved from benchtop laboratory instruments (Fig. 4.2a) to the manufacturing of portable miniature-type spectrometers (Fig. 4.2b) to microspectrometers (Fig. 4.2c) . These have been driven in part by the requirement of end users who want to have the facility to bring the “spectrometer to the samples” rather than the “samples to the laboratory”. This has, for example, opened up opportunities for pre-harvest fruit and vegetable inspection. Such equipment may also include the added functionality of integrated global positioning system (GPS) measurements which are acquired simultaneously with the infrared spectra. Such facilities can allow for the “mapping” of produce quality in situ, thereby allowing the producer to make corrective decisions. In such applications, interference of the environment, such as ambient light and fluctuating temperatures, should be either minimized or accounted for by appropriate data processing (Nicolaï et al. 2007) . Another related emerging platform technology is hyperspectral imaging. It has the advantage of acquiring both spectral and spatial information of sample simultaneously . It has shown considerable potential in the pharmaceutical industry in terms of mapping active ingredients in tablets. Its potential as a PAT tool in the food industry is discussed in detail in Chap. 9.

Fig. 4.2
figure 2

A selection of a benchtop and b–d miniature, microspectrometer and portable spectrometers

Full size image

Another significant consideration of the end user must be the development and maintenance of calibration equations. Many portable spectrometers rely on the end user to provide and maintain the calibration equations required. However, this can require a substantial investment in time, labour and cost. Companies are emerging, however, which offer transferable NIR calibration solutions. Such companies have developed calibration models over many years using benchtop spectrometers, and they license them out for transfer to portable spectrometers . A service contract can also be entered into whereby the company maintains, updates and ensures set accuracy levels for the calibration model over time .

Continued research into the development of robust, fast miniature and microspectrometers will facilitate the continued adoption of this technology as a PAT tool in the food industry.

4.4 Infrared Spectroscopy as a PAT Technology

Infrared spectroscopy has been widely investigated as a rapid non-destructive assessment tool for food products . Fruit , vegetable, dairy and meat products have been the most widely investigated. However, the majority of these studies have been laboratory based. The greatest advantage in the use of infrared-based technology as PAT tools will be their implementation in the form of on-line/at-line process analysers, which take advantage of rapid analysis times and the minimal sample presentation required. However, it should be noted that the requirements for laboratory-based analysis will differ in comparison with on-line technology. Infrared spectroscopy also has the capacity to predict numerous indices of a material simultaneously. In order to realize the potential of such data-rich tools in the food industry, appropriate data analysis (Chap. 2) and data management strategies (Chap. 3) are required. Food quality, however, cannot be considered as a single, well-defined attribute. In fact, it encompasses a number of properties or characteristics, which are often referred to as quality indices, of the product under test (Abbott 1999) . While infrared spectroscopy can offer a solution to this challenge, one must ensure that the basis for the prediction of quality is fully understood, as well as its inherent limitations.

4.5 Applications

4.5.1 Dairy

The dairy industry has seen significant advances towards automation of production processes. For example, the move to closed commercial cheese vats versus the traditional open cheese vat drove the desire for on-line milk coagulation monitoring systems .

4.5.1.1 Raw Material

Milk composition and quality can vary depending on a number of factors, including animal genetics, health, and (in some countries) season. Such variability could significantly impact the quality of the final product. For example, milk fat to protein ratio will significantly affect a number of processing steps (coagulation, syneresis) which ultimately affects cheese quality and quantity. Therefore, it is usual that processors would standardise milk fat and protein content prior to use. Therefore, the use of infrared spectroscopy to facilitate the production of high quality of milk has been investigated in applications ranging from monitoring rumen metabolism through to standardisation of milk in the milk processing plant (Fagan et al. 2009b) .

Off-line rapid analysis of milk composition using the FTIR measuring principle has been successfully commercialized with products such as the MilkoScan™ FT 120 (Foss Analytical, Denmark). It utilizes FTIR technology to analyse up to 600 samples/h and can be used for routine analysis, such as fat, protein, lactose, total solids and solids-non-fat, density, freezing point depression, urea and casein analysis, in compliance with International Dairy Federation (IDF) and Association of Analytical Communities (AOAC) standards.

The further development of on-line determination of milk composition and quality would be advantageous as such knowledge is essential for the efficient management of dairy herds. Brandt et al. (2010) , however, stated that while a number of sensors are available or in development which can be used for management support in improving mastitis detection, monitoring fertility and reproduction and adapting individual diets, there is still a requirement to adapt these sensors to the particular requirements of on-farm utilization such as robustness , calibration and maintenance , costs, operating cycle duration, and high sensitivity and specificity.

Tsenkova et al. (2001) examined the potential of predicting somatic cell count (SCC) of milk using NIR transflectance spectra obtained using a benchtop spectrophotometer. They stated that the results indicated that NIR spectroscopy would be a suitable screening tool in such an application as the differentiation between healthy and mastitic milk samples was possible. More recently, an NIR spectroscopic sensing system for on-line monitoring of milk quality during milking has been developed (Kawamura et al. 2007) . The system was installed between a teatcup cluster and a milk bucket of a milking machine. The authors developed models for the prediction of fat, protein, lactose, SCC and milk urea nitrogen (MUN) during milking with sufficient precision and accuracy (R 2=0.82-0.95), although only four cows were monitored over time. Following this study, the sensing system was installed in an automatic milking system. The system recorded diffusion transmittance spectra (600-1050nm) with a 1-nm interval every 10 s during milking. Seventeen cows were used in this study. The models developed for fat, protein, lactose, SCC and MUN had R 2 values of 0.95, 0.83, 0.72, 0.68 and 0.53, respectively. The authors used the SCC calibration model to discriminate between healthy cow samples and other cow samples. The resulting classification gave a probability, for classifying correctly, of 82 %. In both studies, the samples were divided into calibration (2/3) and validation (1/3) sets. Further validation of the models is therefore recommended in conjunction with testing on a wider range of animals.

MIR spectroscopy has also been explored for the offline determination of milk traits (Cecchinato et al. 2009; Dal Zotto et al. 2008; De Marchi et al. 2009) . Milk coagulation properties (MCP) will vary depending on a number of factors, including heritable parameters (Cassandro et al. 2008) . However, if this information is to be fully exploited, there would be a requirement for a rapid method of determining MCP in milk-recording systems. Dal Zotto et al. (2008) found that MIR spectroscopy could predict the rennet coagulation time (RCT) of milk samples albeit with an R 2 of 0.73, which suggested approximate quantitative predictions were possible. De Marchi et al. (2009) carried out a further examination of this approach. Using a dataset of over a thousand samples, RCT was predicted with an R of 0.79. In both studies, the range error ratio (RER) was similar: 9.2 and 10.6. Cecchinato et al. (2009) investigated the variation of MCP predictions obtained by MIR spectroscopy, as well as estimating the expected response from a breeding program focusing on the enhancement of MCP using MIR predictions as indicator traits. They found that estimated genetic correlations between measure and predictions of RCT were very high.

4.5.1.2 Process Monitoring

NIR technology has been successfully applied at laboratory and commercial scales for monitoring processes during cheese manufacture. In particular, the milk coagulation process during cheese production has received a great deal of attention, and cutting the coagulum either before or after the optimum point results in losses of curd and fat. An increase in cheese moisture also occurs if the gel is too firm when cut. Originally, the determination of the cutting time was established by the cheese maker. Although accurate this method is not feasible in closed commercial vats and, together with an increased desire for automation in the cheese industry, has led to the need for an on-line objective method for the monitoring of milk coagulation . Instruments have been developed based on several technologies to this end. Ideally, a sensor to monitor milk coagulation could be installed on-line to allow for automation of the production process, without causing damage to the forming curd, and NIR sensors meet these requirements. Early methods, which utilized the changes in the optical properties of the milk , were reflection photometry (Hardy and Fanni 1981) and absorbance (McMahon et al. 1984) . Although the reflection photometry and absorbance methods were found to monitor coagulation, they found little usage. However, developments in fibre optics have overcome many of the problems associated with these techniques. Light in the NIR spectral region can be transmitted through a fibre optic bundle and diffuse reflectance or transmission monitor. As the gel is formed, reflectance will increase while transmission will decrease. Payne et al. (1993) developed a method based on changes in diffuse reflectance during milk coagulation. Reflectance was measured using a fibre optic probe, utilizing a photodiode light source at a wavelength of 940 nm. The time to the inflection point (t max) was determined from the first derivative and was found to correlate well with Formograph cutting times. Linear prediction equations, which were considered to be of the form required for predicting cutting time, were also developed using t max. This technology has been commercialized as the CoAguLite sensor (Reflectronics Inc, Lexington, KY) (Fig. 4.3a). This technology could also be used in conjunction with other sensors; for example, the FiberView dairy waste sensor system (Fig. 4.3b) could be used to monitor waste streams in dairy facilities. This enables the location, occurrence or concentration of the discharge to be determined. It monitors solids concentration in dairy plant effluents in the range of 0–1 % solids (or higher), and due to its quick response to loss events, it allows operators to take corrective actions.

Fig. 4.3
figure 3

a The CoAguLite sensor for predicting the optimal cutting time. b The FiberView Dairy Waste Sensor System (Reflectronics Inc, KY)

Full size image

Syneresis is a critical phase in cheese manufacture, with the rate and extent of syneresis playing a fundamental role in determining the moisture, mineral and lactose content of drained curd and hence that of the final cheese (Lawrence and Gilles 1980; Pearse and Mackinlay 1989) . Therefore, research is ongoing into the development of a syneresis control technology. A number of potentially non-invasive technologies have been investigated for such an application, including ultrasound and computer vision (Everard et al. 2007; Fagan et al. 2008a; Taifi et al. 2006; Tellier et al. 1993) and NIR sensing (Castillo et al. 2005a; Fagan et al. 2009a; Fagan et al. 2007a) . Initial studies focused on offline optical sensing of whey samples (Castillo et al. 2005b). An adaption of this technology led to the development of a sensor which could be installed in the wall of a cheese vat for on-line continuous monitoring of both coagulation and syneresis (Fagan et al. 2007a) . The sensor operated at 980 nm and was sensitive to casein micelle aggregation and curd firming during coagulation and to changes in curd moisture and whey fat contents during syneresis . This sensor was also used to predict whey fat content (i.e. fat losses), curd yield and curd moisture content with standard error predictions (SEPs) of 2.37 g, 0.91 and 1.28 %, respectively (Fagan et al. 2008b). Further work used a wider spectral range (300–1100 nm) in conjunction with PLS regression to predict whey fat and curd moisture with root mean square error of cross-validation (RMSECV) values of 0.094 and 4.066 %, respectively (Fagan et al. 2009a). Mateo et al. (2009) developed another set of models which predicted the yield of whey (R 2 = 0.83, error = 6.13 g/100 g) using three terms, namely light backscatter, milk fat content and cutting intensity. These studies were carried out in laboratory-scale cheese vats (7–11 L). Therefore, further scaling up and development under commercial conditions of the technology would be required if it is to become viable at a commercial scale.

NIR spectroscopy has also been investigated as a process control tool in yogurt production. Cimander et al. (2002) studied the potential of NIR spectroscopy to monitor yogurt fermentation in a 4.2-L laboratory-scale vat. A sensor signal fusion approach was adopted with NIR (400–2500 nm), electronic nose , and standard bioreactor sensors installed as part of a multi-analyzer setup (Fig. 4.4a). While the electronic nose followed changes in galactose, lactic acid, lactose and pH, the NIR sensor signal correlated well with the changes in the physical properties during fermentation. Therefore, the signals from the sensors were fused using a cascade artificial neural network (ANN) as detailed in Fig. 4.4b. Results suggested that the accuracy of the neural network prediction was acceptable. This approach was further investigated by Navrátil et al. (2004) under industrial conditions in a 1000-L vat. Signal responses from NIR and electronic nose sensors were subjected to PCA separately. The scores of the first principal component from each PCA were then used to make a trajectory plot for each fermentation batch. PLS regression of the NIR spectra was also used to predict pH and titratable acidity (expressed as Thorner degrees, °Th) during fermentation with reasonable success (SEPs of 0.17 and 6.6 °Th, respectively). MIR spectroscopy has also been employed to monitor the sorghum fermentation process (Correia et al. 2005) . They used FTIR spectroscopy to detect differences due to the effect of lactic bacteria on sorghum fermentation. They found it was possible to differentiate between samples which used natural yogurt and Lactobacillus fermentum as inocula due to variations in protein and starch structure.

Fig. 4.4
figure 4

The multi-analyzer setup a applied during yogurt fermentations: C1, compensator bottle 1, to trap condensating vapour and to compensate for minor flow rate variations; C2, compensator bottle 2 and b the neural network topology used for sensor fusion. The primary network received six input signals from the electronic nose and was cascaded by the secondary network, which received seven input signals: the output signals from the primary network for pH and lactic acid, four second-derivative NIRS signals (1402–1408 nm) and the first derivative of the reactor temperature signal. A logic gate made the final decision for the state variable. (Cimander et al. 2002)

Full size image

4.5.1.3 Final Product Quality and Authenticity

It is stated in Chap. 12 that “Quality attributes for dairy products can be both the chemical composition of a given product like protein, moisture and fat content, and the sensory quality attributes like taste, smell and consistency.” Therefore, the integration of sensing technologies which provide information on such attributes is critical. Table 4.3 summarizes a number of studies which have examined the potential of NIR and MIR spectroscopies to predict the composition of dairy products. These have primarily been laboratory based. Offline laboratory-based infrared sensing systems which provide rapid compositional analysis of dairy products are available. These systems should conform to relevant standards such as ISO Standard 21543:2006 (ISO 2006). A number of studies have also investigated the prediction of sensory quality attributes . Downey et al. (2005) predicted the maturity and sensory attributes of Cheddar cheese using NIR spectroscopy. Generally, second derivative spectra in the region of 750–1098 nm produced the most accurate models with age predicted with an RMSECV of 0.61 months, while the most successfully predicted sensory texture attributes were rubbery, chewy, mouthcoating and massforming with RER values of 8.8, 6.3, 7.6 and 8.5, respectively. NIR spectroscopy was also employed to predict both the sensory and instrumental attributes of processed cheese using NIR spectroscopy (Blazquez et al. 2006) . In general, they found that the models developed for predicting sensory texture in processed cheese were stronger than those for Cheddar cheese, with rubbery, chewy, mouthcoating and massforming predicted with RER values of 9.1, 12.0, 8.1 and 8.1, respectively. Fagan et al. (2007b) compared the NIR models developed by Blazquez et al. (2006) to models developed using MIR spectroscopy, which also predict sensory texture parameters of processed cheese . NIR spectroscopy was better at predicting creamy, chewy, and melting, with the R 2 values of the NIR models indicating excellent predictions as opposed to the good predictions of the MIR models. The RER values for the NIR reflectance models indicated a high utility value, whereas the RER values obtained by Fagan et al. (2007b) had a good practical utility. However, the MIR-derived fragmentable model had better accuracy than the NIR model, with excellent and good predictions, respectively.

Table 4.3 Application of near- and mid-infrared spectroscopy in cheese and yogurt composition analysis. (Modified from Woodcock (2008)
Full size table

The requirement to demonstrate the authenticity and safety of dairy products has also led to research into the use of infrared technology for such applications. Determination of the geographic origin and manufacturing conditions of cheese has received a great deal of attention (Boubellouta et al. 2010; Cattaneo et al. 2008; Karoui et al. 2004, 2005a, b, 2007a, 2008; Kocaoglu-Vurma et al. 2009; Pillonel et al. 2003). For example, Pillonel et al. (2003) studied the potential of MIR and NIR spectroscopies to discriminate between Emmental cheeses (n = 20) based on geographic origin. Samples were obtained from six regions, and they found that MIR transmission spectra could be used to discriminate (i.e. 100 % correct classification) Swiss cheese from the other regions, while NIR spectra classified the samples by the six regions of origin. Karoui et al. (2007a) also examined MIR spectroscopy to determine the authentication of 25 Gruyère “protected designation of origin” (PDO) and L’Etivaz PDO cheeses. They found that the spectral regions 3000–2800 cm−1 and 1500–900 cm−1 were most useful with 90.5 and 90.9 % correct classification results achieved, respectively. MIR spectroscopy (supplemented by partial 16S rDNA sequencing) has also been employed to monitor the population dynamics of microorganisms during cheese ripening (Oberreuter et al. 2003) .

4.5.2 Cereal Grains and Seeds

NIR spectroscopy has been widely used in routine quality control analysis in the grain industry since the 1960s (Scotter 1990) . This has included the assessment of moisture and protein content (Downey and Byrne 1987; Norris and Williams 1979; Williams 1979; Williams and Cordeiro 1979, 1981) . More recently, developments in this area have focused on assessment of grain quality at harvest, grain quality classification and sorting and grain blending.

Kawamura et al. (2003) developed an automated rice quality inspection system which utilized both visible and NIR technology. The objective was to develop a system which measured not only moisture content but also other rice quality indices in order to grade rough rice according to quality when it arrives at the drying facility. The system they developed consisted of a rice huller, a rice cleaner, an NIR instrument and a Vis segregator. This system enabled rough rice transported to a rice-drying facility to be classified into six qualitative grades.

Grain quality at harvesting is also a critical parameter as there can be significant within-field variability of grain quality parameters, for example, protein and moisture content. Maertens et al. (2004) described some of the requirement for online grain quality assessment at harvest. They included the use of a robust NIR spectrometer, design of a measurement configuration that guarantees a constant grain sample presentation while also avoiding dirt and blockages, that the sensor should be calibrated on the harvester and not under simulated conditions in the laboratory and finally that appropriate signal processing techniques should be employed to filter the spectral data, both in the time and wavelength domain They also studied the potential of an NIR sensor mounted on the bypass of the grain elevator of a combine harvester for online prediction of wheat moisture and protein content. They found that the average prediction errors were 0.56 and 0.31 % for protein and moisture content, respectively, where moisture content was below 18 %.

Detection and removal of internal insects and fungal contamination from seeds (grains , beans and nuts) are important control measures for ensuring storage longevity, seed quality and food safety (Pasikatan and Dowell 2001) . NIR spectroscopy has been applied to the detection of infestation of such products. NIR spectroscopy has been used to differentiate among individual wheat kernels that are uninfested, those infested with weevil larvae or pupae, or those that contain a parasitoid pupa (Baker et al. 1999) . Wang et al. (2002) recorded single-seed NIR spectra of a total of 1600 soya bean seeds, i.e. 700 sound seeds and 900 seeds damaged by weather, frost, sprout, heat or mould. The regions 750–1690 nm and 450–1690 nm gave the best classification of seeds into “sound” and “damaged” categories. They also found that an optimally developed neural network (parameters: momentum = 0.6, learning rate = 0.7, learning cycles = 150,000, wavelength region = 490–1690 nm) could classify seed according to six categories, i.e. “sound” (100 %) and five damage categories, “weather” (98 %), “frost” (97 %), “sprout” (64 %), “heat” (79 %) and “mold” (83 %), with reasonable success.

Aflatoxin B1 is recognized by the International Agency of Research on Cancer as a group 1 carcinogen for animals and humans, and Fernández-Ibañez et al. (2009) investigated the potential of Fourier transform NIR spectroscopy to detect aflatoxin B1 in cereal grains. They analysed maize and barley samples (n = 152) and developed models (R 2 = 0.82–0.85) for prediction of the presence of aflatoxin B1, which suggested that NIR spectroscopy could be a suitable alternative for fast detection of aflatoxin B1 in cereals .

4.5.3  Fruit and Vegetables

The application of NIR and MIR to quality assessment of fruit and vegetables has been widely studied (Table 4.4) . In terms of infrared spectroscopy’s role as a PAT tool in this industry, it could be employed for the optimization of harvesting, defect identification, disease control, process control applications and overall quality classification .

Table 4.4 Examples of applications of near- and mid-infrared spectroscopy to quality assessment of fruit
Full size table

4.5.3.1 Harvest Optimization

Prediction of the optimal harvest time of apples will minimize the occurrence of quality losses . Peirs et al. (2001) predicted the optimal harvest date of apples harvested no more than 8 weeks before the commercial picking date using Vis–NIR spectra collected post harvest in the laboratory (measurements were carried out on the same day or the day after picking). They stated that it was possible to measure apple maturity for harvest of individual cultivars within an orchard and that the number of days before the optimum harvest date was well predicted (R = 0.90–0.93). Further work examined the potential of Vis–NIR spectroscopy to estimated apple pre- and post-storage quality indices at harvest (McGlone et al. 2002) . The apples were harvested 1–3 weeks before and up to 1 week after the commercial harvest period. Spectral analysis in this case took place between 16 and 24 h after harvest. The authors found that although models were developed to predict quality indices of the apples they were still very poor in terms of prediction accuracies. Therefore, they were unlikely to be useful for sorting or grading due to the high rate of prediction errors that would result. They also stated that the prediction models, with the exception of soluble solids content, may be almost solely dependent on changes in the apple chlorophyll level and not have any direct sensitivity to the constituents or properties of interest .

Clark et al. (2004) examined the potential of Vis–NIR spectroscopy to predict the storage potential of kiwifruit. They employed canonical discriminant analysis (CDA) to optimize the separation between the two categories, i.e. “sound fruit” and “fruit developing a disorder during storage”. They estimated that the overall incidence of disorders could have been reduced from 33.9 to 17.9 % and 14.7 to 8.5 % depending on the harvest or when using all harvests from 13.7 to 6.8 %.

A similar approach has also been investigated for mango (Saranwong et al. 2004) . Vis–NIR spectra of mango were collected on the day of harvest and models were developed to predict harvest and eating quality using multiple linear regression and PLS regression. They stated that the calibration equations developed were sufficiently accurate to determine the harvest quality, dry matter and starch content of hard green mango fruit non-destructively. Using this information, the soluble solids content of the ripe fruit, which is an eating quality index, could be precisely predicted at the time of harvest .

It has also been demonstrated that infrared technology can be used for fruit assessment prior to harvesting. Pérez-Marín et al. (2009) used a handheld micro-electro-mechanical system (MEMS) spectrometer and a diode-array Vis–NIR ­spectrophotometer to collect the spectra of nectarine during on-tree ripening (n = 144). They developed models to quantify changes in soluble solids content, flesh firmness, fruit weight and diameter. Both instruments provided good precision for soluble solids content (R 2 = 0.89; SEP = 0.75–0.81 %) and for firmness (R 2 = 0.84–0.86; SEP = 11.6–12.7 N). The diode-array instrument predicted the two other physical parameters well (R 2 = 0.98 and SEP = 5.40 g for fruit weight and R 2 = 0.75 and SEP = 0.46 cm for diameter), while the handheld MEMS instrument proved less accurate in this respect (Pérez-Marín et al. 2009) .

A portable non-invasive instrument based on NIR spectroscopy has also been developed to measure the ripeness of wine grapes (Larrain et al. 2008) . It was used to predict three ripeness variables with excellent success for Brix and pH (R 2 = 0.87–0.93) and with less accuracy (R 2 = 0.56–0.80) for pH .

4.5.3.2  Defect Identification

Burks et al. (2000) applied NIR spectroscopy to the sorting and classification of figs. They classified the figs according to the number of categories (“passable”, “infested”, “rotten”, “sour”, “dirty”) with correct classifications ranging from 83 to 100 %. However, 20 PLS factors were required which might limit the robustness to the models. Vis–NIR spectroscopy in both transmission and reflectance modes has been employed to detect brown heart of pears (Fu et al. 2007) . They found that, using discriminant analysis, they could discriminate between brown heart pears and non-brown heart pears. Transmission spectra were more successful than reflectance spectra in this classification: a classification rate of 91.2 % using transmission spectra.

A conceptual view of an NIR transmission-based system for apple assessment (Fig. 4.5) has been proposed by McGlone and Martinsen (2004) . They employed two prototype on-line NIR transmission systems to determine the percentage of internal tissue browning in apples. One prototype used time-delayed integration spectroscopy (TDIS) in which light transmitted through a moving object was electronically tracked as it moved through the spectrometer’s field of view. The other used a large aperture spectrometer (LAS) in which the light from the object is accumulated in a series of one-shot measurements as the fruit progresses through the field of view (McGlone and Martinsen 2004). The systems operated 500 mm s−1. The LAS system gave the best results (R 2 = 0.9) for fast on-line assessment of apples .

Fig. 4.5
figure 5

A conceptual view of NIR transmission system. As the fruit passes through a relatively large field-of-view in the TDIS system (a), a detector simultaneously accumulates many sequential points over three apples. In contrast, the LAS system (b) takes a simple snapshot, like a camera, over a much shorter time for a small portion of one fruit (McGlone and Martinsen 2004). (Reprinted with permission from Journal of Near Infrared Spectroscopy 12(1), 37–43 (2004). Copyright: IM Publications LLP 2004)

Full size image

Further developments in defect identification have focused on the use of multispectral or hyperspectral imaging (Ariana et al. 2006; Blasco et al. 2007) . This emerging platform technology is discussed in Chap. 9.

4.5.3.3 Quality Classification

A key quality characteristic of fruit is SSC. As fruit ripen, there is conversion of insoluble starch into soluble solids, to which the simple sugars (glucose, fructose and sucrose) make the largest contribution (Martinsen and Schaare 1998) . Numerous studies have investigated infrared spectroscopy to predict this parameter non-destructively and have been summarized in Table 4.4. The majority of such studies have utilized NIR spectroscopy. A study by Lammertyn (2000) compared two optical configurations, i.e. a bifurcated and a 0°/45° optical configuration. They found that while the former configuration gave slightly better performance for the prediction of SSC, they recommended 0°/45° configuration for commercial applications as it had a lower cost and could be used for non-contact measurements. However, bifurcated reflectance-based instruments have found an array of applications (Fig. 4.6). It should be noted that numerous variables (e.g. cultivar, geographic origin, etc.) can affect the performance of such predictive models, and therefore studies which have independently validated models, for example, over and within seasons, are crucial to an assessment of model robustness (Golic and Walsh 2006) . Golic and Walsh (2006) collected NIR spectrum of peaches, nectarines and plums and found that model performance for SSC was acceptable when peaches and nectarines were combined, but it was best if a separate plum model was employed. They also stated that model performance was stable over several seasons in terms of R 2 (typical R 2>0.8) .

Fig. 4.6
figure 6

A NIR (LabSpec) with bifurcated fibre optic probe for contact reflectance measurement

Full size image

4.5.4 Meat and Poultry

4.5.4.1  Fresh Meat

A number of studies examined the application of infrared spectroscopy to fat extracts to predict meat quality as fatty acid composition of meat can determine its processing quality. Villé et al. (1995) developed a method for the determination of total fat and phospholipid content in intramuscular pig meat using FTIR spectroscopy. They employed an extraction using chloroform and methanol. FTIR spectra were subsequently recorded in transmission mode, and utilizing selected regions of the FTIR spectra related to the C = O bond (1785–1697 cm−1) developed linear regression equation to predict total fat (R 2 = 0.99). A study has also examined the use of FTIR spectroscopy in the NIR and MIR regions of fat extracts and non-processed pork to determine the fatty acid content in fat slices and fat extracts (Ripoche and Guillard 2001) . They found that MIR spectra using an attenuated total reflectance samples accessory (R 2 ~ 0.91–0.98) and NIR transmission spectra (R 2 ~ 0.85–0.96) of fat extracts could be used to predict saturated fatty acids (SFA), monounsaturated fatty acids (MUFA), polyunsaturated fatty acids (PUFA), palmitic acid (C16:0), oleic acid (C18:1) and linoleic acid (C18:2). However, with 9–15 latent variables included in the models, they may not be very robust . While NIR reflectance spectroscopy successfully predicted SFA, PUFA, C18:1 and C18:2 from spectral measurements of the back and breast fat, MUFA and C16:0 could not be predicted. Mitsumoto et al. (1991) used NIR spectroscopy in reflectance and transmittance mode to predict the quality of beef cuts Warner–Bratzler shear value (tenderness) (R = 0.798−0.826), protein (R = 0.822−0.904), moisture (R = 0.895−0.941), fat (R = 0.890−0.965) and energy content (R = 0.899−0.961) were successfully predicted using both modes. Park et al. (2001) also developed models for predicting the tenderness, i.e. Warner–Bratzler shear value of beef using NIR reflectance spectra and principal component regression (PCR). The coefficient of determination of the developed models were of a similar order (R 2 = 0.612 – 0.692). This technology has also been commercially investigated with instruments such as the QualitySpec BT Spectrometer from Analytical Spectral Devices (Fig. 4.7) .

Fig. 4.7
figure 7

The QualitySpec BT Spectrometer from Analytical Spectral Devices Inc. for measuring meat quality

Full size image

NIR spectroscopy has also been investigated at laboratory scale for determination of the maximum temperature to which beef had been subjected to during a heat treatment (Ellekjaer and Isaksson 1992) , species identification (Ding and Xu 1999) and authenticity assessment (Fumiere et al. 2000) . Other applications of NIR such as the detection of faecal contamination on poultry have been studied. Windham et al. (2003) applied Vis–NIR spectroscopy to discriminate between uncontaminated poultry breast skin and faeces . They found that the developed model could successfully classify faecal-contaminated material due to spectral differences between faecal colour and myoglobin and/or hemoglobin content of the uncontaminated breast skin. However, hyperspectral imaging (Chap. 9) has also been utilized for such an application (Heitschmidt et al. 2007; Liu et al. 2007; Park et al. 2006a, b, 2007) .

4.5.4.2 Ground Meat Quality

The quality of ground meat used as a raw material in products such as burgers and sausages is critical as processors must comply with product-type-dependent restrictions, i.e. chemical composition and origin of raw materials (Togersen et al. 1999) . Togersen et al. (1999) utilized an on-line NIR sensor to determine the fat, water and protein contents in industrial-scale meat batches (beef and pork) in an industrial environment. The NIR sensor was installed at the outlet of a large meat grinder. The models developed had RMSECV of 0.82–1.49 %, 0.94–1.33 % and 0.35–0.70 % for fat, water and protein, respectively. Togersen et al. (2003) went on to predict the chemical composition of industrial-scale batches of frozen beef using a similar system. The resulting RMSECVs were 0.48–1.11 % (fat), 0.43–0.97 % (moisture) and 0.41–0.47 % (protein).

NIR spectroscopy has also been investigated as a tool for detecting adulteration of hamburgers (Ding and Xu 2000) . They found it was possible to predict the level of adulterants in hamburgers with errors of 3.33, 2.99, 0.92 and 0.57 % for the adulterants mutton, pork, skim milk powder and wheat flour, respectively .

4.5.4.3  Meat Emulsion

Optical sensors have also been developed to monitor meat emulsion stability (Alvarez et al. 2007, 2009, 2010a, b) . Initial work focused on prediction of meat emulsion stability using reflection photometry (Alvarez et al. 2007). They found that L* values increased at the beginning of chopping associated with reduced cooking losses, following 8 min of chopping there was a reduction in L* and b* values and an associated increase in cooking losses, which suggested the feasibility of an on-line optical sensor technology to predict the optimum end point of emulsification in the manufacture of finely comminuted meat products. These authors then recorded light backscatter intensity from beef emulsions manufactured with different fat/lean ratio and chopping duration using a dedicated fibre optic prototype (Alvarez et al. 2009). They found several optically derived parameters to be significantly correlated with fat loss during cooking. In subsequent work, they found normalized intensity decreased with increased chopping time as a result of emulsion homogenization, and with increased distance, chopping time had a positive correlation with fat losses during cooking, which in turn had a negative correlation with normalized light intensity and loss of intensity. Therefore, they suggest that light extinction spectroscopy could provide information about emulsion stability (Alvarez et al. 2010) .

4.6  Future

Infrared spectroscopy has been demonstrated to be an excellent PAT tool for monitoring critical processes and prediction of quality indices during food processing. Advances in equipment design will assist in the deployment of infrared ­spectroscopy-based technologies as PAT tools in the food industry. This will include improvements in robustness , cost and advances in microspectrometers. However, where studies have primarily been at laboratory scale, further research is required to ensure appropriate scaling up and transfer of the technology to industry. The combined acquisition of spectral and spatial information through the use of hyperspectral imaging has a number of potential applications. However, further developments are required to reduce the cost and increase the acquisition and processing speed for it to be fully exploited in food quality and safety applications.