**5. Novel emerging tools for quality & safety: non-destructive technology in the food chain**

Traditional analytical techniques for food and feed quality inspection and compositional assessment are typically invasive and time-consuming, requiring extensive sample prepara‐ tion, thus being unsuitable for applications in the highly demanding, fast-paced food proc‐ essing segment. Recently, novel techniques have been investigated for fast, reliable and chemical-free food quality assessments. Near-infrared (NIR) hyperspectral imaging has emerged as an efficient and advanced tool, combining both computer vision techniques and NIR spectroscopy, which can be used for continuous monitoring, process control and quality assessments of agricultural products, food and feed materials. Because most food quality features are related either to the external appearance of the product or its chemical composi‐ tion, either computer vision or NIR spectroscopy alone is adequate for monitoring organic samples in a fast, reliable manner. However, such techniques are still strongly dependent on other reference methods. Prediction of physical characteristics and chemical composition using NIR spectroscopy and/or computer vision methods has been reported on meat (chicken, pork, beef, lamb), cereals and grains (corn, wheat, soy, rye, coffee, cocoa), and fruits and vegetables (apple, citrus, berries). More specifically, there has been major interest in this technique in quality control, food safety and security, i.e., detection and prediction of contamination in agricultural products.

A study [72] compared NIR calibration methods for predicting protein, oil and starch contents in both whole and ground maize samples in the spectral range of 1100–2500 nm for reflectance and 680–1235 nm in transmittance modes. While the best models were obtained for the reflectance spectra of the ground samples, it was suggested that the transmittance mode for whole grains might be more useful due to its greater speed of analysis. Another study [73] developed a rapid single kernel NIR sorting instrument for maize and soybean. Prediction models for moisture of both seed types, and protein contents for soybeans were developed utilizing a spectrometric range from 906 to 1683 nm.

NIR reflectance and transmittance technologies have been investigated for contamination assessments of a range of cereal grain physical quality and chemical traits, and detecting and predicting levels of mycotoxins. Numerous applications have been developed, and cover almost all cereals in the globally important food grains, i.e., corn, wheat, rice and barley. An additional application has been to demonstrate the value in sorting grains infected with fungus or mycotoxins, such as deoxynivalenol, fumonisins and aflatoxins [74].

A shortwave infrared (SWIR) hyperspectral imaging system in the wavelength range between 1000 and 2500 nm was used to assess the potential AFB1 contaminants on the surfaces of healthy corn kernels. Key wavelengths that can indicate AFB1 and are used to differentiate levels of AFB1 were identified. A minimum classification accuracy of 88 % was achieved for the validation set and verification set, indicating that hyperspectral imaging technology could be used to detect AFB1 at levels as low as 10 ng/g, when applied directly on the corn surface [75]. Another study assessed the applicability of NIR for the rapid identification of mycotoxigenic fungi and their toxic metabolites produced in naturally and artificially contaminated products. Two hundred and eighty corn samples were collected in north-central Italy and analysed for fungal infection, ergosterol, and FB1 content. The results indicated that NIR could predict the incidence of kernels infected by *F. verticillioides* and also the quantity of ergosterol and fumonisin B1 in the meal. The best predictive ability for the percentage of global fungal infection and *F. verticillioides* was obtained using a calibration model utilizing corn kernels (r 2 0.75 and SECV 7.43) and maize meals (r 2 0.79 and SECV) 10.95), respectively [76].

A recent study on the quality assessments of meat products [77] reported the application of NIR reflectance as a potential method to predict quality attributes of chicken breast (*Pectoralis major*). Spectra in the wavelengths between 400 and 2500 nm were analysed, presenting clear differences between different quality grades of chicken (Figure 6). PCA performed on the NIR dataset revealed the influence of muscle reflectance (L\*) influencing the spectra. PCA was not successful to completely discriminate between pale, soft and exudative (PSE) and pale-only muscles. High-quality PLSR were obtained for L\* and pH models predicted individually (R2CV of 0.91 and 0.81, and SECV of 1.99 and 0.07, respectively). Sample mincing and different spectra pre-treatments were not necessary to maximize the predictive performance of the models. The results suggest that NIR spectroscopy may represent a useful tool for the quality assessment of chicken meat.

The contamination of meat products has also been investigated. NIR transflectance and Fourier transform-infrared (FT-IR) attenuated total reflectance spectra of intact chicken breast muscle were collected and investigated for their potential use in the rapid, non-destructive detection

A study [72] compared NIR calibration methods for predicting protein, oil and starch contents in both whole and ground maize samples in the spectral range of 1100–2500 nm for reflectance and 680–1235 nm in transmittance modes. While the best models were obtained for the reflectance spectra of the ground samples, it was suggested that the transmittance mode for whole grains might be more useful due to its greater speed of analysis. Another study [73] developed a rapid single kernel NIR sorting instrument for maize and soybean. Prediction models for moisture of both seed types, and protein contents for soybeans were developed

NIR reflectance and transmittance technologies have been investigated for contamination assessments of a range of cereal grain physical quality and chemical traits, and detecting and predicting levels of mycotoxins. Numerous applications have been developed, and cover almost all cereals in the globally important food grains, i.e., corn, wheat, rice and barley. An additional application has been to demonstrate the value in sorting grains infected with fungus

A shortwave infrared (SWIR) hyperspectral imaging system in the wavelength range between 1000 and 2500 nm was used to assess the potential AFB1 contaminants on the surfaces of healthy corn kernels. Key wavelengths that can indicate AFB1 and are used to differentiate levels of AFB1 were identified. A minimum classification accuracy of 88 % was achieved for the validation set and verification set, indicating that hyperspectral imaging technology could be used to detect AFB1 at levels as low as 10 ng/g, when applied directly on the corn surface [75]. Another study assessed the applicability of NIR for the rapid identification of mycotoxigenic fungi and their toxic metabolites produced in naturally and artificially contaminated products. Two hundred and eighty corn samples were collected in north-central Italy and analysed for fungal infection, ergosterol, and FB1 content. The results indicated that NIR could predict the incidence of kernels infected by *F. verticillioides* and also the quantity of ergosterol and fumonisin B1 in the meal. The best predictive ability for the percentage of global fungal infection and *F. verticillioides* was obtained using a calibration model utilizing corn kernels (r

2 0.75 and SECV 7.43) and maize meals (r 2 0.79 and SECV) 10.95), respectively [76].

A recent study on the quality assessments of meat products [77] reported the application of NIR reflectance as a potential method to predict quality attributes of chicken breast (*Pectoralis major*). Spectra in the wavelengths between 400 and 2500 nm were analysed, presenting clear differences between different quality grades of chicken (Figure 6). PCA performed on the NIR dataset revealed the influence of muscle reflectance (L\*) influencing the spectra. PCA was not successful to completely discriminate between pale, soft and exudative (PSE) and pale-only muscles. High-quality PLSR were obtained for L\* and pH models predicted individually (R2CV of 0.91 and 0.81, and SECV of 1.99 and 0.07, respectively). Sample mincing and different spectra pre-treatments were not necessary to maximize the predictive performance of the models. The results suggest that NIR spectroscopy may represent a useful tool for the quality

The contamination of meat products has also been investigated. NIR transflectance and Fourier transform-infrared (FT-IR) attenuated total reflectance spectra of intact chicken breast muscle were collected and investigated for their potential use in the rapid, non-destructive detection

utilizing a spectrometric range from 906 to 1683 nm.

154 Food Production and Industry

assessment of chicken meat.

or mycotoxins, such as deoxynivalenol, fumonisins and aflatoxins [74].

**Figure 6.** Average spectra for dark, normal and pale chicken samples (see [76] - "Reprinted from Food Chemistry, 168, Douglas Fernandes Barbin, Cintia Midori Kaminishikawahara, Adriana Lourenco Soares, Ivone Yurika Mizubuti, Moi‐ ses Grespan, Massami Shimokomaki, Elisa Yoko Hirooka, Prediction of chicken quality attributes by near infrared spectroscopy, 554-560, 2015, with permission from Elsevier").

of spoilage. PCA and PLS2-DA regression correctly classified 8 and 14 day samples (TVC days 8 and 14= 9.61 and 10.37 log10 CFU g−1) with several correlations that highlight the effect of proteolysis influencing the spectra. These correlations indicate that an increase in free amino acids and peptides could be the main factor in the discrimination of intact chicken breast muscle.

These studies have demonstrated that NIR methodology can be applied to monitor bacterial and fungal contamination in postharvest grains and fresh meats and to distinguish contaminated from clean batches to avoid cross-contamination with other materials during storage. However, there is still a demand for the development of cost-effective technolo‐ gies for high-speed sorting. In the area of food safety, it is important to create robust prediction models based on reference methods by including a wide range of samples from different regions. For instance, it is well known that a major drawback of this technology is the application of ready-to-use prediction models from one country into samples from another region of the world. Prediction models are usually built in developed countries, but they are not useful for samples originating in developing countries, mostly due to inherent differences in sample composition, cultivation methods, climate and soil character‐ istics, etc. Once researchers overcome these obstacles, this technology will benefit farm‐ ers, the industry and consumers if it enables contaminated grain and other food samples to be identified and removed from the food chain.
