Consumers are conscious of the effects of food quality and composition on nutrition and well-being. Different types of allergies and intolerances have become more common, and ethical questions about food production raise concerns.
In addition to increasing consumer requirements, food producers have a legal obligation to control the quality and safety of their products. Failures in the production process can cause considerable financial losses and even risk consumer safety.
From labs to realtime inspection
There are several technologies available to monitor the quality of food production. Unfortunately, many of them require laboratory analysis which is labor-intensive and slow. From measurement to results can take days or even weeks. Also, the sample is often destroyed during the inspection.
Food production process control must work with short response times. Quicker and more sophisticated inspection and analysing methods are required in monitoring food production in-line to ensure consistent quality and safety.
Spectroscopy vs hyperspectral imaging in food analysis
Spectroscopy offers a fast and nondestructive method for quality inspection, and it is a well-established method in the food industry. It allows the determination of the composition (e.g., moisture or protein content) and physical characteristics (e.g., the particle size) of samples.
Many relevant chemical bonds in food absorb light at particular wavelengths in the SWIR (900–2500 nm) range. These absorption features reveal the chemistry of the sample and can be detected by spectroscopy.
Spectroscopy is an ideal method for analyzing the average composition of homogeneous samples such as flour. Its disadvantage is that it only provides information on a single sample point. Therefore, it is unsuitable for analyzing more complex multi-component food products.
Hyperspectral imaging combines spectroscopy with imaging. This technique allows for simultaneous pixel-by-pixel analysis of samples in real time. Like spectroscope, hyperspectral imaging doesn’t require contact with the sample and doesn’t destroy or contaminate the food.
Measurements of individual absorbance bands or calibrations for full spectra provide information on the composition, which can be mapped to measure the distribution of components, such as moisture or fat.
Hyperspectral imaging also enables complex, multi-component product analysis, which is difficult with other techniques.
Hyperspectral sensors provide information on hundreds of narrow wavelength bands. Identifying the absorbance bands for different chemical bonds enables the development of calibration models for specific metrics relevant to food production.
It is significantly more information than traditional RGB or spectral sensors can provide. However, to create an application, the user must know, which wavelength bands are of interest and recorded for analysis, and which can be left out.
Campden BRI pioneers hyperspectral analysis for the food industry
Campden BRI is a UK-based company whose history traces back to the year 1919 when it was opened as a Fruit and Vegetable Preserving Research Station. Today, Campden BRI is the world’s largest membership-based food and beverage research organization, with over 2,400 members in over 80 countries. Members include companies like Arla Foods Ltd., Kellogg’s, Coca-Cola, Heinz, and Nestlé.
As a pioneer, Campden BRI explored the opportunities of hyperspectral imaging to strengthen its food analysis methods and expand its food imaging capabilities already 15 years ago.
“SWIR spectroscopy was already well established in the agri-food sector to rapidly analyze foods and their ingredients. Hyperspectral imaging is a unique way to measure the distribution of moisture and fat in complex food samples, for example. It also provides the opportunity to apply this approach to new applications”, states Dr. Martin Whitworth, Principal Scientist at Campden BRI, responsible for leading image analysis research.
Push-broom imaging for real-time inspection
For inline production, the short imaging time is a critical requirement. It is also essential for offline analysis because the intense illumination can damage, i.e., dry or melt the sample. In addition to imaging time, Campden BRI had a set of other criteria the system needed to fulfill.
“It had to operate on 900–2500 nanometers, be suitable for different product sizes, measure in a few seconds, and be transportable to different production locations,” says Martin Whitworth. Based on these criteria, Campden BRI decided to go with the Specim SWIR push-broom hyperspectral camera and has used the same sensor since 2008.
The push-broom camera detects the full spectrum of several narrow bands and scans the total sample in real-time on a production line. The imaging time can go down to milliseconds since it scans only one narrow line at a time. The ability to apply the same method used in the laboratory to inline production was another reason for selecting the push-broom hyperspectral imaging system.
Today, Campden BRI has been offering hyperspectral imaging analysis services for the food industry for almost two decades. They have applied the system to analyze various products, including bread, biscuits, grain, meat, fish, confectionery, and fried products.
Much of Campden BRI’s work is contract analysis for individual food manufacturers. It involves using models and calibrations developed exclusively for clients’ products. They analyze the samples in their laboratories or transport the instruments to client sites to analyze fresh samples. In addition to providing product development information, Campden BRI conducts preliminary trials to evaluate the method’s suitability for inline applications.
Moisture distribution on food products
A typical application of hyperspectral imaging is to measure moisture distributions in products. Moisture distribution is an essential attribute of many food products. It affects texture and conditions favorable for microbial activity.
Moisture can change over shelf life and is therefore associated with product freshness. The moisture distribution is not always uniform, and it is difficult to detect without imaging technology. For example, in baking, it can be used to study the effect of production conditions on the uniformity of final product moisture.
Moisture distribution in bread
Figure 1 below shows moisture distribution in a fresh slice of white bread from a study by Campden BRI. First, a calibration was built, and then this model was applied to actual samples.
The purple and blue color in the image indicates low moisture content, while the yellow and red color indicates high moisture content. It is easy to see that the moisture level increases rapidly towards the center of the bread, while the outer crust has a low moisture level.
Detecting moisture migration in multicomponent products
Hyperspectral imaging can also be applied to detect moisture migration in multicomponent products, which are traditionally more challenging than single-component products.
It is used, for example, to measure changes in moisture distribution over shelf life, which can be essential for products with multiple components of differing water activity, such as a low-moisture product with a high moisture filling.
Qualitative analysis of food composition with hyperspectral imaging
Since hyperspectral imaging provides pixel-by-pixel information on a sample, it is ideal for multicomponent product composition mapping.
Certain qualities like fat, moisture, or crystalline sucrose have clear spectral features in the shortwave infrared (SWIR) range. By calibrating against reference samples, quantitative measurements can be made.
Fully quantitative measurements require the development of separate calibration models for each component to be identified. This is appropriate for applications requiring regular analysis of the same sample type, for example, fat content in mixed meat.
However, useful comparative information can be obtained even without a full calibration for shorter-term applications or where reference samples are unavailable.
Some hyperspectral applications imaging use spectral data to identify and classify features of different compositions in an image. For many food applications, the full benefit of the method is a quantitative hyperspectral analysis to measure the concentration of particular compounds.
This is achieved by creating calibration models based on the comparison of hyperspectral images with reference measurements made by traditional methods for a series of samples.
There are plenty of different algorithms available (Partial Least Squares, Support Vector Machine, and Neural Networks, to name a few) to build these calibration models, which map the desired parameters in the sample to hyperspectral data output.
Different absorbance bands will be significant for calibration modeling depending on what materials or properties one wants to detect. For example, crystalline sucrose has a characteristic absorbance peak at 1435 nanometers. Lipid contains CH2 bonds with absorbance bands at 1724 and 1762 nanometers. OH bonds in water molecules have several SWIR absorbances, including at 1925 nanometers.
In some cases, qualitative assessments can be made using images at these specific wavelengths. However, best results are achieved using multivariate calibrations for a range of wavelengths, including for properties where the relevant choice of absorbance bands is unknown or where there are multiple overlapping bands.
After the calibration model is built, it can be applied to the hyperspectral images of unknown test samples or samples on the production line to map these parameters quickly.
The composition of chocolate bars
Below is an example of a comparative study of commercial chocolate bars. A full calibration would require access to reference samples of each component material type. However, instead of mapping the strength of particular absorbance bands, a useful comparison can be made.
Figure 2 below shows maps of three absorbance bands. A band for CH2 is typically associated with differences in fat content, another is characteristic of the presence of crystalline sucrose. The absorbance band for OH is typically associated with differences in moisture content.
Regions high in fat are shown in red, crystalline sucrose in green, and moisture in blue. Combinations of components are shown as mixed colors.
Nuts are shaded red indicating high-fat content. Caramel appears blue or purple, indicating high moisture content with varying fat content. Chocolate appears in green, yellow, or orange indicating varying combinations of fat and crystalline sucrose. The figure shows how the composition of the different ingredients differs between different commercial products.
A quantitative calibration for these properties could be developed for more detailed analysis for inline inspection.