Qualitative Applications. One of the first studies using this technology in the agricultural sector focused on the detection of meat and bone meal in compound feedstuffs using HSI in the NIR focal plane MatrixNIR (Malvern
Instruments Ltd., Malvern, UK; wavelengths 900-1,700 nm at increments of 6 nm) (59). The method was developed to enforce food legislation adopted after the European mad cow crisis connected with Creutzfeldt-
Jakob diseases in humans. Thousands of spectra in a massive space had to be collected simultaneously. For detection, the chemometric support vector machine (SVM) tool was used. This alternative method of NIR focal
plane MatrixNIR was suggested as being more effective than methods used at that time, which were cumbersome and required a specialist. Similar studies were carried out by Riccioli et al. (60) for discriminating
between terrestrial and fish species in animal protein by-products used in livestock feed. The samples were analyzed by NIR chemical imaging (NIR-CI) in the 1,000-1,700 nm wavelength range. Four algorithms—Mahalanobis distance, Kennard-Stone, spatial interpolation, and binning—were applied in order to select an appropriate subset of pixels for further partial least squares discriminant analysis (PLSDA). For the four algorithms used, the classification accuracy obtained was higher than 99.61%. Kim et al. (61) used fluorescence HSI for the detection of skin tumors (ulcerous lesions surrounded by a rim of thickened skin and dermis) on chicken carcasses, replacing the time-consuming, expensive, and uncomfortable organoleptic inspection method. They used an HSI system from the U.S. Department of Agriculture’s Instrumentation and Sensing Lab (ISL-HSI), which includes a CCD camera, a spectrograph, a sample transport mechanism, and lighting sources (wavelengths 425-711 nm). The detection rate was 76%, indicating that the method needs to be improved; some spots were irrelevant for tumors and some carcasses were not filtered out in the spatial classifier, giving a false-positive rate that was too high. Various studies were conducted to detect the contamination of poultry carcasses with visceral content (62, 63). In order to demonstrate that it is possible to detect fecal and ingested contaminants using NIR-HSI, Lawrence et al. (62) and Park et al. (63) used an imaging camera consisting of a focusing lens, a prism-grating spectrograph, and a high-resolution CCD camera. Park et al. (63) used the region of interest (ROI) algorithm at wavelengths of 290-1,000 nm and obtained an accuracy of 96.6% using principal component analysis (PCA). The imaging system operated from wavelengths of about 400 to 900 nm (62). Similar studies
were carried out by Wang and El Masry (64) for apple bruise detection based on physical and chemical changes compared with unbruised fruits (wavelengths 400-1,000 nm). They developed a model using different
algorithms: minimum noise fraction transform (MNF), ROI, PCA, PLS, and artificial neural networks (ANNs).The 750-, 820-, and 960-nm wavelengths were chosen for bruise detection. In order to determine the potential of
the selected wavelengths for bruise detection, PCA was conducted with successful results, such as 93.25% of the variance between normal and bruised spectral data (principal component 1 [PC1]: 70.01% and principal
component 2 [PC2]: 23.94%). Nagata et al. (65) and Nagata and Tallada (66) worked successfully on strawberry bruise detection. The LDA algorithm was used at a range of 825 and 980 nm and the rate of discrimination was
greater than 90.70% for the calibration model, whereas for validation it was greater than 86.50%. Other applications of the ISL-HSI system were used for vegetables. The system was used successfully for detecting
cucumber chilling injury, with recognition rates of 93.30% for injured cucumbers and 88.30% for uninjured cucumbers (67). With this system, the hyperspectral images were acquired at wavelengths of 448-951 nm, with a
4.5-nm interval. For detection, PCA and Fisher’s linear discrimination (FLD) were used. NIR-HSI is also used to measure food quality, particularly fruit quality. For consumer acceptance and fruit shelf life, firmness is very important and it is necessary for the industry to use a nondestructive sensing system to evaluate it. High scattering from a surface depends on the cell structure of the food and is related to the texture. Scattering profiles can therefore be used to predict fruit firmness. A VIS/NIR-HSI system based on a Varispec Liquid Crystal Tunable Filter (LCTF, Cambridge Research and Instrumentation) was used for firmness detection, and LDA, normalized difference (ND), and ANN algorithms were used to analyze the spectra (66).
For strawberry firmness detection, wavelengths of 665-685, 755-870, and 955-1,000 nm were shown to be optimal (standard error of prediction [SEP] around 0.258 in the case of 70% to fully ripe strawberries and 0.350
in the case of 50% to fully ripe strawberries) (66). A similar technique was used by Lu and Peng (68) to determine peach firmness. The most important bands for predicting peach firmness were found to be around 677,
710-850, and 950 nm. The Lorentzian distribution (LD) parameter combinations for firmness calibration of two types of peaches (Red Haven and Coral Star) were chosen. The coefficients of determination (R2) obtained with multilinear regression (MLR) were between 0.51 and 0.58 and between 0.67 and 0.77, respectively. In both cases, further analysis was needed to obtain better results. Beef color and tenderness are two major parameters of beef quality.