Near-infrared spectroscopy (NIRS) is a widely used technique for performing quality control in laboratories that are associated with pharmaceutical [1], petroleum [2] and agriculture. Now commonly used for various applications, NIR, which absorbs at wavelengths of 800–2500 nm (12,500–4000 cm⁻¹) of the electromagnetic spectrum, was developed by Karl Norris, United States Department of Agriculture (USDA), in the 1960s for the purposes of quality assessment of agricultural products ^[3][4][3,4]. In 1962, Norris was already ahead of his time, using NIR data to build calibration models using advanced statistical methods ^[5][6][7][5,6,7]. The use of NIR has since been extended to applications such as quality assurance of agricultural products. For example, the dairy industry uses NIRS to perform measurements of lactose, protein and fat in milk ^[8][9][10][8,9,10]. There are also numerous reports of NIRS used in the grains industry, for example the discrimination among cultivars of wheat kernels [11], detection of fungal infestation [12] and prohibited additives ^[13][14][13,14]. Plant stress [15] and nutritive value [16] can also be measured using portable NIR spectrometers in the field for plant breeding programs. The high-throughput nature of these phenotyping imaging tools has allowed large-scale implementation in breeding by genomic selection [17] in agriculture. NIRS is also widely used in the meat industry for applications such as classification of poultry carcasses infected with disease in real time ^[18][19][18,19] and determination of sensory and texture characteristics of beef [20], as well as meat properties and chemical composition [20]. Applications in plant-based industries include fruit, grain and seed quality, particularly pathogen infestation and varietal purity, as well as chemical composition. The list of applications continues to grow rapidly ^{[21][22][23][24][25][26][27][28][29][30][31][32][33][34][35][36][37]}[21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37].

The main advantage of NIRS is that the methodology is non-destructive and high throughput. NIRS platforms provide spectral information, usually an average of a few selected points or the mean spectrum of a larger area. However, this is not so useful when objects are non-uniform and the features of interest are restricted to a relatively small but unknown part of the object [38]. When NIRS is coupled to imaging techniques, such as multispectral and hyperspectral techniques, both spatial and spectral information can be obtained. NIRS-multispectral imaging captures images within a small number of spectral bands, usually a maximum of 10 [39]. NIR-hyperspectral imaging (NIR-HSI), an extension of NIRS-multispectral imaging, captures hundreds of contiguous wavelength bands in the NIR region for each pixel. Coupling NIRS with an imaging technique was first introduced in the late 1990s [40] and the technology has advanced such that it is now routinely applied in fields such as military surveillance [41], astronomy [42], environmental monitoring [43] and agriculture.

The non-destructive nature of NIRS imaging allows quality control in agriculturally important fruit (e.g., apples, peaches and apricots) to be obtained, ^{[44][45][46][47][48]}[44,45,46,47,48] particularly for detecting external defects. Typically, fresh fruits are graded into categories based on parameters that affect quality, such as external defects, size, shape and colour, to comply with the Organisation for Economic Co-operation and Development (OECD) standards [49]. External defects are the most difficult to distinguish in automated systems; however, with the advent of NIRS imaging and, more recently, hyperspectral imaging, applications for detecting surface defects in fruit have been rapidly increasing. For example, NIRS and hyperspectral models for detecting surface defects originating from fungal infestation, insect damage on oranges [50], mangoes [51] and peaches [46], as well as damaged almond nut [52], have been reported. More recently, fruit ripening and maturity are also being assessed using NIRS imaging such as ripeness in grapes and firmness and maturity in lime and mango by determination of the total soluble solid and titratable acidity [53]. NIRS for sorting fruit based on sweetness is in use in more than 1000 packing houses in Japan using single-point measurements [54] and conceivably superseded with developments driven in NIR imaging as a visualization technique for sugar content and distribution ^{[54][55][56][57]}[54,55,56,57].

It has become increasingly evident that hyperspectral imaging can be applied to the non-destructive qualitative and quantitative determination of the desired features of selected samples, without contact [58]. Thus, it is very much suited for routine diagnostics such as food quality assessments and safety analyses [59]. A further advantage is that the NIR-HSI spectrum is collected at each pixel in the image, providing both the distribution and chemical composition of individual components [58], whereas NIRS only acquires a single spectrum for the sample [40]. This makes NIR-HSI suitable for heterogenous samples or for exploratory analyses, where the composition of the sample is largely unknown [38]. Despite the time and financial investment required for method development and the large data sets obtained, once a method is developed, it is inexpensive and routine with the benefits of reduced labour, turnaround and cost, compared to traditional methods used for inspecting and testing agricultural food products ^[60][61][60,61].

2. Hyperspectral Imaging Instruments

The components, configuration and design of a hyperspectral system are essential in the acquiring of reliable and high-quality images and data. A typical hyperspectral imaging system consists of a light source, wavelength dispersion device and an area detector (camera).

HSI systems are generally either active or passive systems. Active HSI systems are equipped with an active light source, such as those described in Table 1. A passive HSI system relies on ambient radiation, e.g., sunlight ^[62][66]. The light source is an important aspect in the excitation of the target sample and the quality of the images strongly depends on a well-balanced light intensity. Light sources include halogen lamps, light emitting diodes (LED), lasers and tuneable light sources. Fluorescent lamps are not recommended as there are inconsistencies in energy levels for different wavelengths ^[63][67].

Broadband light sources, including the halogen lamp, LED and tuneable light, are used with a wavelength dispersion device, such as a prism, grating, or filter, a key component in hyperspectral instruments ^[64][68]. It is placed between the detector and sample and it is used to disperse broadband light into different wavelengths. A filter is generally used for multispectral imaging systems whereas a prism and grating are widely used in hyperspectral systems. The final component of the system is the camera which has the role of quantifying the intensity of the light generated by the light source by converting photons into electrons. Charge-couple devices (CCDs) and complementary metal-oxide-semiconductors (CMOSs) are the two major cameras ^[64][65][68,69]. The CCD image sensor is superior to the CMOS device and generates high-quality image data.

The typical hyperspectral components utilised in the development of seed quality applications, include a halogen lamp light source, imaging spectrographs that covers VNIR, SW-NIR and NIR spectral ranges and, mostly, CCD-type image sensors.

Table 1.

 Advantages and disadvantages of using various light sources in hyperspectral imaging systems.

Light Source	Application	Advantages	Disadvantages	Example References
Halogen lamps	VIS NIR SW-NIR Broadband white light	Delivers smooth and continuous spectrum in the spectral range High light intensity	Short lifetime High heat Unstable 1 (operating voltage fluctuations) Sensitive to vibration	[66][70]
LED	From UV to SW-NIR, while some LEDs emit light from LW-NIR to MIR Broadband white light Excitation mode (fluorescence)	Small size Low cost Fast switching Long lifetime Minimal bulb replacement Low heat generation Low energy consumption Robust	Low spectral resolution Sensitive to wide voltage fluctuations High junction temperature Low light intensity	[62][66]
Laser excitation	Emission of fluorescence and Raman Narrowband pulsed light	Composition detection at pixel level High intensity light Narrower bandwidth than LED Signals are not interfered by carbon or water absorption	Detection of weak Raman signals is challenging due to high-fluorescence background	[67][71]
Tuneable light source (Quartz–Tungsten Halogen lamp)	Near UV VIS NIR	Area scanning Weak illumination (using wavelength dispersion) reduces heat damage of samples	No point or line scanning	[68][72]

¹ Low heat–load illumination is also available and provides an evenly distributed illumination line while emitting very low heat compared with the typical halogen lamp.

3. Image Acquisition Methods

Hyperspectral imaging collects information as images using well-defined spectral bands. The images are combined to form a three-dimensional (x, y, λ) hyperspectral data cube for processing and analysis, where x and y represent the two spatial dimensions of the scene and λ represents the spectral dimension (wavelengths) ^[69][70][64,73]. The resolution is characterized by the number of spectral bands of a range in the electromagnetic spectrum. Thus, multispectral images have low spectral resolution, given they have a maximum of 10 spectral bands, whereas HSI sensors have high resolution. A hyperspectral image comprises thousands or even millions of pixels. An ordinary image with 320 × 320 pixels contains 102,400 pixels. However, a hyperspectral image with 200 spectral bands has 20,480,000 pixels ^[71][74]. Spatial resolution, defined as the smallest distinguishable detail in an image [58], is also a factor; the size of the pixel or spatial point influences the signal-to-noise ratio ^[69][70][64,73].

Acquisition methods of a hyperspectral data cube include the following: (1) whisk-broom—one spectrum of a single point at a time; (2) push-broom—spectra of points of one spatial line at a time ^[72][75]; (3) tuneable filter—one waveband image at a time, much similar to a 2D photographic image (x, y); and (4) snapshot—full-waveband image at a time, much similar to capturing a photographic image with a third spectral dimension (x, y, λ). Push-broom is the most commonly used acquisition mode for applications in the food and agricultural industry. Push-broom imaging works either by the movement of the sample, for example, on a conveyer belt, or by directing the beam and detector to the region of interest ^[73][76].

Hyperspectral image acquisition is carried out in various sensing modes, including fluorescence ^[74][77], or one of four spectral modes typically utilized in the visible (400–700 nm) and NIR region—reflectance, absorbance, transmittance and interactance. These various imaging modes may be selected depending on the nature of the sample and the parameters being assessed. The light source for each optical mode is positioned differently to best capture the image based on the mode with minimal interference.

Hyperspectral Fluorescence Imaging (HSFI) detects chemical components that produce a fluorescence emission in the visible region (400–700 nm) when excited with short wavelengths (e.g., ultraviolet (UV) radiation or monochromatic laser light) ^[74][77]. These include components such as chlorophyll and some pigments in seed samples ^[75][78]. There are many examples of fluorescence imaging used in assessing physical and chemical quality parameters in food products ^[76][79], as well as faecal contamination in apples ^[77][80]. However, HSFI is unable to measure many of the food quality attributes that are detected by the optical modes (e.g., reflectance), such as soluble solid content, fruit pH and maturity discrimination ^[77][80].

Reflectance is the most common hyperspectral mode used in agriculture due to the ease in obtaining responses from higher light levels. Quality parameters such as size, shape, colour and surface defects are generally detected in reflectance mode ^[73][78][76,81].

Some applications show better calibration models and prediction outcomes in interactance and transmittance modes ^[79][80][82,83]. In transmittance mode, light is transmitted through the sample, with a detector opposite to the light source to capture the light that has passed through the sample; the strength is often weak and sample dependent but is considered to be a more valuable response in relation to internal components and defects ^[81][84]. In interactance mode, the light source and detector are located on the same side; however, the received light is sealed from the environment to prevent interference ^[82][85]. This particular mode is thought to be a combination of reflectance and transmittance as it can penetrate the sample, extracting more information than reflectance mode. The conformation of the light and detector is important to avoid refraction, specular reflectance and scattering in all three optical modes.

Another technique used in hyperspectral image acquisition is absorbance. Although there are limited reports of its use in agriculture, it is a preferred method for the quantitation of chemical constituents, for example, protein and oil contents of wheat grain ^[83][86].