Grains show different spectral response characteristics in the near-infrared region when containing different contents of nutritional ingredients, so the nutritional attributes can be quantified based on spectral features. Shi et al. [114] collected NIR spectra of 48 wheat samples at 680 to 2500 nm and established the optimal NIR model between 1400 and 2500 nm using mean-centered partial least square regression (PLSR). The R² of crude proteins is predicted to be 0.98. Amanah et al. [115] determined diffuse reflectance spectra of 310 soybean samples in the waveband ranging from 1000 to 2500 nm and studied characteristic absorption bands of different nutritional ingredients of soybeans. In addition, they also built the PLS multi-variate analysis model for isoflavones and oligosaccharides in soybeans, and the predicted R₂ scores of the model for isoflavones and oligosaccharides are separately 0.80 and 0.72. Nutritional ingredients also differ in different varieties of grains in a same type, so the different varieties can be discriminated according to differences in nutritional attributes. Through principal component analysis (PCA), Wadood et al. [127] screened the spectral variables at 950 to 1650 nm and established a classification model for linear discriminant analysis (LDA). The model shows the highest classification accuracies of 100% and 73% separately for the origin and production year of wheat, and the highest classification accuracy of 98.2% for the genotypes of wheat grains. Fan et al. [128] coupled machine learning methodologies with a spectral dimensionality reduction algorithm to establish an assessment model for seed vigor, the accuracy of which exceeds 84.0%. Based on PCA, Basati et al. [129] classified healthy and unhealthy wheat kernels at the accuracy of 100%, and the two types of wheat kernels are more effectively discriminated at wavelengths of 839 nm, 918 nm, and 995 nm. Zhang et al. [130] eliminated interference wavelengths successively using the equidistant combination-partial least squares and wavelength step-by-step phase-out PLS. On this basis, the constructed novel purity analysis model of rice seeds shows a predicted correlation coefficient of 0.920. Schütz et al. [131] established a global classification model for the geographical origin of grain maize by adopting the SVM classification technique, which achieved an overall classification accuracy of 95%. Qiu et al. [132] differentiated viable and unviable maize seeds based on spectral data within the waveband ranging from 1000 to 2500 nm and the established partial least square-discriminant analysis (PLS-DA) classification model shows the classification accuracy of 98.0%. By using the random forest algorithm, Wang et al. [133] screened spectra of soybeans in the range of 5793 to 12,489 cm⁻¹ and selected important variables to establish the classification model, which achieved an 80% classification accuracy of soybeans. Kusumaningrum et al. [134] used the variable importance in projection method to screen 146 optimal wavelengths from 1000 to 2500 nm and built the soybean seed viability prediction model combining PLS-DA. The model shows the prediction accuracy approaching 100% for viable and unviable soybean seeds.

HSI acquires spectral information when collecting image information, so it can be used to detect the nutritional attributes of grains. HSI allows segmentation of hyperspectral data from single grain kernels and also can segment hyperspectral data of different parts of single grain kernels, so it exhibits unique advantages in quality analysis of single grain kernels. Caporaso et al. [31] collected hyperspectral images of individual wheat kernels from 980 to 2500 nm and the R² scores for the calibration and validation sets of the established detection model for the protein content based on PLSR analysis are separately 0.82 and 0.79. Fu et al. [116] selected 23 characteristic wavelengths of oleic acids and 21 characteristic wavelengths of linoleic acids in hyperspectral images of soybean kernels in the range from 600 to 1000 nm. On this basis, they established the PLS-DA classification model of different oil contents, which shows the classification accuracies separately of 98.6% and 86.7% for the calibration and validation of oleic acids. Liu et al. [117] sampled hyperspectral images above and below the embryos of individual maize kernels in the range from 930 to 2500 nm and used the competitive adaptive reweighted sampling (CARS) method to select characteristic wavelengths. They established the PLSR and non-linear statistical data model using the ANN algorithm and its predicted R-value for maize starch content is 0.96. Weng et al. [135] collected hyperspectral images from 10 high-quality paddy varieties in China at 400 to 1000 nm. The classification model for paddy varieties based on PCA networks using spectral and morphological features exhibits classification accuracies separately of 98.66% and 98.57% in the training and prediction sets. Guo et al. [136] obtained hyperspectral images of the mixture of rice produced in Wuchang City and Jiangxi Province, China from 380 to 1000 nm and the established full-wavelength SG1-PLSR model shows a predicted R-value of 0.9909. The R-value of the simplified PLSR model established based on 15 important wavelengths selected according to the weighted regression coefficient is 0.9769. Miao et al. [137] imaged 800 maize kernels in eight varieties in the range from 386.7 to 1016.7 nm, and the t-distributed stochastic neighbor embedding (t-SNE) model pre-processed via PCA exhibits the highest classification accuracy of 97.5%, which is far higher than other models (75% for PCA, 85% for kernel principal component analysis, and 76.25% for local linear embedding). Li et al. [54] collected 3600 hyperspectral images of soybeans in the range from 866.4 to 1701.0 nm and established the one-dimensional CNN model of soybeans. The classification accuracies of the model for four soybean varieties (Zhonghuang 13, 37, 39, and 57) exceeded 98%.

The molecular vibration and rotation information can be obtained based on Raman spectra of nutritional ingredients in grains at different frequencies of incident light, thus realizing the quantitative detection and qualitative analysis of the quality [138,139]. RS can be used for qualitative analysis such as the varietal classification of grains including wheat, paddy, and maize. Liu et al. [118] collected Raman spectra of eight wheat varieties with waxy proteins from 700 to 2000 cm⁻¹, measured spectra in several specific regions, and established the discriminant analysis (DA) classification model. The accuracies of the model in the calibration and validation sets are 94.4% and 94.6%, respectively. Zhu et al. [119] acquired Raman spectra of 107 paddy varieties in China from 200 to 1500 cm⁻¹, used PCA for preliminary evaluation, and established the soft independent modelling of class analogy model. The model can identify high-quality and poor-quality rice at an accuracy of between 80 and 100%. Liu et al. [140] collected Raman spectral data from 760 maize seeds from four different varieties from 400 to 1800 cm⁻¹, selected characteristic wavelengths through CARS and established the MCARS-genetic algorithm-SVM model in 13 wavebands. The calibrated classification accuracy and prediction accuracy of the model are 99.29% and 100%, respectively. The grain quality can be quantitatively assessed based on characteristic peaks on Raman spectra of different nutritional ingredients in different grains. Pezzotti et al. [120] proposed and verified the quantitative algorithm for calculating the amylose and amylopectin components based on spectral differences across 830 to 895 cm⁻¹. They also constructed the Raman equation for quantitatively determining the phenylalanine (PHE) and tryptophan components based on the significant features of PHE at 1003 cm⁻¹ and the breathing vibration of indole rings of tryptophan at 768 cm⁻¹. Singh et al. [121] collected transmitted Raman spectra of soybean flour from 400 to 1950 cm⁻¹ and established the multi-variate linear regression classification model based on the spectral measure. The predicted R₂ scores of proteins and fats in the soybean flour are separately 0.87 and 0.87. Yang et al. [122] studied Raman spectral characteristics of maize seeds from 380 to 180 cm⁻¹ according to samples of standard chemical components of maize, including maize starch, protein, oil, and bran. They found that the characteristic peaks at 477, 1443, 1522, 1596, and 1654 cm⁻¹ in their Raman spectra are separately correlated to the maize starch, oil-starch mixture, zeaxanthin, lignin, and oil.

Based on the NMR principle, a series of detectable inductive signals with attenuated intensity are generated when exciting atomic nuclei with fixed magnetic moments in samples using a specific impulse train. The intensity of the relaxation signals is in direct proportion to the number of nuclear spins in the detected samples (a quantitative basis) and the signal attenuation process is closely related to the constituent structure of the detected substances (a qualitative basis). The inverse analysis of signals using a mathematical method can yield information about various components and microstructures that fail to be obtained using other approaches, thus realizing the detection objective. Li et al. [123] identified the mass and oil content of individual maize kernels using an NMR sorting system. The mass of individual maize kernels is listed (in descending order) as haploid, diploid, and aborted kernels on the whole, with an extremely significant difference between diploid and aborted kernels and insignificant differences among other kernels; the oil contents are listed (in descending order) as diploid, haploid, and aborted kernels, and the oil contents are distributed in gradients in the three types of kernels.

In X-ray fluorescence spectrometry, various elements in samples release X-ray fluorescence with characteristic energy after being excited by irradiating these samples using X-rays. By measuring the wavelength and intensity of these X-rays, the elements present in the samples can be quantified. X-rays can be utilized to detect trace elements in grains. Lemmens et al. [124] adopted pearling, synchrotron X-ray fluorescence microscopy mapping, and X-ray absorption near-edge structure imaging of wheat to reveal changes in mineral elements such as Zn, Fe, and S during germination of wheat. This provides an opportunity for improving the nutritional quality in the food processing process. Chen et al. [125] investigated wheat flour samples by using the energy dispersive X-ray fluorescence spectrometry and extracted characteristic variables using the CARS. They also identified 12 energy variables (corresponding to mineral elements) to characterize the geographical attributes of samples. Through PCA and quadratic discriminant analysis, they established a non-linear model, which has the identification accuracy of 84.25% for origins of wheat flour through five-fold cross validation. Nguli et al. [71] estimated concentrations of Mn, Fe, Cu, and Zn in soybean kernels based on total reflection X-ray fluorescence, and their results show that soybean contains sufficient trace elements. Except for Fe which is present in high concentrations, concentrations of other elements are relatively low yet within the necessary range.

Detection based on acoustic features is used to evaluate grain quality according to the vibration responses of grains under certain excitation. The interactions between acoustic vibration and grains include reflection, scattering, transmission, and absorption. The propagation mode of acoustic vibration is determined by the vibro-acoustic characteristics of grains, which are related to the mechanical and structural properties of grains. Responses of grains to vibration are dependent on the elastic modulus, Poisson’s ratio, density, mass, and shape of grains, and the mechanical and structural properties of grains change with the growth and metabolism of cells. Detection based on acoustic features is widely used to assess attributes that are related to the mechanical and structural properties [93]. Khorshidi et al. [141] detected wheat varieties through transmission of longitudinal ultrasonic waves at five frequencies and established the PCA and generalized linear model thereof. Their results imply that, compared with other frequencies, ultrasonic measurement at 10 MHz can better identify wheat varieties and the ultrasonic phase velocity and longitudinal storage modulus at 10 MHz are the optimal discrimination factors for wheat varieties. Khorshidi et al. [141] also verified the capacity of low-intensity ultrasound at 10 MHz in identifying wheat varieties using 23 red, hard, spring wheat varieties. They confirmed through PCA that the ultrasonic parameters are significantly correlated to the protein content, falling number, and grain ash content of wheat. An acoustic detection method can be applied to quality detection and variety classification of grains, while it is susceptible to external environmental interference and can only be applied to a limited number of scenarios.

Mechanical properties of grains refer to mechanical features thereof under all types of applied load. Determining the mechanical parameters (rupture force, rupture energy, etc.) of grains under load can reduce the loss of grains in the harvest, transportation, storage, and processing processes. Research into the organizational structures and mechanical properties of grains provides a certain physical basis for improving the nutritive value and taste of grain products. Mechanical properties of grains are also an important factor that determines their processing quality, and determining the type and calculating the value of load borne by grains in the processing process can provide a basis for precision-design of relevant agricultural machinery [95]. Feng et al. [63] designed an orthogonal rotation combination test scheme. By using a multi-parameter controllable thin layer drying test bench, they dried maize under different conditions (temperatures of 30 to 60 °C, relative humidity of 30% to 60%, air velocities of 0.46 to 0.94 m s⁻¹, initial moisture contents of maize of 20% to 30% w.b., and tempering ratios of 0 to 3). Thereafter, they employed a texture analyzer to measure various mechanical properties of dried maize kernels, including the rupture force, rupture energy, elastic modulus, and brittleness. They then established the model for relationships of the rupture force, rupture energy, elastic modulus, and brittleness with the drying conditions. As the drying temperature is increased from 30 °C to 60 °C, the rupture energy, elastic modulus, and brittleness of maize kernels separately are increased by 19.11%, 11.76%, and 4.02%; when the relative humidity during drying is increased from 30% to 60%, the rupture force, rupture energy, elastic modulus, and brittleness separately are increased by 15.07%, 13.74%, 20.73%, and 3.31%. Delwiche et al. [50] explored the relationships of super-soft wheat with three compressive strength properties (maximum stress, Young’s modulus, and Poisson’s ratio), with the correlation coefficients separately being 0.76, 0.66, and 0.75. Grain quality based on mechanical properties was detected mainly on the basis that the quality distribution of grains influences organizational structures of grains, which are then shown as the difference in mechanical properties. At present, the correlation between nutritional ingredients and organizational structures of grains remains unclear, and the grain quality detection based on mechanical properties is still in the initial exploratory stage.

Lu et al. [126] measured the original voltametric signals of rice-flour slurry using a self-made E-tongue and built a CNN+BpNN model. The training and prediction accuracies of the model for physical and chemical indicators including the degree of chalkiness, amylose, protein, starch, and total metal content are separately between 84.3% and 92.0% and 81.9% and 89.5%. Zhao et al. [40] assayed rice filtrate by using an E-tongue (α-Astree, and Alpha MOS, France). Discrimination factors DF1 and DF2 separately account for 97.24% and 97.58% of the total variability of paddy in the storage period, and the DF1 value of storage devices shows an increase trend while the DF2 value first increases, then decreases. The results show that rice samples under different storage conditions can be distinguished. Guo et al. [142] acquired current response signals of wheat leachates during five different storage periods using a self-made large amplitude pulse voltammetry (MLAPV) based E-tongue system. They also built a detection model based on wavelet packet transform, improved artificial fish swarm algorithm, and extreme learning machine. The model achieved the training and testing accuracies separately of 96% and 92% for aged wheat stored for five different periods.

NIRS, HSI, and RS can all be used to detect multiple quality indices; NMR detection can detect oils; X-rays can be employed to detect trace elements. In comparison, acoustic and mechanical responses of grains are insignificantly correlated with the nutritional attributes of grains; because grain leachates need to be prepared when detecting grains using an E-tongue, application of E-tongues is, to some extent, limited.

2.3. Non-Destructive Inspection Methods for Safety Attributes of Cereal Grains