Imaging spectroscopy has emerged as a reliable analytical method for effectively characterizing and quantifying quality attributes of agricultural products. By providing spectral information relevant to food quality properties, imaging spectroscopy has been demonstrated to be a potential method for rapid and non-destructive classification, authentication, and prediction of quality parameters of various categories of tubers, including potato and sweet potato. The imaging technique has demonstrated great capacities for gaining rapid information about tuber physical properties (such as texture, water binding capacity, and specific gravity), chemical components (such as protein, starch, and total anthocyanin), varietal authentication, and defect aspects.
Imaging spectroscopy integrates the main features of imaging and spectroscopic technologies, which can simultaneously acquire spatial and spectral information of an object . This technology has been widely used in the quantitative determination and visualization of food physical and chemical values. In a hyperspectral image, each pixel contains a continuous spectrum composed of hundreds of wavebands . The 3-dimension (3-D) spectral image with two spatial dimensions and one spectral dimension can be generated by area scan (tunable filter), line scan (pushbroom), or point scan (whiskbroom) . As the successor of hyperspectral technology, multispectral technology can obtain several discrete spectral data from the test sample to characterize a certain characteristic parameter of the object of interest . The Vis region (380–780 nm) contains spectral information related to color characteristics. The NIR spectrum is mainly in the range of 780–2500 nm, while the MIR spectrum is in the range of 2500–25,000 nm. The far infrared (FIR) spectrum is in the farther spectral range (25,000–300,000 nm). NIR and MIR spectra have higher energy than FIR spectra. These two spectra are more suitable for analyzing fingerprint information related to chemical components . NIR spectrum is used to analyze the stretching and bending of chemical bonds, including O–H, S–H, N–H, and C–H . MIR spectrum is mainly related to basic vibration and rotational vibration structure , which contains characteristic information related to chemical functional groups .
The spectral parameters of the detected object and its physical or chemical properties can be correlated by machine learning. Machine learning uses mathematical algorithms to explore the rules that exist in big data to assist decision-making, involving unsupervised learning and supervised learning. More information about machine learning can be found elsewhere . Based on the establishment of the calibration model, the parameter values of unknown samples can be predicted. Machine learning methods, such as principal component regression (PCR), hierarchical cluster analysis (HCA), support vector machine (SVM), partial least squares regression (PLSR), multiple linear regression (MLR), locally weighted partial least squares regression (LWPLSR), artificial neural network (ANN), and least square support vector machine (LS-SVM), have been widely used in food analysis . Feature variable selection based on genetic algorithm (GA)  , competitive adaptive reweighted sampling (CARS) , first-derivative and mean centering iteration algorithm (FMCIA) , regression coefficient (RC), successive projection algorithm (SPA) , and principal components analysis (PCA) [ 58] help to eliminate the feature overlap of continuous spectral information, which is conducive to the development of more robust and simplified machine learning models . A high-performance model requires higher determination coefficients for cross-validation ( R 2CV ) and prediction ( R 2P ), correlation coefficients for prediction ( R P ), and lower root mean square errors for cross-validation (RMSECV) and prediction (RMSEP). Figure 1 shows the schematic of a general framework for tuber quality determination based on imaging spectroscopy. Detailed applications of the technology are given in the following section.
Figure 1. A typical schematic of imaging spectroscopy for tuber quality determinations.
The concept of agricultural intelligent sensing has attracted widespread attention. In the past few years, many scientists have studied the feasibility of imaging spectroscopy in rapid quality assessments of potato and sweet potato tubers. This section provides an overview of developments and applications of this technology as listed in Table 1 .
|Quality Parameter||Sample Type||Spectral Region||Optimal Model||Accuracy||Reference|
|Freshness, Cultivar||Potato||Vis-NIR||PLSR||0.98 for freshness, 93% for cultivar
|Sprouting activity||Potato||Vis-NIR||KNN, PLSDA||90%|||
|Zebra chip disease||Potato||Vis-NIR||PLSDA||92%|||
|Starch||Potato||Vis-NIR||SVR||RP = 0.93|||
|Starch||Potato||Vis-NIR||PLSR||RP = 0.94|||
|Potato||Vis-NIR||LSSVM||R2P = 0.84 for color, R2P = 0.77 for moisture content|||
|TA, moisture content||Sweet
|Vis-NIR||PLSR||R2P = 0.87 for TA, R2P = 0.86 for moisture content|||
|NIR||PLSR||R2P = 0.95|||
|SSC||Sweet potato||Vis-NIR||SVR||R2P = 0.86|||
|Sulfite dioxide residue||Potato||NIR||SVM||95%|||
|Glucose, sucrose||Potato||Vis-NIR||PLSR||RP = 0.90 glucose, RP = 0.82 for sucrose|||
|Hardness, resilience, springiness, cohesiveness, gumminess, chewiness||Potato, sweet
|MIR||LWPLSR||RP = 0.80, 0.88, 0.58, 0.57, 0.73 and 0.69 for hardness,
resilience, springiness, cohesiveness, gumminess and chewiness
|Moisture content||Potato||Vis-NIR||PLSR||R2P = 0.98 for moisture content|||
|Dry matter, starch||Potato, sweet
|NIR||MLR, PLSR||R2P = 0.96 for dry matter, RP2 = 0.96 for starch|||
|Vis-NIR||MLR||R2P = 0.87|||
|Moisture content, FWC||Sweet
|Vis-NIR||MLR||R2P = 0.98 for
moisture content, R2P = 0.93 for FWC
|Moisture content, color||Potato||Vis-NIR||PLSR||R2P = 0.99 for
moisture content, R2P = 0.99 for colour
|NIR||TBPANN||R2P = 0.97 for VTC, R2P = 0.98 for TCD|||
|NIR||LWPCR||R2P = 0.97 for WBC, R2P = 0.98 for SG|||
|Moisture content||Potato, sweet potato||NIR||PLSR||R2P = 0.94|||
|Potato||Vis-NIR||PLSR||R2P = 0.70 for starch, R2P = 0.51 for
glucose, R2P = 0.70 for asparagine
|Leaf counts, glucose, sucrose, soluble
|Potato||Vis-NIR||PLSR||RP = 0.95 for leaf counts, RP = 0.95 for glucose, RP = 0.55 for soluble solids, RP = 0.95 for sucrose, RP = 0.61 for specific
|Cooking time||Potato||Vis-NIR||PLSDA||R2P = 0.96|||
|Moisture, fat content, color properties, maximum force||Taro chip||NIR||PLSR||R2P = 0.85–0.97|||
LWPLSR—locally weighted partial least squares regression; PLSR—partial least square regression; KNN—k-Nearest Neighbors; LSSVM—least squares support vector machine; PLS-SVM—partial least squares support vector machine; GLCM—gray level co-occurrence matrix; SSC—soluble solid content; SVR—support vector regression; PLSDA—partial least square discriminant analysis; VTC—volatility of tuber compositions; TCD—tuber cooking degree; SMTSM—supervised multiple threshold segmentation model; SVM—support vector machines; MLR—multiple linear regression; BPNN—back-propagation neural network; TBPANN—three-layer back propagation artificial neural network; TA—Total anthocyanin; FWC—freezable water content; RP—correlation coefficient for prediction; R2P—coefficient of determination for prediction.