1. Introduction
To date, the campaign to end hunger has received great attention from nations. With the growth in human population and the ongoing rise in food demand, hunger is a pertinent subject that warrants concern. Furthermore, the global population is expected to continue rising and eventually reach nine billion people [
1]. Moreover, a meta-analysis investigated by Dijk et al. [
2] demonstrated that there will be a rise in food consumption through 2050.
Although crop yields are projected to rise, concerns regarding food security and food safety have been spreading around the globe [
3]. Coronavirus disease (COVID-19) and the war in Ukraine were the pressing problems in 2020–2022, leading to a loss of food security [
4]. The challenge of food security appeared when people tried to deal with social distancing, which forced citizens to stay at home. In this case, people tended to store up their food shelves in huge amounts. Additionally, the fear of viral infection could heighten the concern about food unavailability [
5] caused by inadequate access to agricultural activities. Similarly, the war between Russia and Ukraine significantly affects agricultural processes, from plant growth to food markets [
6]. Further, food safety deals with the safety of the food production line, ranging from on-farm to off-farm practices [
7,
8]. Hazards commonly occur due to chemical contamination [
9], physical contamination [
10,
11], food adulteration [
12,
13], genetic modification [
14], and other processes. Chemical contamination and food adulteration in agricultural products are present as the side effects of food unavailability. Hence, an analytical method, such as spectroscopy, can be used. Spectroscopy imaging techniques are emerging as a powerful tool to directly and rapidly assess food quality [
15], although a model needs to be built using machine learning algorithms before an assessment is conducted. We suggest the reviews by Gowen et al. [
16] and Qin et al. [
15] for detailed applications of spectroscopy imaging techniques in food safety control assessment. A description of current spectroscopy imaging techniques is available in
Section 3.
Plants are considered one of the important food resources for humans and stocks. The consumption of plants can offer various advantages for the human body. This is due to the presence of natural materials in plants, such as vitamins, minerals, and other micro- and macronutrients. However, the quality of food products derived from plants depends on the unique traits during their growth period. Abiotic stresses, such as high salinity [
17,
18], extreme temperature [
19,
20], and drought [
21,
22], play a role in causing yield loss [
23,
24]. Nevertheless, these factors are present naturally during crop growth. Subsequently, numerous symptoms and alterations may occur throughout the growth time and, hence, plant growth can be disturbed. Plant biotechnology has been widely developed by researchers to strengthen various crops’ resistance to stresses, such as Chinese cabbage [
25] and potato [
26]. The term hybridization refers to combining high-resistance and high-yield genes with a target gene [
27,
28]. In this process, a technician is needed to isolate the desired gene and subsequently clone it to the target plant. Nevertheless, some Asian countries still face difficulties when confronting abiotic stresses that threaten crops, for example, heat stress during summer in countries such as Indonesia. In this case, due to the consequences of heat stress, plants are likely to show different responses (in physiological and biochemical aspects), as previously summarized by Hasanuzzaman et al. [
29]. Under an elevating ambient temperature, the stomata tend to close and produce a lot of carbon dioxide, and the photosynthetic rate will fall gradually [
30,
31]. For an in-depth understanding, readers may refer to [
29,
30]. Therefore, a preventive step—crop growth monitoring—should be conducted [
32].
At present, the use of spectroscopy imaging technologies, hyper- and multispectral imaging (HSI and MSI), is more favored than the manual technique. The reason is due to their ease of use while being rapid and nondestructive. Nondestructive refers to the ability to reuse samples without damaging them during analysis. In contrast, chromatography tools, namely gas chromatography [
33] and liquid chromatography [
33,
34], require a lot of sample preparation and are time-consuming [
35]. In such a way, destructive steps need to be performed, including drying, grinding, and extracting before injecting into the instrument [
36]. In agriculture, research on HSI and MSI has been extensively conducted. Some of the publications related to the application of HSI and MSI are listed in
Table 1.
Besides the above cited publications, HSI and MSI are also applicative for monitoring plant growth status. Plant phenotyping is still believed to be time-consuming and labor-intensive work [
49]. The use of HSI and MSI can provide the physical and chemical features of plants [
50] based on their reflectance/absorbance profiles. In this study, we present a systematic literature review of the application of HSI and MSI to monitor plant growth status according to the natural problems encountered by plants. We also provide (i) general information on plant traits; (ii) the interaction of light and plants; (iii) data analyses; and (iv) recent trends of HSI and MSI for monitoring plant growth status. For the review, we collected published papers from Google Scholar with the keywords, “plant abiotic/biotic stress detection using hyperspectral and multispectral imaging”, in 2010–2023.
2. Plant–Light Interaction
To begin, an HSI and MSI system involves analyzing light–plant interaction. Understanding this interaction can help choose the proper sensor [
51]. Light (electromagnetic radiation, EMR) and plant interaction depends on light frequencies [
50]. For healthy leaves, photosynthetic pigments, e.g., chlorophyll, carotenoids, anthocyanin, and xanthophyll, can be observed in the VIS region (400–700 nm), while the NIR region (700–1100 nm) predominantly corresponds to dry matter, and the SWIR region, within 1100–2500 nm, is mainly attributed to water [
50]. Sarić et al. [
52] stated that nitrogen and water can also be observed in the NIR and SWIR regions, respectively. As shown in
Figure 1, the visible region has a lower reflection value. This is caused by the highest absorption at approximately 490 and 690 nm by chlorophyll [
53]. A small peak reflection is also observed in the visible range (~550 nm), which contributes to the green color [
54]. In the NIR bands, a high–sharp reflectance trend occurs, indicating a red edge, in which light absorption by leaf pigments no longer exists [
54,
55]. In vegetables, the response of O–H bands is found at 1190, 1450, and 1940 nm [
10]. The absorption of the first overtone of carbon–hydrogen bond could be detected between 1700 and 1800 nm [
56,
57]. According to Tunny et al. [
10], the response of C–H might be due to the presence of cellulose.
When light passes through a leaf, there will be some interactions between the plant and light. The interactions include (i) reflection, (ii) scattering, (iii) absorption, and (iv) transmission [
50,
52]. Reflection occurs when light is bounced back after entering the leaf’s surface. When light is reflected by a different angle caused by a different shape of the leaf structure, it is called scattering. In terms of absorption (mathematically described as
), electrons accept photons’ whole energy and are displaced to a higher configuration in the form of thermal energy. Otherwise, transmission occurs when atoms pick up the wave, vibrate briefly, transmit the vibration through the body of the leaf, and emit the wave as light at the other end.
Figure 2 illustrates how light interacts with a leaf surface.
3. Hyper- and Multispectral Imaging
It is nearly impossible to monitor a crop’s status through the naked eye. The human eye is limited in its ability to measure or quantify due to fatigue. Consequently, bias results may occur. Moreover, manual laboratory analyses require highly skilled technicians and are time-consuming. Thus, the utilization of nondestructive technologies, such as HSI and MSI, is a promising tool. It offers numerous beneficial aspects, such as being nondestructive and noninvasive with rapid inspection [
15]. Due to its advantages, the use of HSI and MSI has become a necessity. HSI and MSI offer both spatial and spectral information [
40,
52,
58,
59,
60]. For a review of the variability in illumination and camera type, readers can refer to [
49]. To ease readers into the topic, we provide an illustration of HSI/MSI during data collection in
Figure 3.
In terms of application, HSI is used for fundamental research and MSI is more applicable in field works [
15]. In addition, HSI provides high resolution but is time-consuming compared to MSI. To apply MSI, HSI should be performed first to obtain the optimum wavebands [
15]. According to Qin et al. [
15], there are three scanning methods to acquire 3D hyperspectral cubes, namely (i) point scan, (ii) line scan, and (iii) area scan, as presented in
Figure 4. Point scan obtains single-point spatial information by collecting each pixel. This method is time-consuming since it captures two spatial dimensions (x, y). Furthermore, point scan does not cover a wide area. Secondly, line scan is an extension of the former scanning method and captures slit spatial information. This method is suitable for moving samples, i.e., sortation. Additionally, the area scan method instantly captures 2D single-band grayscale images with full spatial information. Moreover, if the target object has undesirable movement, the HSI method is acceptable. In contrast, line and area scans are more suitable for MSI than point scan [
15]. Line scan only works for the selected tracks during scanning time. Finally, area scan is fast and collects images at various wavelengths at once.
3.1. Spectral Correction
During data acquisition, some external disturbances, such as dark noises, unwanted light intensity, and environmental factors, might occur [
32,
40,
42,
61,
62]. Commonly, before performing the measurement, the white and black reference data should be measured [
43]. The white reference is obtained using a white Teflon with ±99% reflectance, and the dark reference (±0% reflectance) can be obtained by covering the lens. Finally, the corrected image’s spectral feature is mathematically quantified using Equation (1):
where
, and
refer to the raw scanned object, the white reference, and the dark reference reflectance, respectively.
The next step after spectral calibration is the selection of regions of interest (ROIs). During data acquisition, the captured images not only show the target plant as a research object but also the background. Therefore, it is necessary to eliminate the background pixels from the calibrated images [
32].
3.2. Preprocessing
Due to the advantages of HSI and MSI, which have been mentioned above, the performance of both instruments is primarily affected by their sensitivity, calibration, physical mechanism, and the surroundings [
63]. Spike is a common phenomenon that can be found within the spectrum body. It presents as extreme rise and fall waves. The complexity of plant geometrics, i.e., spherical, elliptic, wavy, or irregular shape, has made it necessary to correct images with preprocessing, which is inspired by chemometrics [
50]. Firstly, the average of spectral data can be employed since averaging can reduce noises and correct the illumination effect [
49]. In addition, spectral averaging is the simplest preprocessing method. This method has been successfully applied to predict the plant physiology and leaf water content of maize [
64]; the water content, micro-, and macronutrients in maize and soybean [
65]; and the water and nitrogen contents in wheat [
66]. Secondly, other preprocessing techniques, namely multiplicative scatter correction (MSC) and standard normal variate (SNV), can be established to reduce variability due to scattering or baseline shift. Thirdly, smoothing by applying a Savitzky–Golay filter can also be established.
3.3. Chemometric Models
Hyper- and multispectral imaging generates reflectance values from each waveband. Nevertheless, it is indirectly applicable. Chemometrics is an analytical approach originally derived from statistical and mathematical concepts [
67]. This analytical step is favorably employed by researchers to develop prediction models based on reflectance values. Chemometric techniques have been intensively carried out to predict the desired objectives within agri-food production lines, e.g., chemical content, food sensory, and plant monitoring status. Moreover, chemometrics also promotes simple presentation by reducing the dimension and complexity of large datasets, classifies samples, and enables data modeling, with robust accuracy [
68].
Principal component analysis (PCA) has been successfully applied by many researchers as a tool to analyze spectral data. Since spectral data consist of numerical reflectance values for each waveband, dimensional reduction can be performed to display the information clearly in a two- or three-dimensional plot (called PCs). Although PCA simply presents variances in a scatter plot, it does not lose the information. PCA is also an acceptable method for observing outliers and constructing models. For instance, PCA clearly showed that adulterant materials had different reflectance values compared to fresh-cut vegetables [
10]. Fresh-cut vegetables had a similar reflectance value, clustered together as a group, and were separated from adulterant materials. In another study, PCA was used to illustrate the contrast maps of chemical contents in different treated plants. In contrast, Lee et al. [
69] reported that PCA had a lower coefficient of determination (R
2 = 0.69) in predicting total volatile basic nitrogen in pork. This was caused by the fact that PCA only obtained a single waveband (460 nm) to build the TVB-N prediction model.
At present, due to the high dimensionality of spectral data acquired using HSI/MSI, many researchers apply wavelength selection before performing chemometric analysis. The high dimensionality of HSI/MSI spectral data can be further reduced to a low dimension through feature extraction and feature selection. Feature extraction utilizes the original data, while feature selection requires class marking to choose the representative wavelength. However, wavelength extraction can remove intrinsic features [
70].
In optical spectroscopy, outliers may occur during the data acquisition process. This phenomenon may be present due to the environment and operator errors [
71,
72]. Negatively, outlier samples can decrease the accuracy of the built predictive model. Therefore, outliers should be deselected before employing chemometrics. Principal component analysis (PCA) can detect such outliers. In a study by Pandey et al. [
65], two-dimensional PC plots were used to figure out outliers. Since PCA can simply display clusters based on their similarity, sample outliers can easily be detected. Similarly, the effectiveness of PCA in tracing outliers had been proven previously [
73].
To date, PLS-R has been intensively developed to establish predictive models of targeted components based on spectral features [
74]. A latent variable—typically abbreviated as LV—is considered an important parameter to investigate the linear relatedness of both spectrum matrices and reference values. High accuracy of PLS-R depends on the number of chosen LV values, which is determined based on the lowest point of cross-validation root mean squared error [
75]. Nevertheless, the accuracy of PLS-R relates to the validity of the reference matrices obtained using destructive methods. Generally, PLS-R itself is considered to be a linear regression and is mathematically described as shown below (Equation (3)). In Equation (2),
is the reference value,
is the spectrum matrices (
),
is the PLS-R coefficient, and
represents the error:
Similar to PCA, PLS-R has the ability to reduce the high dimensionality of data matrices [
67]. Practically, PLS-R has the ability to predict plant traits [
76,
77], food chemical contents [
78], and other characteristics. In addition, various chemometric techniques, such as partial least squared discriminant analysis (PLS-DA), support vector machine regression (SVM-R), and least squared support vector machine (LS-SVM), can be applied.
Despite the development of the aforementioned chemometric techniques, at present, deep learning (DL) has been favorably developed and used for HSI/MSI data analysis. The concept of DL architecture imitates the principle of the human brain’s visual cortex. DL offers some advantages: (a) it discovers properties on its own; (b) it saves the process by reducing the need for computation; (c) it creates features manually; and (d) it uses large, annotated images that researchers already have access to [
79]. However, supervised DL needs a large-scale dataset. In addition, DL-based spectroscopy is rarely used to quantify plant phenotypic features. Nevertheless, this method can automatically extract raw spectra and improve the performance of prediction models [
80]. Hence, it is still possible to combine DL with spectroscopy during plant phenotyping. In
Table 2, we summarize recent applications of DL for monitoring plant growth status. The detailed information of the use of deep learning in combination with spectroscopy can be found in Wang et al. [
81]. Wang et al. [
81] provides details of major types of DL (convolutional neural network/CNN, fully convolutional network/FCN, tensor learning model/TL, deep belief network/DBN, stacked auto-encoder/SAE, recurrent neural network/RNN, semi-supervised learning, generative adversarial network/GAN, and active learning model/AL) and challenges in the agricultural field.
3.4. Model Validation
Most researchers have commonly used a few mathematical formulas to evaluate the performance of their employed models. The importance is to establish that the proposed models show good satisfaction, as indicated by errors and/or linearity relationship. Based on our review, researchers have often conducted various chemometric methods in comparative study. For instance, different mask segmentations were performed to estimate chlorophyll contents in wheat using MSI combined with PLSR [
87]; various models were applied to predict alfalfa yield using MSI [
88]; and VI models were established to assess nitrogen status in winter oilseed rape using MSI [
89]. The equations that are most commonly used are R-squared (
) and root mean squared error (
), which are expressed in Equations (3) and (4), respectively:
where
corresponds to the total dataset;
and
refer to the estimated and referenced values at the
th element, respectively; and
is the average value of the reference data (destructive value). The value of
ranges between
and 1 [
90]. A model with the closest value to 1 has excellent goodness-of-fit performance. However, we do not recommend evaluating models based on
itself. This is due to the fact that
represents the connection between the
-axis (reference) and
-axis (predicted). Additionally,
can be calculated to make the decision:
In contrast, the decision based on is in selecting a smaller value. Based on our literature review, a higher value is caused by a larger gap between the predicted and observed data. Thus, a higher indicates a poorly built model and, thus, it should not be chosen.
5. Discussion
To date, the use of spectroscopy imaging techniques has a wide range in the agricultural field. In general, we found that hyperspectral imaging is more commonly applied than multispectral imaging. The reason is perhaps that HSI can provide high-resolution data compared to MSI. In food safety and quality assessment, Qin et al. [
15] had come to similar conclusions. Nevertheless, an MSI system can perform snapshot-based imaging and be more practical, as shown in [
32]. Based on our literature review, it can be seen that the use of spectroscopy imaging techniques ranges from seed viability to crop quality evaluation. Additionally, the effects of abiotic stresses can be easily detected using HSI and MSI, which are normally equipped with several chemometrics approaches. In contrast, investigation on heat stress is still limited. This is prominently caused by heat stress that only appears in several countries, particularly tropical countries. However, in fact, HSI and MSI can distinguish biological stresses and quantify nutrient distribution in leaf. Regarding food safety, heavy metal residues within plant matrixes, such as lead (Pb) and cadmium (Cd), have also been observed using these techniques. Such investigations have been carried out due to the high toxicity of heavy metal residues to the human body. Nevertheless, these studies showed low performance and accuracy. This is perhaps due to the fact that not every heavy metal exhibits a distinct spectral response [
123].
As shown in the tables above, various chemometric models have been built using HSI/MSI data. In addition, these models were proposed to predict, classify, and discriminate the spectrum and/or image features acquired by using HSI/MSI. According to our literature review, regression techniques are widely utilized. In
Section 3, we offer a brief introduction about PLS-R. Nevertheless, other regression techniques mentioned above are also able to predict chemical compounds in plants with good satisfactory results. Moreover, model calculation is not only limited to prediction but also includes classification and discrimination. Classification and discrimination models can be developed using supervised and unsupervised methods. PLS-DA is a linear supervised technique in which the basic calculation relies on the PLS technique. Differently, the
matrix contains classes, such as “0” and “1” in [
10] or “2” and “3” in Kim et al. [
32]. The other variation for classification, that is, linear construction, is PCA. PCA is a linear unsupervised technique. As mentioned in
Section 3, users are not required to prepare a response matrix. The information extracted from an example of PCA is illustrated in
Figure 5.
The results obtained by using PCA can be interpreted through the score values. The difference among samples can easily be recognized by separated clusters. Another piece of information gained from PCA is loading. It is useful to describe information about wavelength [
10]. This method relies on spectral features—reflectance or
. Since HSI/MSI provides both spectral and spatial features, we can utilize PLS regression (
) and PC scores to generate a chemical distribution image by multiplying each pixel from the calibrated image. The process of obtaining a PLS-based chemical distribution image can be described using Equation (5) [
40]:
where
denotes the image corrected pixel at the specific band,
corresponds with the beta coefficient for each wavelength, and
is the coefficient value. Application of Equation (5) is presented in
Figure 5 (left).
Furthermore, wavelength selection is also periodically used by researchers before applying chemometric models to achieve higher performance [
125,
126,
127,
128,
129,
130,
131,
132,
133]. In recent years, deep learning (DL) algorithms have also been more likely to be utilized, as mentioned above. In addition, DL could be argued to be a promising method [
49]. Typically, DL consists of two methods for object detection: region-based one-stage method and region proposal-based two-stage method [
134]. Region-based convolutional neural networks (R-CNNs) take a longer time to perform object detection since they are considered to be a two-stage method. Another DL model, “You only look once” (abbreviated as YOLO), is an example of a one-stage method. YOLO takes a considerably shorter time during object detection. Among the various versions of YOLO, YOLOv4 runs faster. As shown in
Table 2 above, various DL models can be used. The task of deep learning algorithms is not only object classification but also prediction. Various tasks in content prediction were summarized in Wang et al. [
81]. In our study, we also revealed similar findings for content prediction as in Sabzi et al. [
82]. However, the challenge of using a DL algorithm is the requirement of huge datasets for the training test to increase result accuracy. In addition, according to Krishnaswami et al. [
84], application of DL is more challenging due to the utilization of mostly remote sensing data.
We also found that model evaluation is not limited to analyzing R-squared and RMSE but also may vary. For example, relative percentage difference (RPD) is also potentially used. In contrast with RMSE, a lower value of RPD indicates a model is worse [
121]. RPD is calculated by dividing standard deviation (SD) with RMSE. Other chemometric models that we identified, such as LS-SVM, use receiver operating characteristic (ROC) curve to test the performance of the models in discriminating the samples. An ROC helps us decide whether a model is applicable or not. A model has great accuracy if the value of the area under the curve (AUC) is close to or equal to 1 [
135]. Similarly, Dharmawan et al. [
44] used an ROC curve to study the performance of PCA-MLP during the authentication of arabica coffee.
The tables presented above show that vegetation indices (VIs) are also used to determine crop stresses. Vegetation indices are based on mathematical formulas that are derived from spectral information. In application, VIs have been used to compare with a built model or for masking an image. Each VI describes a different purpose. For example, normalized difference vegetation index (NDVI) is a vegetation index that can be used to assess the impacts of drought on vegetation. NDVI values above 0.6 generally represent dense vegetation, while values in the range from 0.2 to 0.5 commonly appear in aging plants, shrubs, and meadows [
136]. The equations for different VIs are presented in
Table 8.
6. Conclusions
Hyperspectral (HS) imaging and multispectral (MS) imaging are promising measurement techniques to monitor the growth status of plants. Their feasibility, nondestructive nature, and rapid inspections make HSI/MSI suitable for large-scale plant factories. Additionally, monitoring using a manual technique, i.e., cutting, can damage the plant body, causing physical or mechanical stress, and resulting in a lower yield production. Furthermore, HSI/MSI can be used to observe plant features that cannot be detected with naked human eye, such as chemical compounds. Most of the research studies emphasized that crops were affected by the controlled surrounding conditions, such as abiotic and biotic stresses. Subsequently, HSI/MSI were utilized to monitor the plant growth status.
The spectral data acquired using HSI/MSI were then coupled with chemometric analyses. However, to remove noises from the observation, several preprocessing methods could be performed. Chemometric techniques were used for predicting plant chemical composition and classification, among many other functions. Most researchers used VIs to distinguish plants’ response in a given controlled environment. In recent years, deep learning analysis methods have been widely used to assist HSI/MSI. Further, the input variables (spectral information) are then convoluted and pooled to obtain the results. Moreover, an obtained image can also be processed (image processing) to identify plant symptoms during the growth period.
The applications of HSI and MSI for monitoring crops’ growth status may also vary depending on the aim. Firstly, applications, such as detection of heavy metals in plant tissues as a side effect of pollution, can also be conducted to safeguard human health and food safety. Secondly, the effects of biotic infection during growth season can also be observed. Various types of plant stresses can impact on the subsequent agri-food production lines, such as storage time. However, there is still limitation in research on the use of HSI/MSI for detecting changes in specific chemical compounds under a given stress treatment. Furthermore, a spectral selection algorithm needs to be developed and employed to choose the most representative waveband and increase model robustness and performance.