Snapshot-Based Multispectral Imaging for Heat Stress Detection in Southern-Type Garlic

Abstract

This study aims to develop a model for detecting heat stress in southern-type garlic using a multispectral snapshot camera. Raw snapshot images were obtained from garlic cloves during the garlic bulb enlargement period, capturing the visible (Vis) and near-infrared (NIR) regions. Image preprocessing was applied to obtain a 38-wavelength spectrum by combining a 16-wavelength image in the Vis region and a 22-wavelength image in the NIR region. These spectral data were then utilized to develop models, including PLS-DA, LS-SVM, DNN, and recurrence plots-based CNN (RP-CNN). On average, the LS-SVM model demonstrated the best performance in detecting heat stress during the garlic bulb enlargement period. This is attributed to the nonlinear nature of the spectral differences between groups caused by abiotic stress in garlic. The LS-SVM model is particularly effective at capturing such nonlinear relationships. Among the model images, LS-SVM yielded the best performance, followed by RP-CNN, DNN, and PLS-DA. Therefore, this study confirms the potential of snapshot-based multispectral imaging for measuring changes in garlic crops induced by high-temperature stress.

1. Introduction

Allium sativum, widely cultivated for cooking and medicinal purposes, is an important crop known for its cholesterol-lowering, anti-cancer, and cardiovascular disease-fighting properties [1,2,3,4]. However, garlic plants are vulnerable to various environmental stresses, including high temperatures, which can significantly impact their growth and yield [5,6]. Temperature is the most critical factor affecting garlic growth, with the optimal temperature range for southern-type garlic being 18–20 °C and with growth ceasing at 25 °C or above [6]. Heat stress adversely affects the growth, development, and physiological processes of garlic plants, leading to reduced yield, compromised quality, and economic losses for farmers. When exposed to high temperatures, garlic plants experience disrupted photosynthesis, increased respiration rates, altered water uptake, and impaired cellular functions. These physiological disruptions can result in wilting, leaf chlorosis, premature senescence, and decreased bulb size and weight. Garlic diminishes the photosynthetic capacity and peak photochemical effectiveness of garlic when exposed to temperatures of 25 °C or above [7]. It amplifies the bulb count under high temperatures, yet diminishes bulb weight, leading to compromised commercial viability [8].

Climate change projections indicate a gradual increase in temperatures due to global warming [7,8,9]. High-temperature stress, characterized by temperatures exceeding the optimal range for plant growth, can lead to physiological and biochemical changes in plants, ultimately affecting overall productivity [10,11]. Studies have shown that rising temperatures caused a decrease in maize and wheat yields by 3.8% and 5.5%, respectively in 2011 [12]. In 2013, significant reductions in potato productivity were observed at temperatures above the optimum [13]. Drought and high temperatures also resulted in a 9–10% decline in grain production between 1964 and 2007 [14]. Furthermore, a 2017 study projected that a 1 °C increase in global mean temperature would lead to reductions in wheat, rice, maize, and soybean production by 6%, 3.2%, 7.4%, and 3.1%, respectively, highlighting the vulnerability of crops to climate change [15]. Consequently, heat stress caused by elevated temperatures adversely affects crop growth, necessitating research to mitigate its detrimental effects.

Several studies have investigated the effectiveness of various nutrients, such as zinc, selenium, and boron, in alleviating heat stress [16,17,18]. Heat stress generates reactive oxygen species (ROS) that interfere with crop growth, but the application of boron can enhance the antioxidative activity of crops, minimizing the damage caused by ROS [16,17,18]. Superoxide dismutase (SOD), an antioxidant enzyme, plays a crucial role in mitigating ROS damage, and zinc influences SOD activity, making sufficient zinc intake beneficial for reducing the effects of heat stress [16,19]. Selenium can enhance SOD activity and reduce ROS, protecting crops from the effects of heat stress [16,20,21,22,23].

Currently, precision agriculture is gaining momentum as an economically feasible and environmentally friendly approach. By utilizing crop-related data obtained from various sensors, precision agriculture enables the application of the required amount of inputs, such as fertilizers, thereby minimizing environmental impact and cost [24,25]. Precision farming involves four key steps: (1) data acquisition on crops and their environment, (2) determining crop conditions and identifying suitable solutions based on acquired data, (3) prescribing appropriate actions for crops based on the identified solutions, and (4) analyzing the outcomes of the prescribed actions.

Detecting high-temperature stress not only allows for proactive measures but also reduces labor and costs. Hence, studies have been conducted to detect high-temperature stress in crops to minimize its impact. For instance, experiments were carried out in 2021 to detect high-temperature stress in ginseng, while in 2022, fluorescent hyperspectral imaging was employed to detect high-temperature stress in ginseng and strawberries [26,27,28]. Considering the projected rise in temperature due to global warming, understanding the effect of high-temperature stress on garlic crop growth and developing methods to detect and mitigate it are crucial.

Traditional methods for assessing plant stress rely on time-consuming, subjective visual observations or invasive techniques that often require destructive sampling. In recent years, there has been increasing interest in non-invasive, real-time monitoring techniques that provide accurate and objective assessments of plant stress [29,30]. Hyperspectral and multispectral imaging are non-destructive methods used to obtain images at numerous wavelengths without damaging the crop [26,27,28]. In this study, we utilize multispectral analysis and a multispectral snapshot camera to detect high-temperature stress in southern-type garlic. The snapshot-based approach, along with line-scanning and snap-scanning methods, enables the acquisition of hyperspectral and multispectral images using a single camera. The line-scanning method involves scanning as the camera or the object moves, while the snap-scanning method acquires data by scanning inside the camera without moving it or the object. The snapshot method captures data of multiple wavelengths simultaneously, eliminating the need for scanning time and enabling faster data acquisition, albeit with a relatively limited range of obtainable wavelengths compared to the scanning methods [31,32,33]. Among these techniques, snapshot-based hyperspectral imaging and multispectral imaging have emerged as promising tools for non-destructive plant stress diagnosis, allowing for the simultaneous capture and analysis of visible and near-infrared spectra [31,32,33].

Hyperspectral imaging and multispectral imaging combine spectroscopy principles with digital imaging, facilitating the acquisition of spectral information across a wide range of wavelengths. By analyzing the unique spectral signatures of plant tissues, these imaging techniques enable the detection of subtle changes associated with stress-induced physiological responses. They have been successfully applied in various agricultural applications, including assessing nutrient deficiencies, disease detection, and stress monitoring in crops [26,27,28,30]. In particular, the VIS/NIR wavelengths were chosen to capture the absorption characteristics of pigments, such as chlorophyll, that are vital for photosynthesis, changes in plant cellular structures, water content, and leaf internal composition and indicative of plant health. These wavelengths are known to be sensitive to changes in pigment concentrations and can provide insights into the plant’s photosynthetic activity and stress-induced alterations. In addition, these wavelengths also allow us to assess stress-related changes in cellular structures, such as leaf thickness and internal scattering, which can affect the spectral response.

In the context of garlic crop management, the application of hyperspectral imaging and multispectral imaging holds great potential for evaluating high-temperature stress-induced physiological changes. By capturing and analyzing the spectral response of garlic plants under different temperature regimes, it becomes possible to identify specific spectral features indicative of stress conditions. These features can serve as indicators or biomarkers for evaluating the severity and extent of high-temperature stress on garlic crop growth.

Despite the potential of hyperspectral imaging and multispectral imaging, limited research has focused on their utilization for evaluating high-temperature stress in garlic crops. Most studies have primarily targeted other crops or different stress conditions. Therefore, comprehensive investigations specifically targeting high-temperature stress and its effects on garlic plants using these imaging techniques are necessary. Such studies can provide valuable insights into the physiological responses of garlic plants to high-temperature stress and contribute to the development of effective management strategies.

Various models, including the vegetation index, partial least-squares discriminant analysis (PLS-DA), least-squares support-vector machines (LS-SVMs), deep neural networks (DNNs), and convolutional neural networks (CNNs), have been employed to detect high-temperature stress in snapshot data. The vegetation index quantifies crop vegetation using visible light and near-infrared rays. PLS-DA and LS-SVM are machine learning methods, while DNN and CNN are deep learning methods widely used in various classification tasks [34,35,36,37,38,39,40,41,42]. In this study, we extract spectral data through image preprocessing to detect high-temperature stress in southern-type garlic, using and evaluating the models of vegetation index, PLS-DA, LS-SVM, DNN, and CNN.

The objective of this study is to assess the impact of high-temperature stress on garlic crop growth using a snapshot-based hyperspectral imaging and multispectral imaging system operating in the visible and near-infrared regions. We will analyze the spectral data acquired from garlic plants subjected to different temperature treatments and evaluate the potential of hyperspectral imaging and multispectral imaging as non-destructive and real-time monitoring tools for high-temperature stress. The findings of this research will enhance our understanding of the physiological changes occurring in garlic plants under high-temperature stress and contribute to the development of strategies for sustainable garlic crop management. The specifics of the study are as follows: (1) establish a snapshot-based visible/near-infrared multispectral imaging system to acquire multispectral images of garlic crops, (2) develop an optimal model capable of discriminating garlic crop growth under high-temperature stress using the acquired multispectral images, and (3) analyze the spatial changes of garlic crops caused by high temperature using the developed optimal model images.

2. Materials and Methods

2.1. Garlic

The garlic cultivar used in the experiment was Namdo, a warm-season type of garlic. The test site was conducted at the extreme weather growth chamber (Modified Controlled Environment Extreme Weather Simulator, EGC Co., Chardon, OH, USA) in the National Institute of Horticultural and Herbal Science, Republic of Korea (35.16′ N, 127.02′ E, and 32 m altitude). The garlic crop was sown on 9 September 2021 and grown until 18 May 2022. The pots used in the experiment were 120 cm × 80 cm × 90 cm in width × length × height, and 48 garlic crops were planted in each pot at a planting interval of 10 cm × 15 cm. The high-temperature treatment period was performed for 14 days (2022.4.14~27). The temperature treatment was set at 20 °C during the day and 8 °C at night for the control group, 28 °C during the day and 16 °C at night for the heat level 1, and 36 °C during the day and 24 °C at night for the heat level 2 (12 h/12 h, day/night). The eight pots for the control group and the heat level 1 group and four pots for the heat level 2 group were subjected to temperature treatment for 14 days in the extreme weather growth chamber (

Table 1. Setting temperature.

	Group	Temperature Setting (Day/Night)
Garlic bulb enlargement period	Control group	25 °C/11 °C
	Heat level 1	30 °C/16 °C
	Heat level 2	35 °C/21 °C

).

2.2. Snapshot Camera System

The snapshot camera used in this study comprises a dual-camera system, wherein one camera facilitates the measurement of wavelengths ranging from 462 nm to 623 nm, while the other camera enables the measurement of wavelengths spanning from 631 nm to 870 nm. Using Cube Creator’s specialized camera program, a total of 41 wavelength images captured by the two cameras was successfully correlated with 38 specific wavelengths. Snapshot images were captured using commercial software (Bayspec Spec Frabber, provided by Bayspec Inc., San Jose, CA, USA). A halogen light source was employed, as it provided a continuous spectrum within the wavelength range of 462 nm to 871 nm. A conceptual diagram of the snapshot camera system is depicted in Figure 1 [43].

2.3. Data Acquisition

The multispectral images obtained through the snapshot camera have a resolution of 2048 pixels wide and 1088 pixels high, containing data for all wavelengths. The snapshot multispectral image in the visible (Vis) region consists of 16 wavelengths arranged in a 4 × 4 format, while the snapshot multispectral image in the near-infrared (NIR) region consists of 25 wavelengths arranged in a 5 × 5 format. Figure 2 illustrates a 4 × 4 pixels component of a raw snapshot image in the Vis region and a 5 × 5 pixels component of a raw snapshot image in the NIR region.

To separate the wavelengths in the raw snapshot multispectral images, wavelength images were extracted. We obtained 16 wavelength images in the Vis region and 25 wavelength images in the NIR region. Subsequently, image preprocessing techniques were applied, including setting a region of interest (ROI) and removing the background. This preprocessing resulted in a spectrum composed of 16 wavelengths ranging from 462 nm to 624 nm in the Vis region and a spectrum composed of 22 wavelengths ranging from 631 nm to 871 nm in the NIR region. Thus, a spectrum consisting of a total of 38 wavelengths ranging from 462 nm to 871 nm was obtained.

For each of the three stages, a total of 800 spectra was obtained from the garlic samples, resulting in 2400 data points on the 7th day and 2400 data points on the 14th day.

2.4. Data Analysis

Various vegetation indices have been used for decades as indicators to assess the growth status of crops. In this study, the NDVI (Normalized Difference Vegetation Index), Red Edge Ratio, and PRI (Photochemical Reflectance Index) vegetation indices were used to compare the results with the PLS-DA, LS-SVM, DNN, and RP-CNN models developed using multispectral imaging. The vegetation index was calculated using Equations (1)–(3) as follows:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\frac{{\mathsf{\rho }}_{\mathrm{n}\mathrm{i}\mathrm{r}}-{\mathsf{\rho }}_{\mathrm{r}\mathrm{e}\mathrm{d}}}{{\mathsf{\rho }}_{\mathrm{n}\mathrm{i}\mathrm{r}}+{\mathsf{\rho }}_{\mathrm{r}\mathrm{e}\mathrm{d}}}$$

(1)

$$\mathrm{R}\mathrm{e}\mathrm{d} \mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e} \mathrm{R}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}=\frac{{\mathsf{\rho }}_{\mathrm{n}\mathrm{i}\mathrm{r}}}{{\mathsf{\rho }}_{\mathrm{r}\mathrm{e}\mathrm{d}}}$$

(2)

$$\mathrm{P}\mathrm{R}\mathrm{I}=\frac{{\mathsf{\rho }}_{\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}-{\mathsf{\rho }}_{\mathrm{r}\mathrm{e}\mathrm{d}}}{{\mathsf{\rho }}_{\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}+{\mathsf{\rho }}_{\mathrm{y}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{o}\mathrm{w}}}$$

(3)

PLS-DA is a widely used multivariate statistical technique for classification tasks. It combines the principles of principal component analysis (PCA) and linear regression. By decomposing the spectral data into latent variables known as components, PLS-DA maximizes the covariance between the predictor variables (spectral data) and the response variable (class labels). The number of components is typically determined using cross-validation or other model selection techniques. The PLS-DA algorithm constructs a discriminant model by fitting a linear regression between the components and the class labels [34,35,36]. LS-SVM, on the other hand, is a variant of support vector machines that employs least squares optimization instead of quadratic programming. LS-SVM is particularly useful for handling nonlinear data. By transforming the spectral data into a higher-dimensional space called the feature space, LS-SVM seeks to find an optimal decision boundary on a hyperplane that maximizes the margin between different classes. The algorithm solves a set of linear equations using the least squares method to determine the hyperplane parameters [37,38,39]. DNN, a deep learning method, has shown significant success in various domains, including spectral data analysis. DNNs consist of multiple hidden layers between the input and output layers, enabling them to learn complex patterns and relationships in the data. In our study, the DNN architecture employed an input layer, hidden layer 1, hidden layer 2, and an output layer, as depicted in Figure 3. The hidden layers utilized techniques such as Leaky ReLU activation and channel merging to improve the network’s training performance. The DNN model was trained using labeled spectral data, adjusting the weights of the neurons through optimization algorithms like the stochastic gradient descent [40]. The former generates two groups, while the latter produces three groups.

RP-CNN represents a specialized variant of convolutional neural networks (CNNs) that incorporates random projections. Prior to inputting the spectral data into the convolutional layers, random projections are applied, reducing dimensionality while preserving important properties. RP (recurrence plot) represents one-dimensional (1D) data as an image, as illustrated in Figure 4. RP is created by determining the two-dimensional (2D) spatial trajectory of the spectral data and then constructing a distance matrix using Equation (4) [44].

$${\mathrm{R}}_{\mathrm{i}\mathrm{j}}=\mathsf{\theta }(\mathsf{ϵ}-‖\overrightarrow{{\mathrm{s}}_{\mathrm{i}}}-\overrightarrow{{\mathrm{s}}_{\mathrm{j}}}‖)$$

(4)

A CNN (convolutional neural network) is a deep learning method that involves convolutional, nonlinear activation, and pooling processes [41,42]. The layer structure of the CNN is depicted in Figure 5, comprising the input layer, hidden layer 1, hidden layer 2, hidden layer 3, and output layer. When spectral data is input to the network, the number of channels increases. Batch Norm, Leaky ReLU, and Max Pooling are applied in hidden layer 1 and hidden layer 2. Subsequently, the channels are merged into one channel and three channels in hidden layer 3. The one-channel output undergoes Sigmoid activation, while the three-channel output goes through softmax activation in the output layer, resulting in individual models and an integrated model, respectively. The former generates two groups, while the latter produces three groups.

All machine learning and deep learning models were utilized to build three-group models by incorporating the control group, heat level 1 group, and heat level 2 group. Additionally, two-group models were created by considering 2 groups among the control group, heat level 1, and heat level 2. The calibration model was developed using 75% of the total data, while the remaining 25% was reserved for model performance verification. The accuracy of the models was evaluated using Equation (5).

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{y}\left(\%\right)=\frac{\mathrm{C}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{l}\mathrm{y} \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{e}\mathrm{d} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}\times 100$$

(5)

3. Results

3.1. Model Results

Table 2. Results of all models on the 7th day the of garlic bulb enlargement period.

		Accuracy (%)
		Control Group vs. Heat Level 1	Control Group vs. Heat Level 2	Heat Level 1 vs. Heat Level 2
7th day	NDVI	63.06	55.69	57.06
	Red Edge	60.38	56.38	52.63
	PRI	47.94	50.13	52.38
14th day	NDVI	56.56	54.94	49.75
	Red Edge	60.56	54.19	64
	PRI	52.88	55.31	54.13

presents the results of the vegetation index model on the 7th and 14th days. On the 7th day, with the exception of the PRI results in the control group and Heat level 1, the accuracy of NDVI, Red Edge Ratio, and PRI models was similar. However, the accuracy of all three models remained below 70%. On the 14th day, the Red Edge Ratio model exhibited higher accuracy compared to the other models in heat level 1 and heat level 2. Nevertheless, the accuracy of all three models remained below 70%.

Table 3. Results of machine learning and deep learning models on the 7th day the of garlic bulb enlargement period.

			Accuracy (%)
			Control Group vs. Heat Level 1	Control Group vs. Heat Level 2	Heat Level 1 vs. Heat Level 2
Three-group models	PLS DA	Calibration	62	76.67	55.83
		Prediction	63.75	80.25	55.75
	LS SVM	Calibration	100	100	100
		Prediction	100	97	59.75
	DNN	Calibration	76.08	73.33	61.42
		Prediction	63.5	63.25	50.25
	RP CNN	Calibration	70.75	60.5	69.42
		Prediction	65	55.75	62.25
Two-group models	PLS DA	Calibration	83.17	76.42	77.17
		Prediction	84.75	79.25	74.75
	LS SVM	Calibration	99.67	100	87.33
		Prediction	98.25	100	81.75
	DNN	Calibration	93.42	75.58	82.08
		Prediction	85.75	69	73
	RP CNN	Calibration	91.17	69.92	77.83
		Prediction	87.5	70.25	72

presents the results of the PLS-DA, LS-SVM, DNN, and RP-CNN models created using three groups and two groups on the 7th day of the garlic bulb enlargement period. When examining the accuracy of the models in the control group and heat level 1, it was found that among the three-group models, the LS-SVM model achieved the highest accuracy, while the other models showed no significant difference. Among the two-group models, the LS-SVM model also had the highest accuracy, with no significant difference observed among the remaining models. The PLS-DA, DNN, and RP-CNN models exhibited improved accuracy in the two-group models compared to the three-group models.

Analyzing the accuracy of the models in the control group and heat level 2, the LS-SVM model displayed the highest accuracy among the three-group models, followed by the PLS-DA, DNN, and RP-CNN models. Among the two-group models, the LS-SVM model achieved the highest accuracy, followed by the PLS-DA model. The accuracy of the DNN model did not significantly differ from that of the RP-CNN model. The RP-CNN model demonstrated improved accuracy in the two-group model compared to the three-group model.

When examining the accuracy of the models in heat levels 1 and 2, it was found that the LS-SVM model was overfitted, while the RP-CNN model achieved the highest accuracy. The accuracy of the PLS-DA and DNN models did not significantly differ. Among the two-group models, the LS-SVM models had the highest accuracy, while the PLS-DA, DNN, and RP-CNN models showed no significant difference in accuracy. The LS-SVM model’s overfitting was resolved in the two-group model, and the PLS-DA, DNN, and RP-CNN models demonstrated improved accuracy compared to the three-group model.

Table 4. Results of machine learning and deep learning models on day14 of the garlic bulb enlargement period.

			Accuracy (%)
			Control Group vs. Heat Level 1	Control Group vs. Heat Level 2	Heat Level 1 vs. Heat Level 2
Three-group models	PLS DA	Calibration	76.42	93.17	82.08
		Prediction	76	92.75	81
	LS SVM	Calibration	99.67	99.58	97.5
		Prediction	98	97.75	89
	DNN	Calibration	84.92	92.08	91
		Prediction	76.25	86	85.75
	RP CNN	Calibration	85.17	87.5	96.67
		Prediction	76.25	85	92.75
Two-group models	PLS DA	Calibration	88.83	100	100
		Prediction	85	100	100
	LS SVM	Calibration	99.92	100	100
		Prediction	96.5	100	100
	DNN	Calibration	88	98.42	99.58
		Prediction	79.25	96	98.75
	RP CNN	Calibration	89.08	99.92	100
		Prediction	79.25	98.75	99.75

presents the results of the PLS-DA, LS-SVM, DNN, and RP-CNN models created using three groups and two groups on day 14 during the garlic bulb enlargement period. When examining the accuracy of the models in the control group and heat level 1, it was found that among the three-group models, the LS-SVM model achieved the highest accuracy, while the other models showed no significant difference. Among the two-group models, the LS-SVM model also had the highest accuracy, with no significant difference observed among the remaining models. The PLS-DA, DNN, and RP-CNN models exhibited improved accuracy in the two-group models compared to the three-group models.

Analyzing the accuracy of the models in the control group and heat level 2, it was found that among the three-group models, the LS-SVM model had the highest accuracy, followed by the PLS-DA model. The accuracy of the DNN and RP-CNN models did not significantly differ. For the two-group models, there was no significant difference in accuracy among all models. The PLS-DA, DNN, and RP-CNN models demonstrated accuracy improvement in the two-group models compared to the three-group models.

When examining the accuracy of the models in heat levels 1 and 2, it was found that among the three-group models, the LS-SVM and RP-CNN models achieved the highest accuracy, while the accuracy did not significantly differ between the PLS-DA and DNN models. For the two-group models, no significant difference in accuracy was found among all models. The PLS-DA, LS-SVM, DNN, and RP-CNN models demonstrated improved accuracy in the two-group models compared to the 3-group models.

Figure 6 is the average of the accuracy of each model of the two-group model and the three-group model. Figure 6 shows the results of the LS SVM model with the optimal results in both two-group and three-group models.

3.2. Model Images

Figure 7 illustrates the model images generated using the image data from day 14 of the garlic bulb enlargement period and the two-group model consisting of the control group and heat level 2. In the PLS-DA model images, it is challenging to discern clear differences between the control group, heat level 1, and heat level 2 with the naked eye. However, the LS-SVM model images exhibit distinct dissimilarities among the three groups compared to the PLS-DA model images. The LS-SVM model’s images show greater separation between the groups due to the nonlinear relationship induced by the stress response to high temperatures in the garlic’s multispectral image data.

Regarding the DNN model’s images, visible distinctions between the control group and heat level 2 can be observed in most regions, except for certain garlic leaves located near the multispectral camera. In the RP-CNN model’s images, we can differentiate between the control group and heat level 2, but distinguishing between the control group and heat level 1 is challenging. This limitation arises from the model’s insufficient depth to discern between the control group and heat level 1 in the RP-CNN architecture.

4. Discussion

It is necessary to detect high-temperature stress in garlic to assess the damage caused by abnormally high temperatures resulting from global warming. Therefore, this study aimed to develop models (PLS, LS-SVM, DNN, RP-CNN) capable of detecting high-temperature stress in garlic and compare their accuracy with known vegetation index models. Table 2 indicates that vegetation index models exhibited lower performance with an accuracy below 70% compared to machine learning and deep learning models. This can be attributed to the smaller number of wavelengths used in the vegetation index models compared to the machine learning and deep learning models.

Table 2 and Table 3 demonstrate that the majority of the 14th day models outperformed the 7th day models. This suggests that the effect of high-temperature stress on garlic crops becomes more prominent over time. The higher accuracy of the LS-SVM model compared to the PLS-DA model in Table 3 and Table 4 can be attributed to the phenomenon of high-temperature stress in garlic crops and the non-linear reflectance values exhibited by garlic crops. Abiotic stress, such as high temperature stress, can induce complex and nonlinear changes in the spectral characteristics of plants. These changes manifest as intricate relationships and interactions between different wavelengths in the spectral data. Linear models, such as PLS-DA, assume a linear relationship between the independent variables (spectral data) and the dependent variable (abiotic stress classes), which may not capture the intricate nonlinear relationships present in the data. The LS-SVM model, on the other hand, is well-suited to capture and exploit the nonlinear relationships in the spectral data. It utilizes the kernel trick, which allows it to implicitly map the data into a high-dimensional feature space, where it can find a linear decision boundary that effectively separates the different abiotic stress classes. This transformation enables the LS-SVM model to handle the nonlinear nature of the spectral differences caused by abiotic stress. The LS-SVM model effectively captures the complex and nonlinear variations in the spectral data, allowing it to better discriminate between different classes of abiotic stress in garlic crops. The model’s flexibility in finding nonlinear decision boundaries provides a significant advantage in accurately distinguishing between healthy garlic plants and those experiencing high temperature stress. As the high-temperature stress persisted for 7 and 14 days, the crops adapted to the high-temperature environment, resulting in non-linear reflectivity for garlic crops in high-temperature environments compared to normal environments. However, the deep learning models, DNN and RP-CNN, exhibited lower accuracy compared to the LS-SVM model. This is likely because the DNN and RP-CNN models failed to optimize the model to capture the non-linear relationships among the three groups: the control group, heat level 1, and heat level 2, unlike the LS-SVM model.

Figure 7 presents the optimal images of garlic crops on the 14th day, obtained by applying the model developed using the control group and heat level 2 data. Unlike the high discrimination accuracy observed when analyzing the spectra of each group, the applied model’s images did not reveal clear differences between the groups in all areas and across all models. There was no significant distinction between the control group and the first stage of high temperature in the images, but noticeable differences were observed between the control group and the second stage of high temperature. In particular, the LS-SVM image demonstrated the most pronounced differences among the control group, heat level 1, and heat level 2. This aligns with the accuracy results obtained from the spectral-based models, indicating relatively better performance for the LS-SVM model.

5. Conclusions

In this study, we developed an optimal model for detecting high-temperature stress in southern-type garlic during the garlic bulb enlargement period using a multispectral snapshot camera. The multispectral snapshot camera is capable of capturing 16 wavelengths in the visible region (462–624 nm) and 25 wavelengths in the near-infrared region (603–871 nm), providing a total of 41 wavelength information. Since the data from the visible and near-infrared wavelengths are mixed in the raw images obtained by the multispectral snapshot camera, preprocessing steps are necessary to separate each wavelength and remove the background in order to obtain the garlic spectrum.

We employed PLS-DA and LS-SVM as machine learning models, and DNN and CNN as deep learning models. The LS-SVM model showed the best overall performance, with a small difference between the accuracy of the calibration model and the accuracy of the prediction model, and relatively higher accuracy than other models. This can be attributed to the nonlinear nature of the differences between groups in garlic crops under high-temperature conditions. The LS-SVM model effectively captures these nonlinear relationships. On the other hand, the DNN and RP-CNN models did not outperform the LS-SVM model, mainly because the depth of their layers was not sufficient. Although there was no significant difference in performance between the DNN and RP-CNN models, the RP-CNN model has the potential to show significant improvement by using deeper layers, as it incorporates not only data but also information about the relationships between data.

Despite the rigorous evaluation process, there are certain limitations and potential biases that should be acknowledged. Firstly, the evaluation was performed on a specific dataset collected under controlled experimental conditions. The generalizability of the models to different environments and variations in data collection protocols may vary. Secondly, the choice of performance metrics and evaluation criteria is subjective to some extent and can impact the interpretation of results. It is important to consider the specific goals and requirements of the application when selecting the best model. However, it is crucial to acknowledge the limitations and potential biases in the evaluation process, such as dataset-specific considerations and hyperparameter selection. Further research and validation on diverse datasets are necessary to ascertain the models’ applicability and generalizability to real-world agricultural scenarios.

The findings of our study have significant implications for practical applications in agriculture, particularly in the detection of heat stress and the effective management of garlic crops using snapshot-based multispectral imaging. By leveraging the benefits of multispectral imaging technology, farmers can enhance their ability to detect heat stress in garlic crops and implement timely interventions for better crop management. Firstly, the use of snapshot-based multispectral imaging allows for non-destructive and rapid assessment of heat stress in garlic crops. This technology enables the simultaneous capture of multiple narrow spectral bands, providing valuable information about the physiological status of the plants. By analyzing the spectral responses across different wavelengths, farmers can identify specific indicators or patterns associated with heat stress, such as changes in chlorophyll content, water content, or metabolic activity. This knowledge can aid in early detection and intervention, allowing farmers to mitigate the detrimental effects of heat stress on garlic crops. Secondly, snapshot-based multispectral imaging provides spatial information along with spectral data, enabling the identification of localized areas affected by heat stress. By mapping the spatial distribution of stress indicators, such as temperature gradients or variations in vegetation indices, farmers can precisely target the affected regions for remedial actions. For example, they can implement site-specific irrigation or shading techniques to alleviate heat stress in specific areas, leading to more efficient resource allocation and improved crop yield. In summary, the use of snapshot-based multispectral imaging in the context of heat stress detection and garlic crop management offers several practical benefits for farmers. It enables non-destructive and rapid assessment, facilitates spatial mapping of stress indicators, supports precision agriculture practices, and empowers data-driven decision making. By leveraging this technology, farmers can detect heat stress early, implement targeted interventions, optimize resource allocation, and ultimately achieve improved crop health, yield, and economic outcomes.

Despite the significant contributions made by this study, certain limitations and areas for future research have been identified. One limitation is the reliance on a specific range of wavelengths for analysis. Further investigations into the optimal wavelength range and the inclusion of additional spectral bands could provide more comprehensive insights into heat stress detection in garlic crops. Furthermore, the study focused on a specific geographical region and a limited number of garlic cultivars. Extending the research to different regions and incorporating a wider range of cultivars would enhance the generalizability and applicability of the findings. Additionally, the study primarily focused on heat stress detection in garlic crops. Exploring the application of snapshot-based multispectral imaging in the detection and management of other stresses, such as water stress or nutrient deficiencies, would expand the scope of research in this field.

In conclusion, this study significantly contributes to the understanding of heat stress detection in garlic crops through the utilization of snapshot-based multispectral imaging and advanced modeling techniques. The findings highlight the superiority of the LS-SVM model and the potential of this imaging technology for precision agriculture applications. The identified limitations and areas for future research pave the way for further advancements in the field, facilitating the development of more robust and comprehensive tools for stress detection and crop management in garlic cultivation.