Study on the Detection of Chlorophyll Content in Tomato Leaves Based on RGB Images

Abstract

Chlorophyll is a key substance in plant photosynthesis, and its content detection methods are of great significance in the field of agricultural AI. These methods provide important technical support for crop growth monitoring, pest and disease identification, and yield prediction, playing a crucial role in improving agricultural productivity and the level of intelligence in farming. This paper aims to explore an efficient and low-cost non-destructive method for detecting chlorophyll content (SPAD) and investigate the feasibility of smartphone image analysis technology in predicting chlorophyll content in greenhouse tomatoes. This study uses greenhouse tomato leaves as the experimental object and analyzes the correlation between chlorophyll content and image color features. First, leaf images are captured using a smartphone, and 42 color features based on the red, green, and blue (R, G, B) color channels are constructed to assess their correlation with chlorophyll content. The experiment selects eight color features most sensitive to chlorophyll content, including B, (2G − R − B)/(2G + R + B), GLA, RGBVI, g, g − b, ExG, and CIVE. Based on this, this study constructs and evaluates the predictive performance of multiple models, including multiple linear regression (MLR), ridge regression (RR), support vector regression (SVR), random forest (RF), and the Stacking ensemble learning model. The experimental results indicate that the Stacking ensemble learning model performs the best in terms of prediction accuracy and stability (R² = 0.8359, RMSE = 0.8748). The study confirms the feasibility of using smartphone image analysis for estimating chlorophyll content, providing a convenient, cost-effective, and efficient technological approach for crop health monitoring and precision agriculture management. This method helps agricultural workers to monitor crop growth in real-time and optimize management decisions.

1. Introduction

Tomato (Solanum lycopersicum L.) [1] is an annual herbaceous plant belonging to the Solanaceae family and the Solanum genus, with plants reaching up to 2 m in height. As one of the most widely cultivated fruit vegetables worldwide, tomatoes play a significant role in major producing countries such as the United States, Italy, and China. Tomatoes are not only rich in nutrients, containing carotene, vitamin C, and B vitamins [2], but also have a unique flavor, making them suitable for both raw consumption and various cooking methods, such as boiling and stir-frying. In addition, tomatoes are widely used in the food processing industry, where they are made into products like beverages, dried fruits, and seasonings. They hold a crucial position in both global agriculture and the food industry [3].

Chlorophyll is a key substance for photosynthesis in green plants and an important indicator for evaluating plant nutritional status [4]. Obtaining real-time and accurate chlorophyll content data is crucial for scientific fertilization, pest and disease control, and crop management [5], facilitating healthy crop growth. In the process of crop growth, key indicators of plant health include chlorophyll fluorescence, chlorophyll content, and nutritional and water status. Among these, the chlorophyll level in the leaves shows a significant correlation with the plant’s growth condition and overall health [6]. Therefore, as an important economic crop, studying the chlorophyll levels in tomato leaves is crucial for ensuring the healthy growth of tomato plants.

Traditionally, chlorophyll content is measured through chemical analysis, typically using specialized instruments in a laboratory setting [7]. In contrast, the portable handheld chlorophyll meter (SPAD meter) is based on the principle of Beer’s Law and uses two characteristic wavelengths, 650 nm and 940 nm, for simultaneous detection. It enables rapid and non-destructive measurement of chlorophyll content in plant leaves. This instrument is low-cost and easy to carry and operate, and obtains SPAD values that are highly correlated with the chlorophyll content of crops [8]. Uddling et al. [9] studied the relationship between SPAD values and chlorophyll concentrations in birch, wheat, and potato leaves, demonstrating significant correlations. Similarly, Markwell et al. [10] measured the chlorophyll concentration in soybean and corn leaves using organic extraction and spectrophotometry, and compared the results with the output values from the SPAD-502 chlorophyll meter. The study showed a significant correlation between the two methods. However, SPAD meters have limitations, such as their small measurement area during a single test, making it difficult to meet the demand for large-scale rapid detection. This constrains their application and widespread adoption in practical production scenarios [11]. Li et al. [12] used a SPAD meter to measure the chlorophyll content of Hami melon leaves and verified the relationship between RGB image features and the measured chlorophyll content. The results indicated that RGB images can effectively predict chlorophyll content.

With the advancement of sensor technology, chlorophyll detection methods based on spectrometers or spectral imagers have been extensively studied. By acquiring near-infrared spectra [13], chlorophyll fluorescence spectra [14], multispectral images [15], and hyperspectral images [16,17] of crop leaves or canopies, and extracting spectral features, chlorophyll content can be predicted. Hyperspectral imaging technology, which combines imaging and spectroscopy, enables rapid and non-destructive detection of chlorophyll content, addressing the limitations of traditional manual measurement methods. For instance, Qiao Lang et al. [18] used drone remote sensing technology to collect aerial images of field corn. By constructing color features and building a maize canopy chlorophyll content detection model based on a BP neural network, their model achieved an R² of 0.72. However, spectrometers are expensive, and their data processing and analysis procedures are complex and time-consuming, limiting their widespread application [19]. The aforementioned methods all rely on specialized and costly instruments, resulting in high measurement costs. Additionally, traditional methods not only require destructive sampling of the leaves, but also involve complex procedures and significant delays [7], making large-scale application difficult. Therefore, it is crucial to find a simple, cost-effective, and easily scalable method for estimating chlorophyll content in plant leaves.

The combination of digital image processing technology and visible light devices enables the fast and non-destructive segmentation of leaf images. By utilizing the red, green, and blue (RGB) channels, different color models are used to quantify the color information in digital images [20], allowing for the rapid, non-destructive detection of chlorophyll content in crops. Related studies have shown that changes in chlorophyll content and other pigments within plants lead to alterations in the color of crop leaves. Therefore, by combining digital image processing technology to extract color features highly correlated with chlorophyll content, rapid and non-destructive detection of chlorophyll content in crops can be achieved [21].

In the past, researchers primarily used the RGB color model to monitor plant health. Most studies employed digital cameras to capture leaf images and analyzed the relationship between the R, G, and B values and the chlorophyll and nitrogen content [22]. The color characteristics of plant leaves are primarily influenced by their internal spectral absorption and reflection properties. Chlorophyll a and b have the strongest absorption capabilities in the blue light (430–450 nm) and red light (640–660 nm) bands, while their absorption of green light (500–600 nm) is weaker. This results in leaves primarily reflecting green light, giving them their green appearance [23]. This absorption characteristic allows for the use of R, G, B values and their combined color indices to estimate chlorophyll content and plant health. For instance, Riccardi et al. [24] found that, based on the best-fit regression model for individual color component variables, in amaranth, the R value shows a higher correlation with chlorophyll content, while in quinoa, the G value has a higher correlation with chlorophyll content. Hu et al. [25] used a SPAD meter to measure leaf chlorophyll content and found that RGB color features such as R, G, R + G + B, R−B, R + B, and R + G were significantly correlated with chlorophyll content. Sulistyo [26] used a standard digital camera to directly capture wheat field images, extracted 12 statistical features from the RGB color space, and established a neural network model for predicting wheat leaf nitrogen content. Amaral et al. [27] extracted color features of tropical tree seedling species during the nursery stage using RGB images. The study found that R, 2R(G − B)/(G + B), and 2G(G − B)/(G + B) were most beneficial for analyzing the physiological status and quality of tropical seedlings. Guo et al. [28] extracted color features from maize RGB images and analyzed their correlation with chlorophyll content. Based on this, they developed chlorophyll content estimation models using a backpropagation neural network (BP), support vector machine regression (SVM), and random forest (RF), with the highest coefficient of determination (R²) reaching 0.85.

Research on extracting color features from RGB images has also made progress in plants such as cotton [29], microalgae [30], and citrus [31]. Additionally, studies on crops such as spinach [32] and willow [33] have confirmed that color features extracted from RGB images can achieve rapid detection of leaf chlorophyll content.

With the widespread popularity of smartphones, phone cameras have become a common feature. The RGB images captured by these cameras are not only easy to obtain but also relatively low in cost [34]. By providing spectral information of color images such as red, green, blue (RGB), and hue, saturation, intensity (HSI), constructing different color indices based on RGB and HSI helps researchers estimate the chlorophyll concentration in leaves. Therefore, the use of RGB image processing technology to obtain plant chlorophyll content has gradually gained widespread application due to its non-destructive, real-time, simple, and accurate characteristics. Mohan et al. [35] obtained images of rice leaves and their chlorophyll content using a smartphone and a SPAD meter, and employed digital image analysis methods to detect chlorophyll content. The study found that the RGB model and the DGCI-rgb model both performed well in predicting chlorophyll content, with the latter achieving a correlation coefficient of 0.63. Similarly, Li et al. [36] proposed a low-cost method for acquiring plant leaf RGB images using a smartphone combined with an auxiliary photographing device, and for instantly detecting chlorophyll content. Through color difference correction and optimization with various machine learning models, they constructed a prediction model. The study showed that the SVR model had higher accuracy and better performance, with a correlation coefficient reaching 0.73. Li et al. [12] used a smartphone to measure the chlorophyll content in Hami melon leaves. They selected (B − G − R)/(B + G) and (G − B)B/(R + G) as feature combinations. Through model prediction verification, the results showed that the random forest regression model performed the best.

The aforementioned studies provide references for extracting image features from mobile devices and modeling chlorophyll detection. At the same time, color correction [36] technology corrects and adjusts the color differences in images through mathematical models and algorithms, ensuring that the image displays consistent colors under different devices and lighting conditions. Due to differences in sensors, lenses, and color processing algorithms used by different models of smartphones and cameras, images of the same object captured on different devices may exhibit color deviations. Therefore, color calibration technology plays a key role in the image acquisition and processing process, providing compatibility for different smartphone models when capturing images. It prevents color deviation issues caused by device differences, thereby ensuring the consistency and accuracy of image data.

However, most current research focuses on estimating chlorophyll content in field crops such as wheat and corn, typically conducted in controlled environments, which limits these studies’ applicability and flexibility. In comparison, research on non-destructive, close-range estimation of chlorophyll content in tomato leaves in natural environments is relatively scarce. Different crops exhibit distinct mathematical fitting behaviors during chlorophyll content estimation, and significant differences exist between species in this regard [37]. Therefore, research methods need to be adjusted according to the specific plant species and its growth stage.

Greenhouse cultivation, as an efficient and sustainable agricultural production method, has gradually gained widespread application due to its ability to effectively control environmental factors and improve crop yield and quality. However, the spatial limitations of greenhouse environments make it challenging for traditional drone spectrometers to efficiently capture crop leaf images, which poses a challenge for precise monitoring of crop growth. In contrast, the complex and variable background of crop leaves presents additional difficulties for the application of digital image technology. Although previous studies [12,35,36] have explored digital image-based chlorophyll measurement methods, most existing approaches rely on controlled environments or expensive equipment, lacking a universal solution suitable for greenhouse environments and complex backgrounds. To address this research gap, this paper proposes an innovative, smartphone-based, rapid, non-destructive, and environmentally adaptable chlorophyll content recognition method. By using black cardboard to block complex backgrounds, non-destructive RGB images of tomato leaves are captured, and the correlation between image features and chlorophyll content is analyzed. Machine learning techniques are then employed to construct prediction models. This method does not rely on controlled environments and can be widely applied to various practical scenarios. It not only significantly improves the convenience of chlorophyll content recognition but also ensures the integrity of the plants, making it easier to conduct further research on other parameters at multiple time points. The results of this study can be widely applied in smart agriculture practices, enabling real-time monitoring of tomato plant growth, optimizing fertilization management, and improving agricultural resource utilization efficiency.

2. Materials and Methods

2.1. Leaf Data Collection

The experiment was conducted in 2024 at the cucumber science and technology yard in Juye, Shandong Province. The tested tomato variety was “Shuang fei 8.” Thid study focuses on tomato leaves at the flowering stage (leaf age 22–28 days), a critical phase in tomato growth and development. During this period, the physiological state of the leaves and the measurement conditions are optimal, making it an ideal time for chlorophyll content determination. The flowering stage is typically associated with high photosynthetic efficiency, and previous studies [38] have shown a significant correlation between leaf physiological status and fruit set rate. In contrast, younger leaves (around 15 days old) may not be fully developed, while older leaves (over 40 days old) often exhibit varying degrees of chlorophyll degradation. In the experimental design, a random sampling method was employed: from 114 tomato plants at the flowering stage, one healthy leaf with uniform color and no visible pests or diseases was selected from each plant as a sample. All samples were collected using non-destructive methods to ensure that the normal growth and development of the plants were not affected. To ensure the reliability of the experimental materials, the selected leaves should have a consistent color distribution across the entire area, with minimal color variation. There should be no spots, shadows, yellowing areas, or color gradients present. Data collection included measuring the chlorophyll content (SPAD value) of tomato leaves and capturing images of the leaves. The images were taken immediately after measuring the SPAD value to ensure each leaf image corresponded precisely to its SPAD value.

Images were captured using an iPhone 13 Pro smartphone (Foxconn Zhengzhou Factory, Zhengzhou, China). To ensure consistency in RGB values, photographs were taken under stable lighting conditions on clear, windless days during two time windows—9:00–11:00 a.m. and 3:00–5:00 p.m.—reducing the impact of light fluctuations on RGB values. At the same time, to further improve the accuracy and consistency of the image colors, color calibration technology was applied to process the captured images, correcting color deviations caused by differences in device sensors, lenses, and color processing algorithms. During shooting, the camera lens should be kept perpendicular to the leaf surface to avoid color distortion caused by angle deviations, thereby enhancing the reliability and comparability of the image data.

To reduce the interference of complex backgrounds on image feature extraction, a narrow slit was created above the black cardboard. During the image capture, the petiole of the leaf was passed through the slit, and the leaf was laid flat on the surface of the cardboard. This method effectively highlights the leaf’s color features while blocking out interference from soil, weeds, and other factors in the field environment. As a result, RGB images are captured non-destructively, providing a reliable foundation for subsequent color extraction and analysis. The captured leaf image is shown in Figure 1. Additionally, filtering and image smoothing techniques were applied to reduce noise caused by background glare or reflections. All images were stored in JPEG format with a resolution of 3024 × 4032 pixels to ensure high-quality data acquisition. Each image was named using a unique identifier corresponding to the leaf number for subsequent analysis.

2.2. Chlorophyll Content Measurement

After capturing each image, the SPAD-502 chlorophyll meter (Beijing Zhongke Weihe, TYS-4N, Weihe, China) was used to measure the chlorophyll content. To minimize measurement error, measurements were taken by avoiding large veins. For each leaf, six points were sampled as shown in Figure 2, and the average value was calculated as the chlorophyll content for that leaf. A total of 114 samples were measured during the experiment.

2.3. Leaf Color Feature Parameter Extraction

Initially, the resolution of the images was reduced, and the Mask-RCNN algorithm was used to segment the leaf images, removing the black background and completing image preprocessing. Python 3.11, along with the OpenCV library, was then used to further process the segmented tomato leaf images. Gaussian filtering and morphological closing operations were applied to eliminate noise and enhance the images, with the enhanced images shown in Figure 3. Next, image processing techniques were used to extract the region of interest (RoI) from the images. The RGB channels of the RoI images were extracted, and the mean values of R, G, and B were calculated. Subsequently, additional color features were computed through RGB algebraic operations, and the data were converted into other color spaces such as HSI and Lab. A total of 42 color features were obtained. The detailed color features are shown in

Table 1. Selected color features.

Color Space	Number	Color Features	Formula and Meaning
RGB	19	R	R
		G	G
		B	B
		R/B	R/B
		G/B	G/B
		B/R	B/R
		B/G	B/G
		R − B	R − B
		G − B	G − B
		R − B − G	R − B − G
		(G + B − R)/(2B)	(G + B − R)/(2B)
		(G + B − R)/(2G)	(G + B − R)/(2G)
		(G + B − R)/(2R)	(G + B − R)/(2R)
		(B − G − R)/(B + G)	(B − G − R)/(B + G)
		(2G − R − B)/(2G + R + B)	(2G − R − B)/(2G + R + B)
		(R − G − B)/(R + G)	(R − G − B)/(R + G)
		(R − G − B)/(G + B)	(R − G − B)/(G + B)
		(B − G − R)/(R + B)	(B − G − R)/(R + B)
		(B − G − R)/(G + R)	(B − G − R)/(G + R)
Normalized RGB	8	r	R/(R + G + B)
		g	G/(R + G + B)
		b	B/(R + G + B)
		r/b	r/b
		g/b	g/b
		r − b	r − b
		g − b	g − b
		r − g	r − g
HSI	3	H	H channel average value, Represents the type of color (Hue).
		S	S channel average value, Represents the purity of the color (Saturation).
		I	I channel average value, Represents the overall brightness of the color (Intensity).
Lab	3	L	L channel average value, Represents the brightness (Lightness).
		A	A channel average value, Represents the green-red axis (Green-Red).
		B_1	B channel average value, Represents the blue-yellow axis (Blue-Yellow).
Others	9	RGBVI	(g2 − r2)/(g2 + r2)
		GRVI	(g − r)/(g + r)
		ExG	2g − r − b
		GLA	(2g − r − b)/(2g + r + b)
		ExR	1.4r − g
		ExGR	ExG − 1.4r − g
		MGRVI	(g2 − br)/(g2 + br)
		VARI	(g − r)/(g + r − b)
		CIVE	0.441r − 0.881g + 0.3856b + 18.78745

Note: r, g, b: Normalized R, G, and B values, respectively. H, S, I: Hue, saturation, and intensity in the HSI color space. L, A, B_1: Lightness and two opposing color dimensions in the Lab color space. RGBVI: RGB Vegetation Index. GRVI: Green–Red Vegetation Index. ExG: Excess Green Index. GLA: Green Leaf Area Index. ExR: Excess Red Index. ExGR: Excess Green–Red Index. MGRVI: Modified Green–Red Vegetation Index. VARI: Visible Atmospherically Resistant Index. CIVE: Color Index of Vegetation.

. High-correlation color features were then selected for subsequent analysis and modeling. Based on this foundation, an ensemble learning approach was adopted, integrating multiple machine learning models using the Stacking method to further enhance model performance. The experimental workflow is illustrated in Figure 4.

2.4. Construction of Chlorophyll Prediction Models

First, the Pearson correlation between each color feature and chlorophyll content was analyzed to identify color features with high correlations. Then, multiple modeling methods were applied to the dataset, including multiple linear regression (MLR), support vector regression (SVR), ridge regression (RR), random forest (RF), and ensemble learning (Stacking). The performance of these models was compared. During the modeling process, 70% of the data was randomly selected as the training set and the remaining 30% as the test set.

The performance evaluation of the prediction models is critical for assessing their accuracy and stability, which are key factors in determining the quality of the models and the precision of prediction results. Common indicators for evaluating model performance include the coefficient of determination (R²), root mean square error (RMSE), and normalized root mean square error (NRMSE), with their respective formulas shown in Equations (1)–(3). A higher R² value indicates better predictive modeling performance, while lower RMSE and NRMSE values signify stronger model stability.

$$R^2=1-{\displaystyle \frac{\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{\sum _{i=1}^n{\left(y_i-{\overline{y}}_i\right)}^2}}$$

(1)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$

(2)

$$NRMSE={\displaystyle \frac{\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}}{{\overline{y}}_{max}-{\overline{y}}_{min}}}$$

(3)

where $y_i$ is the actual observed value; ${\widehat{y}}_i$ is the predicted value from the model; ${\overline{y}}_i$ is the mean of the actual observed values; $n$ is the total number of samples; ${\overline{y}}_{max}$ and ${\overline{y}}_{min}$ represent the maximum and minimum actual observed values, respectively.

3. Results and Analysis

3.1. Correlation Analysis Between Color Features and Chlorophyll Content

The Pearson correlation between color features and chlorophyll content was analyzed. The Pearson correlation coefficient ranges from −1 to 1, where a positive value indicates a positive correlation, meaning the two variables change in the same direction, while a negative value indicates a negative correlation, meaning the two variables change in opposite directions. By converting RGB, HSI, and Lab color spaces and applying algebraic operations, 42 color features were calculated. The results of the correlation analysis with SPAD values are shown in

Table 2. Correlation analysis between color features and SPAD values.

Color Features	Correlation Coefficient	Color Features	Correlation Coefficient	Color Features	Correlation Coefficient
R	0.27	R − B	−0.43	CIVE	0.75
G	0.14	G − B	−0.51	RGBVI	−0.75
B	0.82	R − B − G	−0.68	GRVI	−0.39
r	−0.35	(B − G − R)/(B + G)	0.62	ExG	−0.75
g	−0.75	(2G − R − B)/(2G + R + B)	−0.75	GLA	−0.75
b	0.66	(R − G − B)/(R + G)	−0.44	ExR	0.32
g/b	−0.71	(R − G − B)/(G + B)	−0.35	ExGR	−0.62
r/b	−0.61	(B − G − R)/(R + B)	0.70	MGRVI	−0.39
R/B	−0.61	(B − G − R)/(G + R)	0.66	VARI	−0.18
G/B	−0.71	(G + B − R)/(2B)	−0.69	L	0.18
B/R	0.59	(G + B − R)/(2G)	0.53	A	0.50
B/G	0.70	(G + B − R)/(2R)	0.34	B_1	−0.53
r − g	0.49	g − b	−0.71	S	−0.70
r − b	−0.59	H	0.18	I	0.14

.

The color features with higher correlation coefficients listed in the table were selected for analysis, and Figure 5 provides a more intuitive illustration of the relationship between these features and SPAD values.

Based on the 42 color features mentioned above, it was found that B, (2G − R − B)/(2G + R + B), GLA, RGBVI, g, g − b, ExG, and CIVE values exhibit a strong correlation with SPAD values. Among them, the B value exhibits a correlation of 0.82 with the SPAD parameter. The scatter plots and regression lines between these eight color features and chlorophyll content are shown in Figure 6.

The analysis results indicate that among the RGB primary colors, the correlation between R and G values with chlorophyll content is relatively low, while the correlation between the B value and chlorophyll content is the highest and positive. For normalized color features (r, g, b), the g value had a higher correlation with chlorophyll content and showed a negative correlation. In terms of color ratio parameters, (2G − R − B)/(2G + R + B) had a strong negative correlation with chlorophyll content. For color difference parameters, g−b demonstrated a higher positive correlation. By converting the RGB color space into HSI and Lab color spaces, it was observed that in both spaces the S value had a higher negative correlation with chlorophyll content. Additionally, in vegetation indices, most color features showed high correlation with chlorophyll content, particularly CIVE, RGBVI, ExG, and GLA, which had strong correlation coefficients. Based on the above color features, constructing a model to estimate chlorophyll content in tomato leaves is statistically feasible and provides a reliable foundation for rapid estimation of chlorophyll content in tomatoes.

3.2. Chlorophyll Content Prediction Model

The superior performance of the Stacking model primarily stems from its ability to integrate the predictive capabilities of multiple base learners and leverage a meta-learner to refine their outputs. The random forest (RF) model excels in handling nonlinear relationships by constructing multiple decision trees and aggregating their results through voting or averaging, effectively capturing complex patterns and nonlinear features within the data. Support vector regression (SVR) performs more stably on small sample datasets, optimizing the prediction boundary by maximizing the margin, which helps mitigate overfitting in limited data scenarios. Meanwhile, multiple linear regression (MLR) is more efficient in handling simple linear relationships, providing quick and accurate linear predictions.

Based on the above analysis, this study proposes a chlorophyll content estimation model for leaves based on the Stacking ensemble approach. Multiple linear regression (MLR), support vector regression (SVR), and random forest (RF) models are selected as base learners, with ridge regression (Ridge) as the meta-learner. A brief introduction to each model is provided below.

(1)

Multiple Linear Regression Model (MLR)

MLR is a model that extends simple linear regression to handle the linear relationships between multiple independent variables (features) and a dependent variable (target value). The calculation formula for MLR is shown in Equation (4).

$$Y={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_p{x}_p+ϵ$$

(4)

In the formula, β₀ is the intercept (constant term); β₁, β₂, …, β_p are the regression coefficients, representing the extent to which each independent variable influences the dependent variable; x₁, x₂, …, x_p are the independent variables (feature variables); p is the number of independent variables; and ϵ is the error term, representing the random error between the dependent variable and the predicted value from the regression model.

(2)

Support Vector Regression Model (SVR)

SVR is based on the principles of SVM (support vector machine). It works by finding the optimal hyperplane to minimize the distance from sample points to the hyperplane, thereby predicting continuous variables. The core idea is to define a specific loss function to differentiate data points based on their distances and apply corresponding penalties. In this study, the SVR model parameters were selected using grid search, and the best penalty coefficient (C), kernel function (kernel), and kernel function coefficient (gamma) are shown in

Table 3. Model Parameter Table.

Model	Parameters	Meaning	Set Values
	C	Penalty parameter	1
SVR	Kernel gamma	kernel function kernel parameter	linear scale
RF	n_estimators max_depth min_samples_split min_samples_leaf	number of decision trees maximum depth of the tree minimum samples required to split a leaf node minimum samples per leaf	80 3 2 8
RR	λ	control the strength of regularization	1

Note: SVR, support vector regression. RF, random forest. RR, ridge regression.

.

(3)

Random Forest Model (RF)

RF is an ensemble learning algorithm that generates multiple decision trees by randomly sampling data using the Bootstrap method and combines the predictions of all trees to make a final decision. This algorithm can efficiently handle high-dimensional data and effectively overcome the overfitting issue commonly found in individual decision trees. In this study, the grid search method was used to optimize and select the key parameters of the RF model, including the number of decision trees (n_estimators), maximum tree depth (max_depth), minimum samples required for node splitting (min_samples_split), and minimum samples required for leaf nodes (min_samples_leaf). The specific parameters are shown in Table 3.

(4)

Ridge Regression Model (RR)

Ridge regression is based on ordinary least squares regression, but it applies L2 regularization to the sum of the squares of the regression coefficients. The loss function formula is shown in Equation (5). A larger λ value means stronger regularization, which makes the model simpler but may sacrifice some fitting ability. In ridge regression, the goal is to find a linear model that minimizes the loss function. In this study, the best λ value for the ridge regression model was selected using grid search, as shown in Table 3.

$$Loss=\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2+\lambda \sum _{j=1}^p{\beta }_j^2={\left|\left|y-x\beta \right|\right|}_2^2+\lambda {\left|\left|\beta \right|\right|}_2^2$$

(5)

In the formula, $y_i$ represents the true observed value of the i-th sample; ${\widehat{y}}_i$ is the predicted value for the i-th sample; ${\beta }_j$ denotes the coefficient for the j-th feature; $\lambda$ is the regularization parameter that controls the influence of regularization.

The training parameters for each regression model are shown in Table 3.

3.3. Model Performance Comparison

Prediction models for the chlorophyll content of tomato leaves were established using optimal parameters, and the performance of each model was compared, as shown in Figure 7.

Through correlation analysis and the results shown in Figure 5, it can be observed that some color features (such as CIVE, RGBVI, and GLA) have a high correlation with SPAD values. However, the Pearson correlation coefficient does not reach 1, indicating that there may be a complex nonlinear relationship between them. The multivariate linear regression (MLR) model assumes a linear relationship between input variables and the target variable [39], which limits its ability to handle nonlinear features. When there are interactions between color features or higher-order nonlinear dependencies, linear models may fail to capture these characteristics effectively, leading to reduced prediction accuracy. Experimental results also confirm this, as the R² value of the MLR model is only 0.7359, indicating weaker prediction performance.

In contrast, nonlinear models such as support vector regression (SVR) and random forest (RF) are better at modeling complex feature relationships. SVR maps data to high-dimensional space using kernel functions, which allows it to learn nonlinear features to some extent. However, it may become unstable in cases of high-dimensional data and excessive feature redundancy. On the other hand, random forest (RF) uses an ensemble of decision trees to adaptively learn the nonlinear dependencies between variables, offering strong feature selection and generalization capabilities. However, RF may be affected by factors such as tree depth, data noise, and hyperparameter settings, which can limit the model’s predictive power in certain cases.

Unlike single models, Stacking combines multiple base models (such as MLR, SVR, and RF) to leverage the advantages of each model across different data patterns, allowing the final meta-learner to capture the complex relationships between input features and SPAD values more comprehensively. Specifically, Stacking employs a layered architecture, where the first layer uses multiple base models to extract different levels of feature information, and the second layer meta-learner further learns the prediction patterns of each base model to optimize the final output. This fusion strategy effectively reduces the bias and variance of individual models, enhancing the model’s robustness in handling nonlinearity, feature interactions, and data noise. As a result, the Stacking model achieves a coefficient of determination (R²) of 0.8359, significantly outperforming the single models. It not only demonstrates higher stability and generalization ability but also more effectively integrates the strengths of different algorithms.

While the Stacking model improves prediction performance, it also has some limitations. First, training multiple base models and the meta-learner is time-consuming, especially when dealing with large datasets. Second, the model’s performance is sensitive to the choice of base models: too little diversity among the base models may result in insufficient variation, while too much difference may increase the risk of overfitting. Additionally, hyperparameter tuning is complex, as each model requires independent tuning, which increases the computational cost.

In summary, although the Stacking ensemble method theoretically improves prediction performance through model combination, its applicability should be carefully evaluated in practical applications. Especially when dealing with large datasets, scenarios requiring strict model interpretability, or poor data quality, it is advisable to consider the balance between computational cost, model complexity, and performance improvement, and select an appropriate modeling strategy.

4. Discussion

Although Stacking demonstrates high accuracy in chlorophyll content prediction, several influencing factors require further optimization. First, the shooting angle of the smartphone may affect the extraction of leaf color features, as different angles can lead to variations in light reflection, thereby impacting the acquisition of RGB color features [36]. To minimize the effect of angular deviation, the camera lens should be kept perpendicular to the leaf surface during shooting, and a fixed stand or calibration tool should be used to maintain a consistent shooting angle. Additionally, the choice of shooting time is equally critical, as different time periods (such as morning, noon, and evening) have varying lighting conditions that may cause color feature shifts, affecting model stability [40]. To reduce the impact of lighting variations, images should be captured during stable light periods (e.g., 9:00–11:00 a.m. and 3:00–5:00 p.m.), and color correction techniques should be applied to enhance data consistency.

Although nonlinear models (such as RF and Stacking) have overcome some of the limitations of traditional linear models, the effectiveness of chlorophyll prediction models is still limited by the selection of color features. Currently, color features primarily come from color spaces such as RGB, HSI, and Lab. While these features can partially reflect changes in chlorophyll, their representational ability may not be sufficient to fully capture the complex variations in chlorophyll content. Therefore, introducing deep learning models to extract hidden features, or exploring more complex color spaces (such as HSV or YCbCr), holds promise for further improving the predictive performance of the model. Moreover, environmental factors such as light intensity, humidity, and temperature have not been fully considered in the current model, despite their dynamic variations potentially affecting chlorophyll content measurement accuracy and increasing prediction errors. Therefore, incorporating these environmental variables into the model is crucial for improving prediction accuracy and stability. Future studies could introduce environmental factors as additional input features or conduct correlation analyses during model training to minimize external influences on prediction results. This approach would not only optimize the model’s generalization ability but also enhance its applicability under varying environmental conditions, further strengthening its value in smart agriculture by ensuring precise crop health monitoring and management.

The proposed model provides a rapid and non-destructive chlorophyll sensing method for precision agriculture, enabling farmers to promptly assess the nutritional status of tomato plants and optimize fertilization management, thereby reducing resource waste. Future research can further refine the Stacking model structure to minimize redundant computations and enhance its adaptability to large-scale datasets. Additionally, existing modeling approaches still have room for improvement, such as integrating more color features, incorporating other types of sensor data, or refining algorithms to further enhance prediction accuracy and model applicability. Although this study focused on a single tomato variety at a specific growth stage, variations in chlorophyll content, leaf color characteristics, and spectral reflectance among different tomato varieties or genotypes may impact the model’s predictive performance [41]. Therefore, future research should expand trials to multiple crops, varieties, and growth stages to validate the method’s robustness and applicability. Furthermore, exploring additional image-based texture features and other potential information could improve the model’s predictive capability, enhancing its adaptability and applicability in complex agricultural environments. Future studies will also integrate Internet of Things (IoT) technology to develop a real-time automated chlorophyll sensing system for greenhouse environments. This system will enable real-time data acquisition and visualization at the regional level, providing a solid technical foundation and data support for precision crop management and large-scale deployment.

5. Conclusions

This study investigates a method for predicting chlorophyll content in greenhouse tomato leaves based on RGB color features extracted from smartphone images. The results show that chlorophyll content is significantly correlated with several color features (such as the B value, (2G − R − B)/(2G + R + B), GLA, RGBVI, g, g − b, ExG, and CIVE), demonstrating the feasibility and effectiveness of using smartphone RGB images for rapid and non-destructive chlorophyll estimation. Based on these sensitive features, multiple predictive models were developed, including multiple linear regression, ridge regression, support vector regression, random forest, and a Stacking ensemble model. Among them, the Stacking model performed the best, achieving the highest prediction accuracy (R² = 0.8359, RMSE = 0.8748), significantly enhancing the estimation capability for chlorophyll content.

Compared to traditional hyperspectral imaging and digital camera technologies, smartphones offer a more practical and scalable solution for chlorophyll detection due to their high portability, low cost, and ease of use. This study not only enables efficient and non-destructive detection of tomato chlorophyll content but also lays a foundation for precise assessment of plant nutritional status and growth dynamics. The proposed method provides strong technical support for AI-driven chlorophyll recognition and precision agriculture management, promoting the advancement of smart farming and showing great potential for application and adoption.