A Comprehensive Evaluation Method for Greenhouse-Grown Lettuce Based on RGB Images and Hyperspectral Data

Abstract

Quality grading of greenhouse lettuce requires rapid external appearance screening and nondestructive internal quality assessment. However, existing detection methods struggle to simultaneously evaluate both external and internal quality while maintaining efficiency, resulting in a lack of scientific and comprehensive integrated evaluation standards for current crop grading. To address this issue, this study leveraged the technical strengths of different sensors to construct separate models: an RGB image-based monitoring model for external quality and a hyperspectral-based estimation model for internal quality. Using a combined objective–subjective weighting method, this approach scientifically integrated external and internal quality monitoring indicators to establish a comprehensive evaluation method for greenhouse lettuce quality. The results demonstrate that features such as canopy projection area, compactness, and color components can be extracted from RGB images. Combined with Ridge regression, this approach achieves high-accuracy estimation of lettuce fresh weight and leaf area (R² ≥ 0.880). For intrinsic quality, by combining hyperspectral data with the CARS and SPA band selection algorithms, a Random Forest (RF)-based inversion model for chlorophyll, soluble sugar, protein, and vitamin C content was developed. The AHP-CRITIC method effectively resolved the weight imbalance caused by an excessive coefficient of variation in appearance indicators, thereby achieving the scientific integration of appearance and internal quality data. The grading outcomes of this integrated evaluation method were highly consistent with industry standards (kappa coefficient: 0.788). This approach establishes an effective link between the rapid monitoring of external and internal quality for comprehensive evaluation, providing a novel technical pathway and scientific basis for nondestructive post-harvest detection and automated grading of greenhouse vegetables.

1. Introduction

Lettuce (Lactuca sativa L.) is a leafy vegetable valued both nutritionally and economically. Owing to its short growth cycle and adaptability to controlled environmental conditions, lettuce has become a primary crop for continuous year-round cultivation in modern plant factories and protected agriculture systems [1]. Lettuce quality affects consumer safety and sensory experience, as well as functions as a key determinant of market value and production profitability [2]. Therefore, the development of rapid and accurate quality evaluation methods is crucial for achieving standardized production and efficient management of greenhouse-grown lettuce. Traditionally, the assessment of vegetable quality predominantly relied on sensory evaluation and physicochemical analyses. Although intuitive, these methods are susceptible to subjective experience and encounter challenges in establishing quantifiable grading standards. Physicochemical testing based on technologies such as chromatography and mass spectrometry exhibits high accuracy; however, they have disadvantages such as high invasiveness, complex pretreatment, and limited timeliness [3]. These constraints make them unsuitable for real-time, nondestructive monitoring of crop growth in controlled environments.

In the fields of precision agriculture and high-throughput phenotyping, optical-based non-contact, non-destructive detection technologies have emerged as a research frontier for crop condition monitoring [4]. RGB imaging technology, with its advantages of high spatial resolution, low cost, and rapid imaging speed, excels at capturing intricate geometric structures and color characteristics of crops. It has been widely applied in assessing external quality attributes such as plant architecture and appearance [5,6]. However, unlike smooth-surfaced, compact-structured fruit crops, leafy vegetables like lettuce possess unique biological characteristics—such as overlapping leaves that cause mutual shading and biomass primarily composed of foliage—making it difficult for visible light to penetrate the phenotypic barrier. This hinders the effective quantification of internal physiological and biochemical indicators. Hyperspectral technology, with its high spectral resolution and continuous band coverage, precisely captures unique spectral information of samples. This provides a technical solution to the limitations of RGB imaging in monitoring intrinsic quality, enabling accurate estimation of chlorophyll and other physiological/biochemical components within lettuce [7,8,9]. Furthermore, it offers robust data support for multidimensional crop quality assessment, ultimately achieving synergistic non-destructive evaluation of both lettuce appearance and internal quality. Current lettuce evaluation standards primarily focus on external quality [10,11] and lack the simultaneous consideration of internal qualities that determine edible value. This results in notable limitations in the quality assessment. Therefore, to enhance the grading accuracy and practicality, developing a dual-dimensional integrated evaluation method that combines external and internal qualities is necessary. This approach enables optimization and upgrading from single-aspect appearance judgment to comprehensive quality assessment.

In this study, we investigated the application potential of multi-source sensing technologies for monitoring leafy vegetable quality using greenhouse-grown lettuce. First, rapid monitoring models for external indicators such as fresh weight and leaf area were constructed based on RGB image feature extraction. Subsequently, spectral inversion models for internal quality indicators, including soluble sugars, proteins, and vitamin C, were investigated using hyperspectral technology combined with machine-learning algorithms. Finally, by integrating the appearance and internal quality monitoring data, a comprehensive lettuce quality evaluation system based on subjective and objective weighting was established. This study is expected to establish a reliable technical approach for quality monitoring of greenhouse-grown lettuce, develop an integrated technical pathway from rapid monitoring to precise evaluation, and provide theoretical and methodological foundations for accurate grading of vegetables post-harvest.

2. Materials and Methods

2.1. Research Area

The experimental site was located in the plant factory of the Agricultural Science Building at Shihezi University in the Xinjiang Uygur Autonomous Region, situated at 44°18′ N, 86°03′ E. The experimental variety used in this study was loose-leaf lettuce. Cultivation was conducted from March to May 2025 and from May to July 2025. Plants were grown in 10 cm diameter pots with a spacing of 10 cm between plants and 10 cm between rows. The greenhouse was maintained at an average temperature of 25 ± 2 °C and relative humidity of 50 ± 10% throughout the day and night. Base fertilizer (a complete soil drip irrigation fertilizer) was applied prior to sowing. Two additional fertilizer applications were administered during the entire growth period, at 30 and 40 d after sowing. A schematic diagram of the test area is shown in Figure 1.

2.2. Data Collection and Measurements

2.2.1. Quality Index Measurement

Fresh weight, leaf area, soluble protein content, soluble sugar content, vitamin C content, and chlorophyll content were measured for each lettuce plant. The specific measurement methods are as follows:

Fresh weight was measured using an electronic balance to weigh the aboveground fresh weight [12], leaf area was determined using digital image processing [13], soluble protein content was measured using the bicinchoninic acid (BCA) microplate assay [14], soluble sugar content was determined using the anthrone colorimetric method [15], and vitamin C content was determined using the Ferrozine Colorimetry method [16] (all kits were provided by Nanjing Jiancheng Bioengineering Institute, Nanjing, China). Chlorophyll content was determined using a spectrophotometer [17]. Three leaves were randomly selected from each lettuce plant, and three leaf discs (0.7 cm² each) were excised. The discs were then immersed in 7 mL of methanol and incubated in the dark at 4 °C for 12 h for pigment extraction prior to measurement. The calculation equations are as follows:

$${Chl}_a=12.25\times A_{664}-2.79\times A_{647}$$

(1)

$${Chl}_b=21.5\times A_{647}-5.1\times A_{664}$$

(2)

$${Chl}_{a+b}={Chl}_a+{Chl}_b$$

(3)

2.2.2. RGB Image and Hyperspectral Data Acquisition

Data were acquired using a Trait-RGB visible-light imaging unit and a Trait-Hyperspec hyperspectral imaging unit within a high-throughput plant phenotyping platform (PhenoTrait Technology Co., Ltd., Beijing, China; Hui Nuo Rui De Technology Co., Ltd., Shenzhen, China). The Trait-RGB visible light imaging unit has a resolution of 4924 × 3283 pixels, employing an exposure time of 5000 us with automatic gain set to 2.0 dB. The Trait-Hyperspec hyperspectral imaging unit covers a spectral range of 400–1000 nm, comprising 448 bands with a full width at half maximum (FWHM) of 1.36 nm, an exposure time of 8 ms, and a default frame rate of 50 Hz. A uniform acquisition distance of 1.0 m was maintained. A schematic of the equipment is shown in Figure 2.

2.2.3. Feature Extraction Methods

(1)

RGB Image Feature Parameter Extraction

The original images were uniformly resized to a resolution of 2500 × 2500 pixels. Lettuce image segmentation was performed using color thresholding, morphological operations, and connected-component analysis. The color and morphological features of the lettuce were calculated from binary images obtained using automatic image segmentation algorithms. This study extracted 21 feature indicators, including 12 color features: red (R_Avg), green (G_Avg), and blue (B_Avg) components; hue (H_Avg); saturation (S_Avg); value (V_Avg); lightness (L*_Avg); green–red difference (a*_Avg); blue–yellow component (b*_Avg), lightness component (Y_Avg), red difference component (Cr_Avg), and blue difference component (Cb_Avg).

Color features were extracted using the first-order moment (Average) of the image. The first-order moment reflects the overall pixel intensity of the image. The formula for the first-order moment is as follows:

$$Avg={\displaystyle \frac{1}{M+N}}\sum _{i=1}^M\sum _{j=1}^{N}P(i,j)$$

(4)

where M and N represent the number of rows and columns of the target image, P(i,j) denotes the pixel intensity value at coordinates (i, j) in the target image, and Avg represents the first-order moment.

Nine morphological characteristics, including canopy projection perimeter (MC_1), canopy projection area (MC_2), convex hull perimeter (MC_3), convex hull area (MC_4), compactness (MC_5), major axis length of the bounding rectangle (MC_6), minor axis length (MC_7), their aspect ratio (MC_8), and circularity (MC_9), were used.

Area and perimeter are typical shape characteristics. The canopy projection perimeter (MC_1) of the lettuce was obtained by calculating the total contour length of the polygon. Furthermore, the area within the enclosed contour boundary was calculated using Green’s theorem to obtain the canopy projection area of the lettuce (MC_2). A convex hull, defined as the smallest convex polygon completely enclosing a target region, is widely applied in image processing and pattern recognition. The Sklansky algorithm was employed to rapidly locate the smallest convex polygon enclosing the target contour, and its vertex coordinates were extracted. Following this method, the convex hull of the lettuce canopy was obtained. The perimeter of this convex hull (MC_3) was calculated directly by summing the Euclidean distances between adjacent vertices, while its area (MC_4) was computed using Gauss’s area formula. Compactness (MC_5) intuitively reflects the fullness of the lettuce canopy and the density of its leaves. This ratio is a value between 0 and 1; results closer to 1 indicate a fuller growth shape with smoother edges. The equation for calculating compactness is as follows:

$$MC\_5={\displaystyle \frac{MC\_2}{MC\_4}}$$

(5)

The axes of the bounding rectangle were aligned to the principal axes of the target. The width and height of this rectangle were extracted and defined as the major axis length (MC_6) and minor axis length (MC_7), respectively. The aspect ratio (MC_8) was then calculated as the ratio of the major axis to the minor axis.

$$MC\_8={\displaystyle \frac{MC\_6}{MC\_7}}$$

(6)

Circularity (MC_9) was calculated from the canopy projection perimeter (MC_1) and canopy projection area (MC_2) of the target area using the following equation:

$$M{C}_{\_}9={\displaystyle \frac{4\pi (MC\_2)}{{(MC\_1)}^2}}$$

(7)

Simple correlations among phenotypic parameters were analyzed using Pearson’s correlation coefficient to identify representative indicators.

(2)

Hyperspectral Feature Band Extraction

In this study, three preprocessing methods were applied: Multiplicative Scatter Correction (MSC) [18], Standard Normal Variate (SNV) [19], and Savitzky–Golay (SG) smoothing [20]. MSC is employed to mitigate scattering effects induced by uneven distribution of solid particle sizes and variations in optical path length, thereby enhancing the spectral absorption features. This technique first calculates the mean spectrum of all samples, which serves as the ideal reference spectrum. For each individual sample spectrum, a linear regression is performed against this mean spectrum using the least squares method to determine the slope and intercept. A linear transformation is then applied to correct the spectrum. In this study, the mean spectrum was derived from the full sample spectral matrix, and each sample spectrum was normalized accordingly. SNV was applied to eliminate the effects of optical path length variations and surface scattering from solid particles. Its algorithm standardizes each spectrum individually by subtracting the spectrum’s mean value and then dividing by its standard deviation. This method was primarily used in this study to correct for baseline drift between samples. The SG smoothing algorithm was used to reduce high-frequency noise while preserving the features of spectral peaks and valleys. It operates by fitting a polynomial to the spectral data within a moving window. The window width was set to 11 points, and the polynomial order to 2. The algorithm filters the spectral data through a convolution process.

Feature-band selection employs a Sequential Projection Algorithm (SPA) and Competitive Adaptive Re-weighted Sampling (CARS). SPA extracts the most representative feature variables by eliminating redundant information and multicollinearity [21]. The Specific Peak Analysis (SPA) algorithm was then employed for feature selection. The procedure begins by using a projection operator to calculate the projection of each wavelength vector onto the orthogonal subspace of the pre-selected wavelength bands. The wavelength corresponding to the maximum projection norm is then selected and added to the subset. To determine the optimal number of bands, the maximum number of bands to be selected was set to 60. For each candidate subset size, a Partial Least Squares (PLS) model with 5-fold cross-validation was built, and the root mean square error of cross-validation (RMSECV) was calculated. Finally, the subset of bands that yielded the smallest RMSECV was chosen as the optimal feature set.

CARS combines Monte Carlo sampling with an exponential decay function mechanism to select the most effective feature variables through iterative competition [22]. For further feature refinement, the Competitive Adaptive Reweighted Sampling (CARS) method was used. This algorithm operates through an iterative process with 100 iterations. In each iteration, Monte Carlo sampling is first employed to select a fixed ratio of samples for building a PLS model. Simultaneously, an exponentially decreasing function (EDF) is used to enforce the removal of a portion of wavelength variables with small absolute regression coefficients. Following this forced removal, the absolute values of the regression coefficients from the PLS model are used as an importance metric for the remaining wavelengths. Adaptive reweighted sampling is then applied to select, based on these weights, the wavelengths that form a new candidate subset for the next iteration. After completing all iterations, 5-fold cross-validation is performed to compute the RMSECV for the subset of bands selected at each iteration. The subset corresponding to the global minimum RMSECV across all iterations is ultimately selected as the optimal feature bands.

2.3. Modeling Approach

This study used the Python version 3.8 programming environment to construct models that employed algorithms, including a Backpropagation Neural Network (BPNN) [23], Partial Least Squares Regression (PLSR) [24], Ridge Models [25], Multiple Linear Regression (MLR) [26], Random Forests (RFs) [27], AdaBoost Algorithms [28], and Bagging [29].

A BPNN is a nonlinear system with strong mapping capabilities. Through its three-layer structure—comprising an input layer, a hidden layer, and an output layer—it can effectively approximate complex nonlinear functions, making it particularly suitable for modeling intricate nonlinear data. PLSR is an advanced statistical method that integrates multiple linear regression, principal component analysis, and canonical correlation analysis. It extracts components for modeling from both the independent and dependent variables simultaneously, thereby overcoming the limitations of ordinary least squares regression by being robust to multicollinearity among the independent variables. As a classical linear method, MLR establishes a linear relationship between multiple independent variables and a dependent variable. While it is structurally simple and offers high interpretability, its application is limited when dealing with complex, nonlinear data. The Ridge model incorporates an L2 regularization term into the loss function. This effectively mitigates the risk of overfitting, which in this study was associated with weak correlations among the features MC_2, MC_5, and a*_Avg. By doing so, the model retains clear biological interpretability while stably quantifying the influence of the selected features on fresh weight and leaf area. This approach achieves superior predictive accuracy compared to the nonlinear models tested, while also significantly reducing computational cost. Consequently, it offers a more practical solution for rapid and efficient monitoring in real-world production settings. RF achieves nonlinear fitting by constructing an ensemble of decision trees. Each tree is trained on a bootstrap sample (i.e., a sample drawn with replacement from the original data), and the final prediction is obtained by averaging the predictions of all individual trees. Similarly, the Bagging algorithm generates multiple training sets via the bootstrap process to build multiple base learners, enhancing model stability through this ensemble strategy. In contrast, the AdaBoost algorithm focuses on combining multiple weak learners to form a strong learner. It iteratively adjusts the weights of the training data to give more focus to hard-to-classify samples, thereby progressively improving the model’s overall performance.

The Content Gradient Method was employed to evenly partition the lettuce samples into a test set (168 samples) and a validation set (42 samples) in an 8:2 ratio. The coefficient of determination (R²) and root mean square error (RMSE) were used as model evaluation metrics, with five-fold cross-validation introduced during training to assess the model performance more robustly.

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

(8)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{x}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}})}^2}{\mathrm{n}}}}$$

(9)

where ${\mathrm{x}}_{\mathrm{i}}$ represents the simulated value; ${\mathrm{y}}_{\mathrm{i}}$ denotes the actual value; ${\overline{\mathrm{y}}}_{\mathrm{i}}$ is the average of the actual values; and n is the number of samples available for verification.

2.4. Comprehensive Evaluation Method

This study introduced a comprehensive lettuce evaluation method that integrated both the external and internal qualities. First, we calculated the maximum, minimum, mean, standard deviation, and coefficient of variation for growth performance indicators, including external and internal quality, while analyzing inter-indicator correlations. Based on this foundation, we comprehensively applied principal component analysis [30], probabilistic grading [31], subjective-objective weighting [32], and k-means cluster analysis [33] to achieve a holistic assessment of lettuce quality. Finally, using the grading results from the standard NY/T4058-2021 [34] “Grades and specifications for agri-product in market information collection-Leafy vegetables” as the benchmark, the overall accuracy (OA) and kappa coefficient were selected as evaluation metrics to assess the model’s classification accuracy and consistency, thereby identifying the most suitable and comprehensive integrated evaluation method.

Principal Component Analysis is a statistical dimensionality reduction technique that transforms a set of correlated variables into a smaller number of uncorrelated principal components through linear transformation. This approach effectively mitigates issues related to computational complexity and multicollinearity that arise from high-dimensional data, while maximizing the retention of original information. Probability-based classification is a method grounded in probability theory. It accounts for uncertainty and randomness in data by using probability distributions to assign data points to categories or classes. The Subjective-Objective Weighting Method combines subjective judgment with objective data analysis. It integrates expert experience and domain knowledge while also leveraging the inherent characteristics of the data to determine indicator weights, thereby achieving a balance between subjectivity and objectivity. Cluster Analysis, a key unsupervised learning method, automatically groups similar data points into clusters, thereby revealing underlying structures and patterns within the data.

$$OA={\displaystyle \frac{N_{correct}}{N}}\times 100\%$$

(10)

$$\mathit{Kappa}={\displaystyle \frac{N{\sum }_{i=1}^n{x}_{ii}-{\mathit{\sum }}_{i=1}^n{x}_{i+}{x}_{+i}}{N^2-{\mathit{\sum }}_{i=1}^n{x}_{i+}{x}_{+i}}}$$

(11)

$N_{correct}$ represents the number of samples correctly predicted by the model; N denotes the total number of samples for verification; n denotes the total number of columns in the confusion matrix; $x_{ii}$ denotes the sample size for the i row and i column of the confusion matrix; and $x_{i+}$, $x_{+i}$ denote the row and column sums of the confusion matrix.

3. Results

3.1. Construction of Appearance Quality Index Models for Greenhouse-Grown Lettuce Based on RGB Images

3.1.1. Feature Parameter Selection

Based on the Pearson correlation coefficients and heatmap analysis results (Figure 3), significant multicollinearity existed among the various phenotypic features of lettuce. Specifically, the correlations between the canopy projected area (MC_2) and canopy projected perimeter (MC_1), convex hull parameters (MC_3 and MC_4), and bounding rectangle parameters (MC_6 and MC_7) were all approximately 1. This indicates a high information redundancy among these indicators in characterizing plant geometric dimensions. As MC_2 exhibited the highest explanatory power for biomass and had the most apparent physical significance, other redundant terms were eliminated, retaining MC_2 as the sole representative variable for dimensionality. By contrast, compactness (MC_5) exhibited extremely low correlation coefficients with the aforementioned size indicators, demonstrating significant independence. This outperformed the circularity (MC_9) and aspect ratio (MC_8), which showed increased internal correlations, thus confirming that MC_5 is a structural feature that indicates plant spatial density. Among the color metrics, a*_Avg exhibited weak collinearity with morphological indicators and was retained as a representative color dimension. Ultimately, the study selected three core indicators (MC_2, MC_5, and a *_Avg) for subsequent modeling.

3.1.2. Monitoring Model for Lettuce External Quality Indicators

A comparative analysis of the predictive performance of the BPNN, MLR, PLSR, and Ridge regression models revealed that linear regression models consistently outperformed neural network models for both fresh weight and leaf area indicators (Figure 4). Specifically, Ridge regression demonstrated the best goodness-of-fit and generalization capabilities. The coefficient of determination (R²) in the fresh weight prediction test set reached 0.897, whereas the R² for leaf area prediction reached 0.880, which was significantly higher than that of the BPNN model. Model validation indicated a highly significant linear relationship between the three selected key image features and the actual physical indicators of lettuce, with determination coefficients exceeding 0.8. This indicates that these features adequately captured core information regarding lettuce appearance and biomass. Because the direct utilization of image features prevents computational delays and error accumulation associated with secondary inversion, this study directly employed these three key feature vectors to represent the lettuce appearance quality in subsequent comprehensive evaluation analyses. The predicted physical values were no longer used as intermediate variables, resulting in the development of an efficient end-to-end evaluation method.

3.2. Establishment of Hyperspectral-Based Intrinsic Quality Indicator Models for Lettuce

3.2.1. Feature Wavelength Selection

During the selection of feature wavelengths for soluble protein, soluble sugar, vitamin C, and chlorophyll contents in lettuce (Figure 5, Figure 6, Figure 7 and Figure 8), the SPA and CARS methods exhibited consistent patterns under different spectral preprocessing conditions. Overall, following MSC and SNV preprocessing, the CARS algorithm selected relatively fewer variables, demonstrated stable model convergence, and produced 23–30 spectral bands. The raw spectra and SG-processed data resulted in increased variable counts, frequently exhibiting oscillating results, with band selections ranging from 84 to 127. The SPA method performed optimally on raw spectra and SG-processed data, selecting 4–24 bands, whereas MSC and SNV processing resulted in increased selections, ranging from 11 to 60 bands. Both methods indicated that characteristic bands primarily clustered in three regions: 491.75–578.31 nm (green light), 660.43–783.15 nm (red-to-near-infrared transition), and 920.30–1004.52 nm (near-infrared). These regions exhibited high sensitivity toward all four physiological indicators, potentially correlating with chlorophyll absorption, leaf structure, water status, and the characteristic absorption of organic compounds such as proteins, sugars, and vitamin C. This demonstrates the strong relationship between the spectral information and the intrinsic quality of lettuce.

3.2.2. Lettuce Intrinsic Quality Detection Model

A comprehensive comparison of the performance of the inversion model across the four quality indicators was conducted. To predict the soluble protein content, the Random Forest (RF) model using SNV preprocessing combined with CARS feature extraction demonstrated optimal performance, achieving a validation set R² of 0.872 and an RMSE of 0.166 (

Table 1. Comparison of soluble protein content prediction performance across different machine learning models.

Feature Extraction Method	Preprocessing Method	Model	Train		Evaluation
			R2	RMSE	R2	RMSE
SPA	Original	RF	0.873	0.166	0.866	0.169
	MSC		0.894	0.152	0.863	0.171
	SNV		0.890	0.154	0.849	0.180
	SG		0.884	0.159	0.866	0.170
	Original	AdaBoost	0.539	0.316	0.537	0.315
	MSC		0.647	0.276	0.655	0.272
	SNV		0.629	0.283	0.602	0.292
	SG		0.532	0.318	0.549	0.311
	Original	Bagging	0.830	0.192	0.866	0.170
	MSC		0.794	0.211	0.862	0.172
	SNV		0.807	0.204	0.832	0.190
	SG		0.835	0.189	0.775	0.220
CARS	Original	RF	0.880	0.161	0.870	0.167
	MSC		0.894	0.151	0.857	0.175
	SNV		0.880	0.161	0.872	0.166
	SG		0.877	0.163	0.866	0.170
	Original	AdaBoost	0.520	0.322	0.481	0.334
	MSC		0.563	0.308	0.556	0.309
	SNV		0.606	0.292	0.551	0.310
	SG		0.571	0.305	0.567	0.305
	Original	Bagging	0.863	0.172	0.831	0.190
	MSC		0.819	0.198	0.846	0.182
	SNV		0.845	0.183	0.845	0.182
	SG		0.817	0.199	0.824	0.194

). For soluble sugar content prediction, the combination of MSC preprocessing and the RF model exhibited the highest accuracy, with a validation set R² of 0.914 and RMSE of 0.138, demonstrating good model generalization (

Table 2. Comparison of soluble sugar content prediction performance across different machine learning models.

Feature Extraction Method	Preprocessing Method	Model	Train		Evaluation
			R2	RMSE	R2	RMSE
SPA	Original	RF	0.912	0.156	0.911	0.141
	MSC		0.917	0.152	0.914	0.138
	SNV		0.918	0.151	0.913	0.139
	SG		0.921	0.148	0.909	0.142
	Original	AdaBoost	0.741	0.268	0.685	0.265
	MSC		0.747	0.265	0.761	0.231
	SNV		0.756	0.260	0.724	0.248
	SG		0.757	0.260	0.724	0.248
	Original	Bagging	0.888	0.177	0.884	0.160
	MSC		0.898	0.168	0.893	0.154
	SNV		0.891	0.174	0.868	0.172
	SG		0.879	0.183	0.835	0.192
CARS	Original	RF	0.904	0.163	0.902	0.147
	MSC		0.909	0.159	0.914	0.138
	SNV		0.910	0.158	0.912	0.140
	SG		0.917	0.152	0.909	0.143
	Original	AdaBoost	0.693	0.292	0.612	0.294
	MSC		0.737	0.270	0.702	0.257
	SNV		0.705	0.286	0.671	0.270
	SG		0.696	0.290	0.628	0.288
	Original	Bagging	0.900	0.166	0.811	0.205
	MSC		0.896	0.170	0.856	0.179
	SNV		0.905	0.163	0.840	0.189
	SG		0.885	0.179	0.863	0.174

). For vitamin C content prediction, the combination of SNV preprocessing with CARS feature extraction and RF resulted in the best results, with a validation set R² of 0.899 and an RMSE of 1.257 (

Table 3. Comparison of vitamin C content prediction performance across different machine learning models.

Feature Extraction Method	Preprocessing Method	Model	Train		Evaluation
			R2	RMSE	R2	RMSE
SPA	Original	RF	0.877	1.431	0.878	1.387
	MSC		0.881	1.408	0.893	1.297
	SNV		0.879	1.420	0.899	1.262
	SG		0.886	1.376	0.865	1.459
	Original	AdaBoost	0.559	2.712	0.483	2.855
	MSC		0.669	2.350	0.614	2.468
	SNV		0.631	2.480	0.600	2.511
	SG		0.583	2.637	0.539	2.696
	Original	Bagging	0.841	1.629	0.852	1.528
	MSC		0.851	1.578	0.840	1.590
	SNV		0.848	1.593	0.827	1.652
	SG		0.860	1.526	0.851	1.535
CARS	Original	RF	0.870	1.471	0.872	1.419
	MSC		0.897	1.312	0.879	1.381
	SNV		0.887	1.371	0.899	1.259
	SG		0.874	1.452	0.855	1.510
	Original	AdaBoost	0.582	2.639	0.538	2.700
	MSC		0.701	2.234	0.668	2.287
	SNV		0.660	2.381	0.613	2.471
	SG		0.607	2.561	0.588	2.548
	Original	Bagging	0.839	1.641	0.870	1.431
	MSC		0.843	1.619	0.819	1.688
	SNV		0.861	1.523	0.817	1.699
	SG		0.845	1.608	0.838	1.599

). In chlorophyll content modeling, the RF model using SG preprocessing combined with SPA feature extraction achieved the highest accuracy, with a validation set R² of 0.958 and an RMSE of 1.292 (

Table 4. Comparison of chlorophyll content prediction performance across different machine learning models.

Feature Extraction Method	Preprocessing Method	Model	Train		Evaluation
			R2	RMSE	R2	RMSE
SPA	Original	RF	0.959	1.223	0.956	1.317
	MSC		0.955	1.284	0.949	1.421
	SNV		0.957	1.258	0.949	1.416
	SG		0.962	1.191	0.958	1.292
	Original	AdaBoost	0.812	2.628	0.797	2.839
	MSC		0.815	2.610	0.845	2.481
	SNV		0.825	2.541	0.837	2.540
	SG		0.809	2.654	0.784	2.928
	Original	Bagging	0.927	1.635	0.944	1.485
	MSC		0.939	1.503	0.922	1.761
	SNV		0.939	1.494	0.928	1.690
	SG		0.943	1.451	0.949	1.424
CARS	Original	RF	0.962	1.188	0.949	1.428
	MSC		0.956	1.267	0.947	1.448
	SNV		0.955	1.292	0.947	1.445
	SG		0.952	1.333	0.947	1.449
	Original	AdaBoost	0.844	2.397	0.793	2.865
	MSC		0.824	2.548	0.804	2.787
	SNV		0.827	2.525	0.819	2.682
	SG		0.807	2.669	0.773	3.003
	Original	Bagging	0.935	1.551	0.932	1.648
	MSC		0.929	1.622	0.947	1.456
	SNV		0.949	1.367	0.945	1.476
	SG		0.946	1.412	0.914	1.849

). Overall, the RF models demonstrated strong stability and predictive capability across different indicators and preprocessing methods, whereas the AdaBoost models performed relatively poorly under all conditions.

3.3. Comprehensive Evaluation Methods for Evaluating Greenhouse Lettuce Quality

3.3.1. Statistical Analysis of Lettuce Quality Traits

The coefficient of variation indicates the extent of genetic variation in the traits. As presented in

Table 5. Statistical analysis of lettuce quality traits.

Index	a*_Avg	MC_2 (cm2)	MC_5	Chlorophyll Content	Soluble Protein Content (mg/g)	Soluble Sugar Content (%)	Vitamin C Content (mg/100 g)
Min.	119.55	23.15	0.55	11.80	3.52	0.13	8.29
Max.	127.89	1424.74	0.90	38.70	6.19	2.33	31.01
Mean	124.95	572.07	0.74	25.38	4.80	1.05	20.76
SD	1.89	334.60	0.07	6.15	0.47	0.51	4.06
CV (%)	1.51	58.49	9.43	24.24	9.69	48.77	19.55

, the quality traits of the lettuce samples exhibited significant variations within populations. Specifically, MC_2 exhibited the highest coefficient of variation (58.49%). The soluble sugar, soluble protein, and vitamin C contents each had a coefficient of variation exceeding 10%. The coefficients of variation for MC_5 soluble protein content and a*_Avg were all <10%, with a*_Avg showing the lowest coefficient of variation.

3.3.2. Correlation Analysis of External and Internal Quality Traits in Lettuce

The correlation analysis of the seven lettuce quality indicators resulted in a correlation coefficient matrix (

Table 6. Correlation results of lettuce quality indicators.

	a*_Avg	MC_2	MC_5	Chlorophyll Content	Soluble Protein Content	Soluble Sugar Content
MC_2	−0.990 **
MC_5	0.287 **	−0.269 **
Chlorophyll content	−0.669 **	0.710 **	−0.177 *
Soluble protein content	−0.299 **	0.290 **	−0.009	0.375 **
Soluble sugar content	−0.368 **	0.358 **	−0.088	0.444 **	0.545 **
Vitamin C content	−0.278 **	0.271 **	0.042	0.353 **	0.498 **	0.441 **

Note: ** Correlation is significant at the 0.01 level (two-tailed); * Correlation is significant at the 0.05 level (two-tailed).

). The results indicated varying degrees of correlation among the indicators. Among the appearance-related indicators, a*_Avg exhibited a highly significant negative correlation with MC_2. Among the internal quality indicators, most pairs showed highly significant positive correlations, with relatively high coefficients observed between the soluble protein content and soluble sugar content/vitamin C content. Additionally, the appearance indicators a*_Avg and MC_2 showed highly significant negative correlations with most internal indicators, such as chlorophyll, soluble protein, soluble sugar, and vitamin C. MC_2 exhibited a strong correlation with chlorophyll content, whereas MC_5 showed weak or insignificant correlations with most internal quality indicators.

3.3.3. Probability Grading Method Based on Principal Component Analysis

(1)

Determination of Lettuce Quality Indicators

Principal component analysis was conducted based on seven indicator data points from 210 lettuce samples (

Table 7. Results of principal component analysis of lettuce quality indicators.

Primary Component	Principal Component Load
	Principal Component 1	Principal Component 2
a*_Avg	−0.866	0.386
MC_2	0.869	−0.391
MC_5	−0.279	0.549
Chlorophyll content	0.822	−0.119
Soluble protein content	0.611	0.553
Soluble sugar content	0.668	0.409
Vitamin C content	0.565	0.557
Eigenvalue	3.397	1.400
Total contribution rate/%	48.535	68.538

). Following the principle of extracting components with eigenvalues ˃ 1, two principal components were identified. The cumulative contribution rate of the first two principal components reached 68.538%, representing most of the information from the seven indicators. Therefore, the seven quality indicators were reduced to two factors. In the first principal component, a*_Avg and MC_2 were visual quality indicators. Because MC_2 had the highest factor weight (in absolute value, the same below) and exhibited significant correlations with the other two indicators, MC_2 was selected to represent the visual quality factor of lettuce. In the second principal component, the vitamin C content exhibited a relatively high factor weight and was defined as the intrinsic quality factor.

(2)

Establishment of Grading Standards for Lettuce Quality Evaluation

Based on these two quality indicators, the lettuce was graded into three levels for each indicator using a probability distribution. The Shapiro–Wilk test (p > 0.05) indicated that vitamin C content was normally distributed; therefore, a probability-based grading method based on the normal distribution was employed. The samples were categorized into three grades, namely special, first, and second, using the thresholds (X − 0.524 × S) and (X + 0.524 × S) as the boundaries for classifying the samples into such grades. However, MC_2 exhibited significant non-normality (p < 0.01), and directly applying the normal distribution assumption resulted in severe distortions in the grading proportions. Therefore, to ensure strict adherence to the preset 30%:40%:30% sample ratio across grades, this study employed a quantile-based method to determine the MC_2 grading thresholds. Specifically, the P30 and P70 quantiles of the sample distribution were selected as boundaries for the first versus special and second grades, respectively. The final comprehensive quality grading results for lettuce are presented in

Table 8. Grading standards for key lettuce indicators.

Index	Special Grade	First Grade	Second Grade
MC_2/(cm2)	>759.75	425.13–759.75	<425.13
Vitamin C content/(mg/100 g)	>22.88	18.63–22.88	<18.63

.

3.3.4. Cluster-Based Grading Method Based on AHP-CRIYIC Combined Weighting

(1)

Determination of Subjective Weight Coefficients Using AHP

A comprehensive evaluation index system for lettuce was established, categorizing decision objectives, criteria, and subjects into target, criterion, and indicator layers based on their interrelationships. The hierarchical structure is illustrated in Figure 9. The criterion layer T1 (External Quality) included three indicators: a*_Avg (C1), MC_2 (C2), and MC_5 (C3), and the criterion layer T2 (Internal Quality) comprised four indicators: chlorophyll content (C4), soluble protein content (C5), soluble sugar content (C6), and vitamin C content (C7).

The decision matrix and consistency test for lettuce comprehensive quality were constructed based on the contribution of each influencing factor to the comprehensive quality of lettuce and the importance of each factor, combined with expert experience. This resulted in a decision matrix (

Table 9. Judgment matrix of the overall criteria layer.

Index	C1	C2	C3	C4	C5	C6	C7
C1	1	2	3	1/3	1/5	1/5	1/7
C2	1/2	1	2	1/4	1/6	1/6	1/8
C3	1/3	1/2	1	1/5	1/7	1/7	1/9
C4	3	4	5	1	1/3	1/3	1/4
C5	5	6	7	3	1	1/2	1/6
C6	5	6	7	3	2	1	1/2
C7	7	8	9	4	6	2	1

Note: The consistency ratio CR = 0.056 (CR < 0.1), the judgment matrix passes the consistency test, and the weights are valid.

) for lower-level indicators in the hierarchical structure (Figure 9) relative to their corresponding higher-level indicators. The consistency test results for the judgment matrix are as follows. The consistency test indicated that the difference between the order n of each matrix and its maximum eigenvalue λ was within the permissible range, indicating that the interrelationships among influencing factors in the constructed judgment matrix were relatively consistent.

Table 10. Scoring method for judgment matrix criteria 1 to 9.

Scale	Definition and Explanation
1	Two elements are of equal importance to a certain attribute.
3	When comparing the two elements, the former one is slightly more important than the latter.
5	When comparing the two elements, the former is clearly more important than the latter.
7	When comparing the two elements, the former is significantly more important than the latter.
9	When comparing two elements, the former element is extremely important than the latter.
2, 4, 6, 8	The importance of the preceding element over the subsequent one falls between that of the calibrated standard.
1/ai,j	The reverse comparison of two elements.

lists the values and meaning of each factor in the judgment matrices.

(2)

Weighting Results

This study employed the AHP (subjective weighting), CRITIC (objective weighting), and combined weighting methods to determine the weights for each evaluation indicator of lettuce appearance and intrinsic quality. The results are presented in

Table 11. Objective, subjective, and combined weighting methods.

Primary Index	Secondary Index	AHP		CRIYIC		AHP-CRIYIC
T1	C1	0.052	0.113	0.226	0.535	0.139	0.324
	C2	0.036		0.162		0.099
	C3	0.025		0.147		0.086
T2	C4	0.105	0.887	0.137	0.465	0.121	0.676
	C5	0.171		0.097		0.134
	C6	0.223		0.131		0.177
	C7	0.388		0.100		0.244

. Among the appearance quality indicators, a*_Avg (C1) had a higher weight, whereas MC_5 (C3) generally received a lower weight. Regarding internal quality, the emphasis on vitamin C (C7) and chlorophyll (C4) content varied significantly across the methods. Chlorophyll content received notably higher weighting in the objective weighting method, whereas soluble protein content (C5) maintained a relatively stable weighting. Overall, the AHP method underscores appearance quality, whereas the CRITIC method gives increased prominence to indicators with increased data variability. The results of the combined weighting method were intermediate, effectively integrating the subjective and objective information.

(3)

Scores and Grading Results

Based on the composite scores obtained from different weighting methods, the samples were classified into special, first, and second grades through k-means clustering analysis, and the results are presented in

Table 12. K-means clustering results.

Method of Weight Assignment	Index	Special Grade	First Grade	Second Grade
	Final Clustering Center	54.23	42.76	20.59
AHP	Score range	>48.48	32.39–48.48	<32.39
	Virtual cluster number	45	105	60
	Final clustering center	193.46	140.69	52.07
CRIYIC	Score range	>166.59	98.14~166.59	<98.14
	Virtual cluster number	47	103	60
	Final clustering center	124.00	92.11	36.81
AHP-CRIYIC	Score range	>108.00	66.40–108.00	<66.40
	Virtual cluster number	46	103	61

. The classification criteria and sample distributions obtained using different weighting methods exhibited significant differences. For example, the CRITIC method applied the strictest criteria for special grades, whereas the AHP and composite weighting methods resulted in relatively similar thresholds. Regarding the sample distribution across grades, the CRITIC method assigned the highest proportion of samples to the first grade, whereas the AHP and composite weighting methods retained more samples from the special and second grades.

3.3.5. Consistency Analysis of Lettuce Grading Standards

The grading results of the two methods exhibited significant differences, as shown in Figure 10. The combined weighting method demonstrated good consistency with the industry standards, achieving a kappa coefficient of 0.788. Although it exhibited a few errors in the misclassification of special- and first-grade samples, the method achieved optimal accuracy for second-grade samples by correctly identifying all 59 cases. In contrast, the principal component analysis method resulted in a kappa coefficient of only 0.512. This produced numerous cross-grade misclassifications in the first-grade samples, misclassifying 69 as special grade and 13 as second grade, resulting in an unsatisfactory grading performance. Although industry standards focus on appearance, this study prioritizes intrinsic quality. The combined weighting method balances the weights of the internal and external indicators by integrating subjective and objective information, thereby effectively aligning with the principle of intrinsic evaluation. Consequently, it demonstrates superior performance in balancing standardization and rationality.

4. Discussion

4.1. Effect of Image Phenotypic Features and Modeling Methods on Monitoring the Appearance Quality of Greenhouse-Grown Lettuce

In this study, the Ridge regression model outperformed nonlinear models such as the BP neural network in monitoring fresh weight and leaf area owing to the strong linear allometric growth relationship between the two-dimensional morphological characteristics and the visual quality of lettuce under controlled environmental conditions. The selected MC_2 directly indicates the plant’s light interception capacity, which forms the basis for biomass development. MC_5 and a*_Avg were effectively adjusted for plant density and leaf color health, respectively. By incorporating a regularization term, Ridge regression reduces overfitting risks caused by weak feature coupling more effectively than ordinary multiple linear regression, while maintaining model interpretability [35]. Our findings indicate that, after feature selection [36], a structurally simple linear model is sufficient to capture crop growth patterns in controlled environments, providing reduced computational costs and increased practical applicability. However, relying solely on RGB two-dimensional images to estimate biomass retains its inherent physical limitations. This study focused on lettuce at harvest maturity, when canopy development was complete and leaf overlap/shading was most severe. Although the model incorporates an MC_5 feature for correction, the 2D projected area inherently cannot penetrate the surface to capture 3D volume information representing the vertical height and internal head density. This leads to prediction errors for samples with similar projected areas but notably different internal compactness. Future work can incorporate depth cameras (RGB-D) or LiDAR technology to acquire point cloud data, supplementing three-dimensional geometric information to overcome the accuracy limitations of two-dimensional imaging in analyzing highly closed canopy plant phenotypes.

4.2. Effect of Spectral Preprocessing, Feature Extraction, and Modeling Methods on Lettuce Internal Quality Inversion

Hyperspectral technology enables the effective inversion of the intrinsic quality of lettuce. The results indicated that RF models combined with CARS or SPA feature extraction outperformed the AdaBoost and Bagging models overall. Owing to the differing biochemical characteristics and spectral response mechanisms among quality indicators, optimal inversion strategies vary considerably. This variation primarily results from differences in the physical properties of the data signals and noise sources. Specifically, the chlorophyll RF model achieved the highest accuracy under SG smoothing preprocessing because SG smoothing effectively removes high-frequency random noise while preserving the steep spectral characteristics in the 680–760 nm red edge region. This region, where chlorophyll absorption peaks intersect with cellular structure scattering, is highly sensitive to content variation. This finding is consistent with those of studies by Mu [37] and Jiang et al. [38] on field crops. In contrast, this study observed that the MSC and SNV pretreatments outperformed SG in terms of soluble sugars and proteins. This advantage is likely attributable to the superior correction of the light-scattering effects resulting from the uneven surface of the lettuce leaves. From a physiological perspective, chlorophyll exhibits strong electronic transition absorption and reflection characteristics in the visible and red edge regions, creating a distinctive red edge effect with a high signal intensity and sharp features [39]. In contrast, internal constituents, such as sugars, proteins, and vitamin C, lack chromophores. Their spectral responses primarily depend on the second- and sum-frequency vibrations of the C–H and O–H bonds in the near-infrared region [40], representing weak absorption signals that are easily masked by background noise, and the inversion accuracy for vitamin C is slightly lower than that of other indicators. This is primarily due to the substantially high water content of fresh lettuce samples, which leads to strong absorption bands in the near-infrared region [41], easily obscuring the weak overtone signals of the O–H and C–H bonds in vitamin C. Future studies should investigate deep-learning algorithms to identify more complex nonlinear spectral features, thereby enhancing the sensitivity and detection thresholds of trace components.

Therefore, the combination of RGB imaging and hyperspectral technology enables the simultaneous assessment of both the external and internal qualities of lettuce. RGB imaging offers a distinct advantage in spatial resolution and precise edge extraction. The accurate determination of appearance quality indicators for lettuce (such as MC_1 to MC_7) demands exceptionally high image clarity and edge sharpness. Conventional RGB cameras can readily capture ultra-high-resolution images, and this high spatial resolution provides a highly reliable data foundation for precisely capturing the complex edge contours and minute morphological features of lettuce leaves [42,43]. This is essential for ensuring the accuracy of appearance parameter calculations. Furthermore, differences exist in the data acquisition methods of the two imaging sensors. Hyperspectral imaging commonly employs push-broom line-scan technology. During this process, leaf tremors induced by minor vibrations from mechanical conveyors or environmental airflow [44] can lead to image drift, spatial distortion, and motion artifacts. In contrast, RGB cameras employ instantaneous fixed-point acquisition, completing the imaging process within milliseconds. This effectively eliminates spatial errors caused by mechanical and environmental jitter, ensuring high stability and consistency in feature acquisition.

4.3. Effect of Subjective-Objective Combined Weighting Method on Lettuce Quality Grading

For developing a comprehensive lettuce quality evaluation method, the AHP-CRITIC combined weighting approach notably outperformed the PCA-based method. This disparity is primarily attributed to the pronounced imbalance in variability between harvest-stage lettuce morphological indicators and intrinsic nutritional metrics. The statistical data revealed that the coefficient of variation for MC_2 was 58.49%, whereas the numerical fluctuations in the intrinsic quality indicators were relatively minor. The PCA method primarily determines weights based on the variance contribution rates [45], which can result in the evaluation results being heavily influenced by discrete appearance indicators. This approach does not consider intrinsic nutritional indicators with minimal variation but equal importance in determining quality. In contrast, the AHP-CRITIC method employed in this study effectively addressed weight imbalance caused by varying indicator variability by introducing the Analytic Hierarchy Process (AHP) to assign reasonable baseline weights to intrinsic quality, while utilizing the Critical Information Retention Criterion (CRITIC) to preserve objective data information [46]. However, the AHP component of this evaluation system remains reliant on expert scoring, with weight assignments constrained to a certain extent by the evaluators’ subjective experience and knowledge background. Future research may consider the incorporation of market transaction price data or large-scale consumer sensory evaluation data to construct a dynamic weight-adjustment model based on market feedback. This enables the evaluation method to indicate the market’s actual demand for facility-grown lettuce more objectively.

5. Conclusions

This study used RGB images and hyperspectral data to rapidly monitor both the external morphology and internal quality of greenhouse-grown lettuce, thereby developing a comprehensive scientific quality evaluation method. The key conclusions are as follows:

(1)

The Ridge regression model based on RGB image features accurately monitored the fresh weight and leaf area of greenhouse lettuce, resulting in test set determination coefficients (R²) of 0.897 and 0.880, respectively.

(2)

Hyperspectral inversion models were developed separately for each intrinsic quality parameter of lettuce. For chlorophyll content, the optimal model was SG-SPA-RF, which yielded the highest inversion accuracy with an R² value of 0.958. The inversion models for soluble sugar, soluble protein, and vitamin C content all achieved R² values exceeding 0.89.

(3)

A comprehensive quality evaluation method for greenhouse-grown lettuce based on AHP-CRITIC combined weighting was developed. The consistency of its classification results with industry standards achieved a Kappa coefficient of 0.788. It also provides an effective theoretical basis and technical support for the intelligent grading and precise management of greenhouse vegetables.