Evaluation and Early Detection of Downy Mildew of Lettuce Using Hyperspectral Imagery

Abstract

This study combines hyperspectral imaging technology with biochemical parameter analysis to facilitate the disease severity evaluation and early detection of lettuce downy mildew. The results reveal a significant negative correlation between the disease index (DI) and the levels of flavonoids (r = −0.523) and anthocyanins (r = −0.746), indicating the role of these secondary metabolites in enhancing plant resistance. Analysis of hyperspectral data identified that spectral regions (410–503 nm, 510–615 nm, and 630–690 nm) and vegetation indices like PRI and ARI2 were highly correlated with DI, flavonoids, and anthocyanins, providing potential spectral indicators for disease assessment and early detection. Moreover, regression models developed using Partial Least Squares (PLS), Random Forest (RF), and Convolutional Neural Network (CNN) algorithms demonstrated high accuracy and reliability in predicting DI, flavonoids, and anthocyanins, with the highest R² of 0.857, 0.910, and 0.963, respectively. The classification model using PLS, RF, and CNN successfully detected early physiological changes in lettuce within 24 h post-infection (highest accuracy = 0.764), offering an effective tool for early disease detection. The key spectral parameters in the PLS-DA model, like PRI, also demonstrated strong correlations with DI. These findings provide a scientific basis and practical tools for managing lettuce downy mildew and resistance breeding while laying a foundation for broader applications of hyperspectral imaging in plant pathology.

1. Introduction

Downy mildew, caused by the obligate pathogen Bremia lactucae, is a highly destructive disease that poses a significant threat to lettuce production worldwide [1]. This disease is primarily spread through windborne spores and rain splash, and it thrives under humid conditions, making it particularly challenging to control in regions with wet climates [2]. Infected lettuce plants typically exhibit symptoms such as yellowing, necrosis, and eventual defoliation, leading to substantial yield losses of up to 80% and reduced market value due to compromised quality [3,4]. The management of downy mildew is further complicated by the evolving nature of pathogen races, which can quickly render existing control measures ineffective [5].

Breeding resistant cultivars is considered a sustainable and effective approach for disease control [6,7]. However, the success of such breeding programs is often limited by the lack of rapid and objective methods to evaluate disease resistance [8]. Traditional methods for resistance breeding rely on the assessment of disease symptoms at advanced stages of infection, which involves manually estimating the infected leaf area, counting diseased leaves, and calculating the disease index (DI) [9]. These methods are not only labor-intensive and time-consuming but also highly susceptible to human bias, potentially delaying breeding progress and reducing evaluation consistency. Moreover, diseases such as downy mildew exhibit a latent period of approximately 5–7 days before visible symptoms emerge [10]. This delay in symptom manifestation poses a significant challenge for early disease detection, limiting timely intervention and hindering efficient resistance screening in breeding programs. Therefore, two critical needs arise: (1) the development of efficient and objective methods for DI assessment to enhance accuracy and reliability, and (2) the capability for early disease prediction, enabling proactive disease management and accelerating the identification of resistant genotypes in breeding programs [11,12].

Recent advancements in high-throughput plant phenotyping (HTPP) platforms have revolutionized the ability to assess plant traits non-destructively and efficiently. By integrating optical sensors, these platforms enable the simultaneous capture of morphological, physiological, and biochemical traits [13,14]. Among these technologies, hyperspectral imaging has garnered significant attention due to its ability to collect high-resolution spectral data across a wide range of wavelengths, including the visible (VIS) and near-infrared (NIR) regions [15]. In plant disease monitoring, the VIS region provides direct visual information on disease symptoms, such as chlorosis, necrosis, and lesion development, which are critical for conventional disease assessment. Meanwhile, the NIR region detects subtle physiological and structural changes in plant tissues that are often imperceptible to the human eye, enabling the identification of early-stage infections before visible symptoms appear [15,16]. This unique advantage makes hyperspectral imaging an invaluable tool for both accurate disease index (DI) assessment and early disease detection.

The utility of hyperspectral imaging in plant disease research has been well documented. Studies have shown its effectiveness in detecting and assessing downy mildew in grape, wheat, cucumbers, and brassica [17,18,19,20], powdery mildew in wheat and rubber trees [21,22], fusarium head blight in barley and wheat [23,24], late blight in potato and tomato [25,26], etc. These studies demonstrated the potential of hyperspectral imaging to capture spectral features associated with physiological and biochemical changes in crops, enabling precise identification of disease presence and severity. For example, diseases like powdery mildew in wheat, downy mildew in grapevine and cucumbers, and late blight in potatoes have been successfully detected using spectral reflectance data and vegetation indices. Hyperspectral imaging was able to detect subtle changes in leaf pigments, structural traits, and moisture content before visible symptoms appeared in these studies. By integrating vegetation indices and spectral features into machine learning models such as Random Forest (RF), Support Vector Machine (SVM), and deep learning architectures such as Convolutional Neural Network (CNN), researchers have achieved high classification accuracies, often exceeding 90% [17,18,19,20,21,22,23,24,25,26,27]. These tools not only offer rapid and non-destructive solutions for disease monitoring but also provide insights into the underlying mechanisms of disease progression and host resistance. Despite progress in other crops, the application of HSI to lettuce downy mildew resistance breeding remains underdeveloped.

Secondary metabolites, such as flavonoids and anthocyanins, play crucial roles in plant defense mechanisms against pathogens [28]. Flavonoids are extensively studied for their regulatory functions in plant development and interactions with commensal microbes [29]. Anthocyanins have been implicated in enhancing plant tolerance to various stressors, including pathogen infections [30]. Additionally, both anthocyanins and flavonoids are known to be involved in plant protection against pathogens [31]. In lettuce, higher flavonoid content has been correlated with enhanced disease resistance [32]. However, the precise relationship between flavonoids and downy mildew resistance remains underexplored. Furthermore, the relationship between hyperspectral features, biochemical resistance markers (e.g., flavonoids and anthocyanins), and disease progression has not been systematically investigated. This knowledge gap hinders the translation of spectral data into actionable breeding metrics, as current models lack links to host–pathogen interactions. Therefore, exploring the relationship between spectral data, biochemical traits, and disease progression is essential to uncover the biological significance of spectral features closely associated with downy mildew. By establishing the spectral basis of disease resistance, this study aims to provide a deeper mechanistic understanding of plant–pathogen interactions while paving the way for the development of automated tools for high-throughput resistance screening.

Therefore, this study employs hyperspectral imaging technology to capture spectral changes in lettuce during the progression of downy mildew infection. By analyzing the correlations between the disease index (DI) and biochemical parameters such as flavonoid and anthocyanin contents, regression models based on Partial Least Squares (PLS), RF, and CNN algorithms were developed, enabling non-contact, high-throughput quantitative estimation of disease severity. For the first time, this study systematically identifies the spectral characteristics and key wavelengths associated with early-stage downy mildew infection, as well as critical vegetation indices, providing a scientific basis for rapid and accurate disease diagnosis. The integration of hyperspectral imaging with machine learning algorithms represents a novel, non-destructive approach to disease detection, while also offering insights into the biochemical and physiological responses of lettuce to downy mildew. This approach significantly enhances the ability to screen for disease resistance traits in lettuce germplasm, accelerating the breeding of disease-resistant cultivars. By providing tools for early warning and precision control of lettuce diseases, this research not only contributes to sustainable crop management but also lays a solid foundation for the application of hyperspectral technology in breeding programs aimed at improving lettuce resilience to downy mildew and other biotic stresses.

2. Materials and Methods

2.1. Experiment Design

This study consisted of two experiments. Experiment 1 aimed to assess the severity of downy mildew across different lettuce accessions using hyperspectral imaging. Experiment 2 focused on the early detection of downy mildew in lettuce through hyperspectral imaging. The details are as follows:

Experiment 1: A total of 150 lettuce accessions were selected from the germplasm repository of the Shanghai Agrobiological Gene Center. For each accession, 30 plants were initially cultivated in plant pots under controlled greenhouse conditions, maintained at a temperature range of 15–20 °C. At the five-leaf and one-heart stage, 15 plants per accession exhibiting uniform growth and free from visible abnormalities were selected for the experiment. The selected plants were inoculated with downy mildew pathogen spores using a standardized spray inoculation method. The spore suspension was evenly applied to the true leaves of each plant until the inoculum began to drip off the leaf surfaces, ensuring uniform coverage. To minimize cross-contamination, the inoculation process was conducted in a controlled environment, and the plants were isolated during the initial post-inoculation period. Following inoculation, the plants were returned to the greenhouse, where environmental conditions, including temperature and humidity, were carefully monitored and maintained to support both pathogen development and plant growth. Additional measures, such as ensuring adequate ventilation, were implemented to prevent secondary infections or the unintended spread of pathogens within the greenhouse. Seven days post-inoculation, hyperspectral images of all the inoculated lettuce were acquired, the disease index (DI) for each accession was investigated, and the biochemical parameters, including chlorophyll, flavonoid, and anthocyanin contents, were measured.

Experiment 2: Two downy-mildew-susceptible lettuce cultivars were selected, with 36 plants assigned to each cultivar. For each cultivar, the plants were divided into two groups: a treatment group (T) and a control group (CK), with 18 plants per group. All plants were cultivated in plant pots under identical environmental conditions in a greenhouse to minimize variability. At the five-leaf and one-heart stage, plants in the treatment group (T) were sprayed with a suspension of downy mildew spores to induce infection, while plants in the control group (CK) were sprayed with purified water to serve as a control. To ensure uniform application, a standardized spray technique was used for both groups, and the plants were briefly isolated to prevent cross-contamination during the inoculation process. Following inoculation, all plants were kept stationary for 24 h in a controlled environment to allow for proper pathogen establishment. Hyperspectral images of all plants were captured one day before inoculation (Day 0) to establish baseline measurements. After inoculation, hyperspectral imaging was performed daily for seven consecutive days (Day 1 to Day 7) to monitor spectral changes over time. By Day 7, visible symptoms of downy mildew infection were observed in the treatment group (T), while the control group (CK) remained symptom-free (Figure 1).

2.2. Data Acquisition

2.2.1. Hyperspectral Image Acquisition

In both experiments, hyperspectral images of lettuces were captured in a darkroom environment. The lettuces were placed on a black light-absorbing background cloth to minimize spectral interference. The hyperspectral imaging system (Figure 2) used in this study consisted of three main components: a hyperspectral imager (Pika L), halogen light sources, an automated translation platform, and a computer. The Pika L imager (Resonon Inc., Bozeman, MT, USA) operates as an external push-broom sensor with a spectral range of 400–1000 nm, a spectral resolution of 2.1 nm, a total of 281 spectral bands, and 900 spatial channels. The spatial resolution of the system allows for the capture of 900 spatial channels per scan. The halogen light sources provide uniform illumination, minimizing the influence of external light interference during imaging. The motorized translation stage facilitates the movement of the imaging system to scan the entire area of interest, ensuring complete and consistent data acquisition. The computer controls the whole system. A Spectralon panel was placed alongside the lettuce to provide a white reference for each spectral scan.

2.2.2. DI Investigation

In Experiment 1, seven days post-inoculation, the disease symptoms of the 15 inoculated plants in each accession were investigated by counting the number of diseased leaves by disease severity grade based on the standardized disease grading scale (

Table 1. Standard for grading downy mildew disease severity.

Disease Severity Grade	Symptom
0	Asymptomatic
1	Slight necrosis at the inoculation site
3	The necrotic spots were obvious and less than 0.5 cm in diameter
5	Necrotic spots account for less than 1/3 of the leaf area
7	The area of necrotic spots accounts for 1/3 to 2/3 of the leaf area
9	Necrotic spots accounted for more than 2/3 of the leaf area, and even dried up

). Then, the disease index (DI) for each accession was calculated using Formula (1).

$$\mathrm{DI}={\displaystyle \frac{\sum \left(i\times n_i\right)}{9\times N}}\times 100$$

(1)

where N: total number of plants investigated; i: disease severity grade; n_i: number of plants in the corresponding disease severity grade.

2.2.3. Biochemical Parameter Measurement

In Experiment 1, biochemical parameters, including chlorophyll (Chl), flavonoid (Flav), and anthocyanin (Anth) contents, were measured using a chlorophyll–polyphenol meter (Dualex Scientific+, Force-A Co., Orsay, France) after the hyperspectral images acquisition. The device provides dimensionless readings (Chl, Flav, and Anth) that represent the relative contents of chlorophyll, flavonoid, and anthocyanin in plant leaves. Measurements were conducted by securely holding each lettuce leaf with the device’s leaf clip and recording readings at two distinct positions on the leaf to account for potential variability in biochemical content in the leaf. All leaves of the 15 lettuce plants in each accession were measured. The average Chl, Flav, and Anth values for each lettuce accession were determined by calculating the mean values across all measured leaves within that accession. These averaged values were utilized as the biochemical parameters in the subsequent analyses.

2.3. Data Process

The raw hyperspectral images were converted to reflectance data using SpectrononPro software (v3.1.1, Resonon Inc., Bozeman, MT, USA) based on the spectrum of the Spectralon panel. Then the reflectance images were processed using ENVI software (v5.6, NV5 Geospatial Inc., Broomfield, CO, USA). A decision tree classification algorithm was applied to segment the lettuce plants from the background in the hyperspectral images. This method enabled the accurate extraction of pure lettuce plant regions. For each plant, the spectral reflectance of all pixels within the segmented area was averaged to obtain a representative spectral reflectance curve for subsequent analysis (Figure 3). In Experiment 1, each lettuce accession was treated as a single sample, correlating with its measured disease index (DI) and biochemical parameters. The spectral reflectance curve for each accession was obtained by averaging the spectral reflectance data from all 15 plants within that accession, resulting in a total of 150 spectral reflectance curves. In Experiment 2, individual plants served as separate samples for further analysis.

To reduce the influence of noise inherent in the hyperspectral imaging system at the edges of its spectral range, only the spectral reflectance data within the range of 410–900 nm were selected for further analysis. The spectral reflectance data were then processed using multiplicative signal correction (MSC) to reduce the influence of spectral noise and baseline shift [33].

Vegetation indices (VIs) are transformations of two or more spectral bands designed to enhance sensitivity to vegetation while minimizing confounding factors such as soil background reflectance. In this study, 34 vegetation indices were tested, and 13 indices (

Table 2. Vegetation indices used in this study.

Vegetation Index	Acronym	Equation	Reference
Green Normalized Difference Vegetation Index	GNDVI	(R780 − R550)/(R780 + R550)	[34]
Green Ratio Vegetation Index	GRVI	R780/R550	[35]
Anthocyanin Reflectance Index 1	ARI1	1/R550 − 1/R700	[36]
Anthocyanin Reflectance Index 2	ARI2	R800 (1/R550 − 1/R700)	[36]
Photochemical Reflectance Index	PRI	(R531 − R570)/(R531 + R570)	[37]
Atmospherically Resistant Vegetation Index	ARVI	(R780 − (2R680 − R450))/(R780 + (2R680 − R450))	[38]
Carotenoid Reflectance Index 2	CRI2	1/R510 − 1/R700	[39]
Difference Vegetation Index	DVI	R780 − R680	[40]
Enhanced Vegetation Index	EVI	2.5((R780 − R680)/(R780 + 6R680 − 7.56R450 + 1))	[41]
Modified Red Edge Simple Ratio Index	mSR705	(R750 − R445)/(R705 + R445)	[42]
Red Edge Normalized Difference Vegetation Index	NDVI705	(R750 − R705)/(R750 + R705)	[43]
Normalized Phaeophytinization Index	NPQI	(R415 − R435)/(R415 + R435)	[44]
Plant Senescence Reflectance Index	PSRI	(R680 − R500)/R750	[45]

Note: R represents the reflectance value in specified bands.

) closely related to downy mildew and biochemical parameters were ultimately selected for analysis.

2.4. Data Analysis

2.4.1. Pearson Correlation Analysis

Correlation analysis was conducted using Pearson’s correlation coefficient (r) to explore the relationships between the disease index (DI), flavonoid content, anthocyanin content, and spectral parameters.

Pearson’s correlation analysis is a statistical method that measures the linear correlation between two continuous variables, and it is widely used to determine the strength and direction of the relationship between variables [46]. In our analysis, we calculated the r to quantify the degree of linear dependence between each pair of variables. The r ranges from −1 to 1, where values close to −1 or 1 indicate a strong negative or positive linear relationship, respectively, and values around 0 suggest no linear relationship. In this study, the threshold for statistical significance of correlations was set at p < 0.01.

This analysis aims to identify biochemical and spectral parameters strongly associated with lettuce DI, as well as to explore the interrelationships between these parameters. This approach provides valuable insights into the physiological and biochemical changes underlying downy mildew infection in lettuce. Additionally, it establishes a robust statistical foundation for subsequent regression modeling to estimate lettuce DI and biochemical parameters from spectral data.

2.4.2. Regression Analysis

Partial Least Squares (PLS), Random Forest (RF), and Convolutional Neural Network (CNN) were employed to construct regression models for estimating the DI and biochemical parameters of lettuce. The goal was to achieve non-contact, rapid estimation of these metrics through spectral analysis, highlighting the potential of both linear and nonlinear algorithms in this context.

PLS regression is a widely used statistical method in chemometrics and multivariate data analysis, particularly well suited for examining relationships between dependent and independent variables in the presence of multicollinearity [47]. PLS constructs a linear model between the independent variable matrix X and the dependent variable matrix Y, taking into account the internal structure of both datasets. This method is especially effective for analyzing datasets with noisy, collinear, or even incomplete variables. Unlike traditional regression approaches, PLS benefits from the inclusion of a larger number of relevant variables and observations, which can improve the precision of the model parameters. The PLS algorithm works by extracting principal components, referred to as latent variables (LVs), from X and Y. These LVs serve to reduce the dimensionality of the data while retaining the most relevant information. The reduced dimensionality enables the construction of a robust multiple linear regression model, which enhances predictive performance, particularly when dealing with small sample sizes or high-dimensional datasets. PLS integrates key features of Principal Component Analysis (PCA) and Canonical Correlation Analysis (CCA), offering results that are both interpretable and rich in analytical depth. Cross-validation was conducted to optimize the number of latent variables, ensuring both the accuracy and generalizability of the model.

RF regression is an ensemble machine learning technique renowned for its robustness and versatility in modeling complex relationships within datasets [48]. It is an extension of the Classification and Regression Tree (CART) method, constructing multiple decision trees during training and aggregating their predictions, typically through averaging, to enhance accuracy and mitigate overfitting. RF excels in handling large datasets with numerous features, including those characterized by high dimensionality and complexity. It is inherently resilient to outliers and requires minimal parameter tuning, with the primary adjustable hyperparameter being the number of trees (n_estimators) in the forest. Each tree is built using a bootstrap sample (random subset) of the training data, and at each split, a random subset of features is considered. This randomness promotes model diversity and improves generalization performance. In this study, RF regression was utilized to predict the Disease Index (DI) and biochemical parameters of lettuce based on hyperspectral data. The number of trees was set as 100 with a minimum leaf size of 3. The TreeBagger function was used in the regression mode to train the model.

CNN regression is a deep learning approach particularly effective for capturing complex spatial and spectral patterns within high-dimensional data. CNNs automatically learn hierarchical representations from raw input data, making them well suited for hyperspectral image analysis [49]. A typical CNN architecture consists of multiple layers, including convolutional layers, pooling layers, and fully connected layers. The convolutional layers apply learnable filters to extract local spectral features, while pooling layers reduce dimensionality and computational complexity, preventing overfitting. The extracted feature maps are then flattened and passed through fully connected layers to generate final predictions. To enhance model performance and stability, batch normalization and dropout regularization were implemented. In this study, the CNN architecture comprised three convolutional layers with ReLU activation functions, followed by max-pooling layers, and two fully connected layers leading to the output. The key hyperparameters of the CNN model included the number of filters, set to 64, 128, and 256 for the first, second, and third convolutional layers, respectively. All convolutional layers utilized a filter size of 3 × 3 to effectively capture local spectral features. The model was trained using the Adam optimizer with a learning rate of 0.001 for 100 epochs, employing a batch size of 32 to ensure efficient learning and stable convergence.

PLS is a linear modeling approach, making it well suited for scenarios where a linear relationship exists between variables. In contrast, RF and CNN are nonlinear modeling methods that excel at capturing complex, nonlinear relationships between variables. By comparing these three models, we aim to evaluate their respective strengths and applicability in the non-contact, rapid estimation of lettuce disease index (DI) and associated biochemical parameters.

The performance of the PLS and RF models was assessed using two key metrics: the coefficient of determination (R²) and root mean square error (RMSE). These metrics quantify the models’ predictive accuracy and reliability. An ideal model exhibits a higher R² value alongside lower RMSE values, indicating superior precision and accuracy. The calculation formulas for R² and RMSE are provided as follows:

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

(2)

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

(3)

where $y_i$ is the investigated DI or measured biochemical parameters, ${\widehat{y}}_i$ is the predicted DI, $\overline{y}$ is the average of investigated DI, and n is the number of samples.

2.4.3. Discriminant Analysis

In this study, PLS Discriminant Analysis (PLS-DA), RF, and CNN were employed to develop classification models for distinguishing between infected (T) and healthy (CK) lettuce. These models enable the early detection of downy mildew infection based on hyperspectral characteristics, facilitating a non-invasive and efficient approach to plant disease diagnosis.

PLS-DA is a multivariate statistical method used for supervised discriminant analysis on the basis of PLS [50]. It operates by projecting both the predictor variables (X) and the response variables (Y) into a new space, with the aim of maximizing the covariance between the projected components of X and Y. This approach allows for the identification of patterns that distinguish between different groups, making PLS-DA a powerful tool for classification tasks. To further evaluate the importance of spectral variables, the Variable Importance in Projection (VIP) scores are calculated. VIP scores measure the contribution of each spectral variable to sample classification and help identify key spectral parameters for discrimination. A VIP score greater than 1.0 is typically used as the threshold for selecting significant spectral features.

In RF classification, multiple decision trees are constructed using bootstrap samples of the training data, and the final class prediction is determined through majority voting among the trees [51]. In this study, the RF classification model uses 100 decision trees, each with a minimum leaf size of 3 to balance overfitting. Out-of-bag (OOB) error estimation is enabled for performance evaluation. The model is trained using the TreeBagger function with classification mode.

CNN-based classification leverages deep learning to automatically extract hierarchical spectral features from hyperspectral data [52]. The CNN classification model used in this study consists of two convolutional layers, each with 16 and 32 filters of size 2 × 1, followed by ReLU activation functions and max-pooling layers (2 × 1, stride 2 × 1). A fully connected layer is used before the Softmax classification layer. The model is trained using the Adam optimizer with an initial learning rate of 0.001, a maximum of 100 epochs, and L2 regularization (1 × 10⁻⁴). Additionally, a learning rate decay strategy is applied, reducing the learning rate by a factor of 0.1 every 80 epochs, to enhance model stability and generalization.

To assess the classification performance of various discriminant models, confusion matrices were employed to analyze their effectiveness across different classes. These matrices offer a comprehensive view of the models’ predictive accuracy, highlighting areas where misclassifications occur. By mapping predicted classes against actual classes, confusion matrices help identify instances where the model struggles to distinguish between categories. The performance of the detection models was evaluated using three key metrics: accuracy, sensitivity, and specificity, all derived from the confusion matrices. These widely used indicators assess the effectiveness of classification models in both parameter space and feature extraction. Accuracy represents the overall classification performance, defined as the ratio of correctly classified samples to the total number of samples. Sensitivity measures the model’s ability to correctly identify positive samples, calculated as the ratio of correctly classified positive samples to the total actual positive samples. Specificity reflects the model’s ability to correctly identify negative samples, determined by the ratio of correctly classified negative samples to the total actual negative samples. An ideal classification model should achieve high accuracy, sensitivity, and specificity, with values approaching 1 indicating optimal performance. These metrics are calculated using the following formulas:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}}}$$

(4)

$$\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(5)

$$\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}}}$$

(6)

where TP denotes the number of true positives (correctly classified infected samples), TN denotes the number of true negatives (correctly classified healthy samples), FP denotes the number of false positives (incorrectly classified healthy samples), and FN denotes the number of false negatives (incorrectly classified infected samples).

3. Results

3.1. Correlation Analysis Between DI and Biochemical Parameters

Pearson correlation analysis was conducted to evaluate the linear relationships between the Disease Index (DI) and various biochemical parameters. The results (Figure 4) revealed significant negative correlations between DI and both Flav and Anth, with r values of −0.523 and −0.746, respectively. Additionally, a significant positive correlation was observed between Flav and Anth, with an r of 0.707. These findings indicate a strong association between the severity of downy mildew symptoms and the concentrations of these phenolic compounds, with lettuce accessions exhibiting higher flavonoid and anthocyanin levels showing milder disease symptoms, highlighting the potential role of flavonoids and anthocyanins in conferring resistance to downy mildew in lettuce. In contrast, no significant correlation was observed between DI and Chl (r = 0.0498), indicating that at this stage of infection, chlorophyll levels measured by the Dualex device were not associated with DI. The correlations between Chl and both Flav and Anth were also found to be non-significant. Consequently, Chl was excluded from further analysis due to its limited relevance to DI.

3.2. Correlation Analysis Between DI, Biochemical Parameters, and Spectral Parameters

3.2.1. Spectral Reflectance

The correlation analysis between DI, Flav, Anth, and spectral reflectance at different wavelengths reveals significant relationships in three common wavelength ranges (Figure 5). Consistent with the analysis results in Section 3.1, the direction of correlation between DI and spectral reflectance is opposite to that of Flav and Anth in these ranges. Specifically, in the 410–503 nm and 630–690 nm ranges, DI exhibits a significant negative correlation with spectral reflectance, while Flav and Anth show significant positive correlations; conversely, in the 510–610 nm range, DI is significantly positively correlated with spectral reflectance, whereas Flav and Anth are significantly negatively correlated.

Further analysis identifies wavelength intervals where the absolute correlation coefficient exceeds 0.5. For DI, these intervals are 508–600 nm, with a peak correlation at 550 nm (r = 0.673), and 634–696 nm, with a peak at 655 nm (r = −0.756). For Flav, the intervals are 510–600 nm, peaking at 540 nm (r = −0.642), and 640–689 nm, with a maximum at 655 nm (r = 0.646). For Anth, the intervals are 510–600 nm, with the strongest correlation at 540 nm (r = −0.747), and 630–700 nm, with a peak at 650 nm (r = 0.847). These results highlight the distinct correlation patterns of DI, Flav, and Anth with spectral reflectance, suggesting their potential as indicators for monitoring the severity of the downy mildew and biochemical characteristics.

3.2.2. Vegetation Indices

The correlation between vegetation indices and DI, Flav, and Anth was analyzed, and Figure 6 shows the indices with r higher than 0.5. The results indicate that the direction of correlation between DI and VI is opposite to that between Flav, Anth, and VI. Specifically, DI exhibits a significant positive correlation with PRI (r = 0.679), while Flav (r = −0.64) and Anth (r = −0.831) are significantly negatively correlated with PRI. Conversely, DI shows significant negative correlations with ARI1 (r = −0.724), ARI2 (r = −0.736), GRVI (r = −0.711), and GNDVI (r = −0.648). In contrast, Flav and Anth exhibit significant positive correlations with these four indices, with Flav showing correlation coefficients ranging from 0.65 to 0.66, and Anth displaying a correlation of 0.724 with GNDVI and coefficients between 0.84 and 0.87 with ARI1, ARI2, and GRVI. These findings underscore the contrasting roles of DI, Flav, and Anth in their relationships with vegetation indices, further emphasizing their potential to differentiate physiological and biochemical characteristics in lettuce downy mildew monitoring.

3.3. Determination of DI, Chl, Flav, Anth Based on Spectral Parameters

Estimation models for DI, Flav, and Anth were constructed using PLS, RF, and CNN algorithms, with spectral reflectance (Ref) and VI as independent variables. The results, as shown in

Table 3. Performance of PLS, RF, and CNN models for estimating DI, Flav, and Anth.

Index	Spectral Parameter	Method	Training		Testing
			R2	RMSE	R2	RMSE
DI	Ref	PLS	0.631	5.297	0.586	6.699
		RF	0.857	2.813	0.748	5.861
		CNN	0.823	3.482	0.709	7.956
	VI	PLS	0.635	5.904	0.563	5.734
		RF	0.789	3.552	0.704	6.586
		CNN	0.758	5.430	0.835	6.947
Flav	Ref	PLS	0.565	0.135	0.557	0.151
		RF	0.892	0.069	0.732	0.109
		CNN	0.910	0.066	0.733	0.107
	VI	PLS	0.522	0.162	0.464	0.146
		RF	0.843	0.083	0.670	0.128
		CNN	0.738	0.145	0.711	0.176
Anth	Ref	PLS	0.818	0.044	0.819	0.061
		RF	0.963	0.019	0.881	0.032
		CNN	0.951	0.023	0.811	0.061
	VI	PLS	0.901	0.036	0.825	0.058
		RF	0.955	0.021	0.864	0.046
		CNN	0.908	0.032	0.853	0.052

, revealed significant differences in the predictive capabilities of the three models across different parameters and spectral data types.

For DI prediction, RF demonstrated superior performance with Ref inputs (test R² = 0.748, RMSE = 5.861), while CNN excelled under VI conditions (test R² = 0.835, RMSE = 6.947), significantly outperforming PLS. In Flav estimation, CNN achieved the highest test R² of 0.733 (RMSE = 0.107) with Ref inputs, surpassing PLS by 31.6% (R² = 0.557) and RF (R² = 0.732). Notably, CNN maintained robust performance with VI inputs (test R² = 0.711), outperforming both PLS (0.464) and RF (0.670). For Anth prediction, RF dominated Ref-based scenarios (test R² = 0.881, RMSE = 0.032), as well as that with VI inputs (test R² = 0.864, RMSE = 0.046). PLS exhibited moderate success in Anth-VI prediction (test R² = 0.825, RMSE = 0.058), suggesting stronger linear correlations for this parameter.

Input parameter selection critically influenced outcomes: Ref inputs favored CNN in Flav and DI predictions, whereas VI enhanced PLS’s Anth accuracy. Despite higher RMSE values in certain cases, nonlinear models consistently outperformed PLS, with RF and CNN achieving 42.9–48.5% R² improvements for DI and 31.6–53.3% gains for Flav. The superior performance of CNN and RF highlights the potential of nonlinear models in spectral analysis for non-contact, rapid estimation of lettuce disease index and biochemical parameters. The ability of CNN to capture complex spectral patterns and the robust generalization ability of RF contribute to their superior predictive accuracy.

Furthermore, as shown in Figure 7, the scatter plots of predicted versus measured DI, Flav, and Anth using RF and CNN based on Ref exhibit a tight clustering of data points around the fitting line, reflecting high predictive accuracy. However, despite the high R² values, the scatter plots of the VI-based models for DI and Flav display a markedly dispersed distribution away from the fitting line, indicating lower prediction accuracy. These findings suggest that caution should be exercised when employing VI as an input variable.

3.4. Early Detection of Downy Mildew of Lettuce

3.4.1. Early Detection Models

Based on Experiment 2, to evaluate the potential of spectral data for early detection of downy mildew in lettuce, PLS-DA, RF, and CNN classification models were constructed using full-band spectral reflectance (Ref) and various vegetation indices (VIs) as input variables. These models were developed for each observation day during the experiment to distinguish infected lettuce (T group) from healthy lettuce (CK group). The classification results and prediction accuracy of the models are presented in Figure 8 and

Table 4. Classification results of the PLS-DA, RF, and CNN models based on Ref and VI from Day 0 to Day 7.

	Method	Ref			VI
		Accuracy	Sensitivity	Specificity	Accuracy	Sensitivity	Specificity
Day 0	PLS-DA	0.542	0.556	0.528	0.486	0.472	0.500
	RF	0.486	0.500	0.472	0.514	0.556	0.472
	CNN	0.542	0.528	0.556	0.472	0.444	0.500
Day 1	PLS-DA	0.764	0.750	0.778	0.681	0.694	0.667
	RF	0.708	0.722	0.694	0.708	0.750	0.667
	CNN	0.764	0.778	0.750	0.681	0.722	0.639
Day 2	PLS-DA	0.861	0.833	0.889	0.806	0.778	0.833
	RF	0.764	0.778	0.750	0.764	0.806	0.722
	CNN	0.847	0.861	0.833	0.819	0.833	0.806
Day 3	PLS-DA	0.903	0.889	0.917	0.861	0.833	0.889
	RF	0.819	0.806	0.833	0.792	0.778	0.806
	CNN	0.903	0.889	0.917	0.861	0.861	0.861
Day 4	PLS-DA	0.944	0.944	0.944	0.903	0.889	0.917
	RF	0.875	0.861	0.889	0.847	0.833	0.861
	CNN	0.931	0.917	0.944	0.903	0.889	0.917
Day 5	PLS-DA	0.972	0.972	0.972	0.944	0.944	0.944
	RF	0.917	0.917	0.917	0.889	0.889	0.889
	CNN	0.958	0.944	0.972	0.944	0.944	0.944
Day 6	PLS-DA	1.000	1.000	1.000	0.986	0.972	1.000
	RF	0.958	0.972	0.944	0.931	0.917	0.944
	CNN	0.986	1.000	0.972	0.972	0.972	0.972
Day 7	PLS-DA	1.000	1.000	1.000	1.000	1.000	1.000
	RF	0.986	1.000	0.972	0.986	0.972	1.000
	CNN	1.000	1.000	1.000	1.000	1.000	1.000

.

On the day prior to inoculation (Day 0), the classification accuracy for both Ref- and VI-based models was below 0.6, indicating that the spectral data could not differentiate between the CK and T groups at this stage. This result suggests that no measurable spectral differences existed between the two groups before inoculation. On the first day after inoculation (Day 1), the accuracy for the models exceeded 0.65, and the misclassification number was below 25. This indicates that significant spectral differences emerged between the CK and T groups within 24 h of inoculation. These findings highlight the potential of hyperspectral data to detect early physiological changes in lettuce induced by downy mildew infection.

From Day 1 to Day 7, the accuracy of the classification models showed a consistent upward trend with the accuracy ranging from 0.708 to 1 and the misclassification number decreasing to 0, reflecting a growing spectral divergence between the CK and T groups as the disease progressed. By Day 6 and Day 7, both PLS-DA and CNN models achieve perfect scores (1.000) across all metrics and 0 misclassifications for both spectral parameters, indicating highly reliable classification capabilities. The RF model, while slightly less performant compared to PLS-DA and CNN, still demonstrates high efficacy, especially on the later days, with accuracy, sensitivity, and specificity nearing or reaching 1.000. These findings underscore the robustness of these models in distinguishing between healthy and infected lettuce over time.

Notably, across all observation days, the classification models based on Ref consistently outperformed those based on VI. This suggests that full-band spectral reflectance provides richer information for capturing the subtle physiological and biochemical changes associated with downy mildew infection compared to the vegetation indices derived from spectral data.

These results demonstrate the feasibility of using hyperspectral data for the early, non-invasive detection of downy mildew in lettuce. They also emphasize the advantage of leveraging full spectral reflectance over simplified vegetation indices for achieving higher model accuracy.

3.4.2. Key Bands for Early Detection

An analysis of the Variable Importance in Projection (VIP) values derived from the PLS-DA models using the Ref for each day revealed consistent patterns throughout the experimental period. As shown in Figure 9, the VIP values of specific wavelength ranges, including 508–510 nm (peaking at 509 nm), 622–655 nm, 685–695 nm (peaking at 640 nm), and 737–758 nm (peaking at 744 nm), were consistently above the threshold of 1 across all observation days. This indicates that the spectral reflectance of lettuce in these bands exhibited significant differences between the T group and the CK group. Importantly, these differences emerged as early as 24 h post-inoculation and persisted stably throughout the disease progression, suggesting that these wavelength ranges are key spectral bands for distinguishing between diseased and healthy lettuce plants at an early stage.

Further analysis revealed that, as shown in Figure 10, except for some bands in the 620–632 nm and 737–758 nm range, most of the identified key bands for early detection also demonstrated strong correlations with DI, as discussed in Section 3.2.1. Specifically, the 508–510 nm range corresponds to the broader 508–600 nm range, which was found to be positively correlated with DI (r > 0.5); The 622–655 nm and 685–695 nm ranges align with the 634–696 nm range, which was significantly negatively correlated with DI (r < −0.5). These results confirm that the identified key bands are closely associated with downy mildew infection. Furthermore, they highlight that most of these bands are sensitive to both the presence of infection and the severity of the disease. The 620–632 nm and 737–758 nm range, in contrast, appears to be more specific to infection status rather than disease severity.

3.4.3. Key Vegetation Indices for Early Detection

An analysis of the VIP values derived from the PLS-DA models using VI for each day identified key indices for early detection of downy mildew. Figure 11 presents the VIP values for VIs with values greater than 1 on at least three separate days. Among these, four indices—EVI, mSR705, NPQI, and PRI—consistently had VIP values above 1 throughout the experimental period. This finding indicates that significant differences in these VIs existed between the T group and the CK group from the first day after inoculation (Day 1) to the seventh day (Day 7), when visible symptoms appeared. These four VIs can thus be considered as key indices for distinguishing infected and healthy lettuce at an early stage of infection. Other VIs showed variable performance across different days, indicating less consistent sensitivity to the infection.

Further insights were obtained by integrating the findings from Section 3.2.2: EVI and PRI demonstrated significant positive correlations with DI, with r values of 0.371 and 0.679, respectively. This suggests that EVI and PSRI are sensitive not only to the presence of infection but also to the severity of the disease. mSR705 and NPQI, on the other hand, showed no significant correlation with DI, implying that these indices are more specific to infection status rather than disease severity.

These results highlight the utility of EVI, mSR705, NPQI, and PRI as critical VIs for early detection of downy mildew in lettuce. While EVI and PSRI are valuable for both infection detection and assessing disease severity, mSR705 and NPQI provide reliable markers for identifying the presence of infection.

4. Discussion

4.1. DI, Flav, and Anth

In this study, a significant negative correlation was observed between DI and the contents of flavonoids and anthocyanins, indicating that plants with higher levels of these secondary metabolites exhibited milder symptoms of downy mildew. This finding supports the role of flavonoids and anthocyanins in plant defense mechanisms, as highlighted in previous studies [28,29,30]. These compounds may enhance resistance to downy mildew through their antioxidant activities and by modulating plant responses to both biotic and abiotic stresses [53]. Additionally, the strong positive correlation between Flav and Anth underscores the synergistic role of these secondary metabolites in plant defense responses. Their interconnected pathways may contribute to a more robust defense against pathogen invasion, further emphasizing their importance in plant immunity. In contrast, no significant correlation was found between DI and chlorophyll content. While previous research has reported that chlorophyll levels can be affected by stress factors such as disease severity, chlorophyll is not directly involved in the stress defense process [54,55]. This may explain the lack of a direct relationship between chlorophyll levels and resistance to downy mildew among different lettuce germplasm accessions observed in this study.

These findings have important practical implications for breeding programs. By selecting plants with higher flavonoid and anthocyanin contents, breeders can develop new lettuce cultivars with enhanced resistance to downy mildew. Moreover, these biochemical parameters can serve as early indicators for resistance screening, thereby accelerating the breeding process and improving efficiency.

4.2. DI, Flav, Anth, and Spectral Parameters

This study revealed significant correlations between spectral parameters (Ref and VI) and the DI, highlighting their potential for monitoring downy mildew infection in lettuce. Spectral reflectance within the 410–503 nm and 630–690 nm wavelength ranges, as well as vegetation indices such as GNDVI, GRVI, ARI1, and ARI2, showed significant negative correlations with DI and positive correlations with Flav and Anth contents. Conversely, reflectance within the 510–610 nm range and PRI exhibited significant positive correlations with DI and negative correlations with Flav and Anth. These findings align with previous studies identifying spectral regions and vegetation indices associated with downy mildew severity in other crops, such as grapevine [56], brassica [27], and watermelon [57]. Furthermore, the results corroborate earlier findings that demonstrated the feasibility of using VIS/NIR spectroscopy for non-destructive estimation of Flav and Anth contents in plant leaves [58]. These outcomes underscore the ability of spectral data to capture subtle biochemical variations, making it a valuable tool for assessing both disease severity and associated biochemical characteristics.

In the early detection of lettuce downy mildew, specific wavelengths (508–510 nm, 622–655 nm, 685–695 nm, and 737–758 nm) and vegetation indices (EVI, mSR705, NPQI, and PRI) played critical roles in distinguishing infected plants from healthy ones. These spectral parameters demonstrated significant differences between treatment (T) and control (CK) groups within 24 h post-inoculation. Among them, reflectance at 508–510 nm, 622–655 nm, and 685–695 nm, as well as PRI, were strongly correlated with DI, suggesting their utility in both detecting infection and gauging disease severity. Similar trends have been observed in other plant–pathogen systems [17,18,26], reinforcing the generalizability of these spectral features for early disease detection.

Notably, PRI emerged as a sensitive indicator across various studies involving plant responses to biotic and abiotic stresses. PRI is designed to estimate photosynthetic efficiency [59], which is often reduced under stress conditions, leading to lower PRI values as stress severity increases. This trend was observed in Experiment 2 of this study, where the PRI of healthy lettuce (CK) consistently exceeded that of infected plants (T). However, in Experiment 1, a unique relationship was observed: lettuce varieties with higher baseline PRI values exhibited greater disease severity (higher DI). This counterintuitive result may be attributed to varietal differences in Flav and Anth contents. Varieties with higher Flav and Anth levels showed stronger resistance to downy mildew (lower DI) but typically had lower PRI values due to reduced absorption of red and blue light associated with anthocyanin-rich leaves. Conversely, varieties with lower Flav and Anth levels exhibited weaker resistance (higher DI) and higher PRI values, as their greener leaves absorbed more light for photosynthesis [60]. In Experiment 2, both lettuce varieties were green and exhibited no baseline differences in photosynthetic efficiency, allowing clear differentiation between T and CK groups based solely on the impact of downy mildew infection.

These findings have significant implications for developing spectral tools for disease monitoring. By identifying spectral regions and vegetation indices closely associated with disease severity and biochemical markers, robust models can be constructed for disease assessment. Such models facilitate the early detection of diseases and accelerate the identification of resistant genotypes in breeding programs. Future research should focus on elucidating the physiological and biochemical mechanisms underlying spectral parameter changes and their connection to plant defense responses. Additionally, exploring the performance of these spectral features across diverse environmental conditions and plant varieties will enhance the universality and accuracy of spectral models.

4.3. Determination and Early Detection Models for Downy Mildew

This study developed models based on Ref and VI for predicting the DI, Flav, and Anth and detecting downy mildew in the early stage. These models hold significant promise for achieving non-destructive and high-throughput assessments of plant diseases and biochemical characteristics. The results revealed that models using full-spectrum Ref as independent variables outperformed those based on VI. This superiority is attributed to the comprehensive wavelength information captured by Ref, spanning from visible to near-infrared regions, which likely contains more intricate details related to plant physiological and biochemical states. In contrast, VI, despite its simplicity in calculation, may lose subtle spectral information relevant to predicting DI and biochemical parameters.

The early detection models for diseases can identify infections during the initial stages of development, enabling timely control measures in agricultural production to mitigate crop losses. In breeding programs, these models facilitate the early identification of resistant varieties, thereby accelerating the breeding process. In this study, classification models were employed to differentiate between lettuce infected with downy mildew and healthy lettuce. The construction and evaluation of these models are critical to understanding the potential of hyperspectral data for early disease detection.

The findings showed that classification models exhibited robust predictive performance as early as 1 day post-inoculation (24 h), with accuracy values exceeding 0.65. This indicates that hyperspectral data could detect early physiological changes in lettuce following inoculation. As the disease progressed, the predictive ability of these models further improved, reaching their peak on Day 7. These results underscore the efficacy of hyperspectral data in distinguishing between infected and healthy lettuce, even at early infection stages. The successful development and validation of these models highlight the potential of spectral data for predicting plant diseases and biochemical parameters, offering rapid and accurate tools for disease management and resistance breeding. However, the limitations of these models lie in their dependence on specific experimental conditions and lettuce varieties. Future research should focus on validating the generalizability and robustness of these models across diverse environmental conditions, lettuce varieties, and growth stages.

4.4. PLS, RF, and CNN Algorithms

The comparative analysis of our models highlights distinct strengths and potential pitfalls across the employed regression and classification algorithms including PLS, RF, and CNN. For instance, the PLS model offers the advantages of simplicity and straightforward interpretability, making it useful for initial screening and rapid assessments. However, its reliance on linear assumptions can limit its effectiveness in capturing complex, nonlinear spectral variations, potentially reducing accuracy when subtle physiological changes are involved. In contrast, the RF algorithm demonstrates robust performance through its ensemble approach, which effectively mitigates overfitting and handles noisy data well. Nonetheless, RF may require careful parameter tuning, and its ensemble nature can obscure the detailed interpretability of individual spectral contributions. Meanwhile, CNN excels at modeling intricate spectral patterns and nonlinear relationships inherent in the hyperspectral data, leading to superior predictive accuracy in both regression and classification tasks. Yet, CNNs are computationally intensive and demand substantial training data to achieve optimal generalization, which could be a limiting factor in resource-constrained scenarios.

The development prospects of the proposed methods, which integrate regression models based on PLS, RF, and CNN algorithms, are promising. The novelty of our approach lies in its comprehensive integration of hyperspectral imaging with multiple machine learning algorithms for the early, non-destructive detection and quantification of lettuce downy mildew—a challenge that has not been addressed in previous studies. This methodology not only establishes a spectral basis for disease resistance by correlating spectral characteristics with biochemical parameters but also paves the way for automated, high-throughput screening in breeding programs. Future work should focus on further optimizing these models, incorporating additional spectral indices, and exploring advanced deep learning techniques to enhance prediction robustness. Such advancements could extend the applicability of this approach to a wider range of crops and diseases, thereby contributing significantly to precision agriculture and sustainable crop management.

5. Conclusions

This study integrates hyperspectral imaging technology with biochemical parameter analysis to advance the early detection and resistance evaluation of lettuce downy mildew. The results demonstrate that hyperspectral imaging effectively captures spectral features associated with physiological and biochemical changes in plants, which are closely related to disease progression and resistance responses. Correlation analysis revealed a significant negative relationship between the DI and levels of Flav and Anth, highlighting the role of these secondary metabolites in enhancing plant resistance. Specific spectral regions and vegetation indices highly correlated with DI, Flav, and Anth were identified, providing potential spectral indicators for disease assessment and early detection. Furthermore, regression models developed using PLS, RF, and CNN algorithms demonstrated high accuracy and reliability in predicting DI, Flav, and Anth, while the classification models successfully detected early physiological changes in lettuce within 24 h of infection, offering an effective tool for the early detection of downy mildew.

Despite these advances, several challenges remain, including environmental variability, high technical costs, complex data analysis, and real-time system integration. Future research should validate model performance under diverse conditions while developing cost-effective sensor systems and streamlined analytical workflows. Moreover, incorporating advanced techniques such as transfer learning and data fusion could further enhance model robustness and adaptability. These efforts will ultimately facilitate the broader integration of hyperspectral imaging into precision agriculture for early disease detection and resistance screening, contributing to more sustainable and resilient crop production systems.

Overall, this study provides a scientific foundation and practical tools for managing lettuce downy mildew and resistance breeding, while paving the way for broader applications of hyperspectral imaging in plant pathology to enhance the efficiency and sustainability of modern agricultural production.