Application of NDVI for Early Detection of Yellow Rust (Puccinia striiformis)

Abstract

Yellow rust is one of the most destructive fungal diseases affecting wheat, significantly reducing yield and grain quality. Early detection is crucial for effective plant protection and disease management. This study aims to develop and validate a methodology for early diagnosis of yellow rust using the Normalized Difference Vegetation Index (NDVI) derived from UAV-acquired spectral data. This research was conducted in an experimental wheat field near General Toshevo, Bulgaria, which is owned by the Dobrudja Agricultural Institute (DAI). A widely cultivated winter wheat variety, Enola, was monitored using UAV-based imaging, and the NDVI values were analyzed to assess the correlation between spectral reflectance and infection severity. The NDVI showed a moderate correlation as an indicator of pathogen-induced stress, with moderate predictive capability (R² = 51.4%) for assessing yellow rust infection severity. The results demonstrated that UAV-based NDVI analysis could effectively detect early-stage infections and monitor the spatial spread of the disease. The proposed methodology enables large-scale, non-invasive monitoring of wheat health, facilitating early disease detection. This approach can help optimize disease management strategies, although ground-based validation remains essential to distinguish between different stress factors affecting vegetation.

1. Introduction

Wheat is the crop of greatest importance for feeding the population in many countries around the world. Maintaining crops in good agrotechnical condition is a prerequisite for obtaining high yields and quality grain. Conditions of constant climatic changes [1,2], such as extreme droughts, unusual cold spells, and heavy rainfall, lead to the rapid development of many diseases, pests, and weeds in wheat. Their timely detection and adequate control are crucial for maintaining sustainable yields in wheat cultivation in the conditions of modern intensive agriculture. Some of the most frequently occurring diseases in wheat are stripe rust, or yellow rust, septoria blotch, fusarium head blight, and others [3,4,5]. Yellow rust (Puccinia striiformis) is a destructive fungal disease that has the greatest consequences on the development of wheat and the reduction in yield and quality of grain [6,7,8]. Traditional invasive detection methods with sampling and laboratory testing are mostly manual and labor-intensive. Strategies such as cultivating resistant varieties, applying fungicides, and practicing good agricultural techniques often fail to effectively identify and respond to wheat rust outbreaks [9].

With the integration of information technology in agriculture, artificial intelligence (AI), and the development of precision agriculture, many well-known traditional technologies in growing crops are undergoing significant changes. One of the main advantages of these precision technologies is the ability to remotely monitor, detect stress and disease severity, quantify, and precisely estimate the impact on the affected plants [10,11,12]. Recent studies have demonstrated that artificial intelligence combined with UAV-acquired imagery significantly supports agronomic decision-making and disease detection, confirming the relevance and efficiency of such approaches [13,14,15].

One of the rapidly developing approaches for remote sensing is the aerial surveillance approach using unmanned aerial vehicles (UAVs) [16,17]. Its advantages include the possibility of achieving large-scale surveillance compared to ground-based surveys. The application of UAVs with multispectral cameras is a widely used practice [18]. Identification and quantification of symptoms of wheat leaf rust and stripe rust using RGB cameras at the UAV scale and ground-based visual observations was proposed in [19], where the RGB-imagery-derived values and observed disease severities at the study sites had correlation coefficients greater than 0.90. Recent applications in vegetation monitoring and estimation of biophysical parameters through machine learning and vegetation indices further highlight the potential of UAVs under different agricultural systems [20].

An interesting method for identifying wheat rust was proposed by [21], which uses data obtained from UAVs to train a semantic segmentation model of the pyramid scene parsing network (PSPNet) to classify healthy wheat, wheat with rust, and bare soil, with the recognition accuracy of the PSPNet model in this study reaching 98%. Identification of yellow rust using a wheat-leaf-rust-monitoring model based on the backpropagation neural network (BPNN) method was proposed by [22]. In [23], a combination of CNNs and RGB-based ultrahigh spatial resolution images from UAVs was used, providing a simple and rapid method for accurate large-scale detection of crop diseases. The reviewed studies primarily focus on developing customized neural networks based on two-dimensional convolutional layers, largely overlooking the potential of their three-dimensional counterpart, which could effectively capture and model interrelated features.

For example, Nguyen et al. [24] developed a machine learning pipeline incorporating a support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP) to classify field plots as healthy, mildly infected, or severely infected. Furthermore, a custom three-dimensional convolutional neural network (3D-CNN) leveraging a feature-learning mechanism was proposed as an alternative predictive approach.

Other approaches for detecting and identifying yellow rust infection using hyperspectral cameras have been explored by [25,26,27]. A promising method with the use of UAV and a multispectral camera to monitor wheat yellow rust disease was proposed by [17], which is focused on the spatiotemporal monitoring of winter wheat inoculated with various levels of yellow rust inoculum during the entire growth season. Zheng et al. [28] conducted a comparative analysis of the Photochemical Reflectance Index (PRI) and the Anthocyanin Reflectance Index (ARI) at different growth stages to evaluate their potential for estimating the severity of yellow rust disease. Hyperspectral imaging provides higher spectral resolution compared to multispectral cameras; however, its limitations include significantly higher costs, the necessity for trained users capable of handling the more complex processing of acquired images, increased data storage and computational requirements, as well as longer acquisition and processing times, which may limit its practical application in real-time field monitoring.

The previous studies established a strong methodological foundation for UAV-based yellow rust detection in wheat, leveraging spectral, spatial, and temporal information through advanced machine learning frameworks. However, much of this research was conducted in controlled experimental environments, often limited to individual plot trials or small areas. This creates a significant gap in practical applicability, as findings may not directly translate to large-scale commercial wheat cultivation.

One of the major challenges in UAV-based disease detection is the trade-off between image resolution and operational feasibility. Many studies achieve high-resolution imagery by maintaining a UAV flight altitude of 30 m; however, this approach is impractical for large-scale wheat fields due to the limited flight endurance and rapid battery depletion of UAVs. While hyperspectral cameras can address this issue by providing detailed spectral information, their high cost and the complexity of data processing present significant barriers to widespread adoption in agricultural practice.

A critical challenge in our research is the detection of yellow rust (Puccinia striiformis) in large-scale commercial wheat fields when using UAV imagery captured from high altitudes. This is particularly relevant in extensive wheat-growing regions such as Southern Dobrudja, Bulgaria, where practical, cost-effective disease-monitoring solutions are urgently needed. Addressing this gap in scientific research is the primary objective of our study.

The main objective of this research is to develop a methodology for the early detection of yellow rust (Puccinia striiformis) using the NDVI derived from UAV-acquired imagery. The specific goals include (1) assessing the correlation between the NDVI values and yellow rust infection severity, and (2) proposing an accessible and scalable approach that enables practical application by farmers using affordable UAVs and sensors. This integrated methodology aims to support timely disease monitoring and improve overall crop protection strategies.

To address the limitations identified in previous research, particularly the lack of scalable, cost-effective, and operationally feasible methods for detecting yellow rust in large-scale wheat fields, our proposed approach emphasizes practicality and accessibility. By relying on NDVI data obtained from commercially available cameras mounted on UAVs flying at higher altitudes, we aim to overcome the challenges related to limited flight time, image coverage, and high equipment costs. This approach is specifically tailored to real-world agricultural conditions in regions such as Southern Dobrudja, Bulgaria, where large wheat-growing areas require efficient and rapid disease monitoring. The methodology balances simplicity, accuracy, and scalability by focusing on the relationship between the NDVI and yellow rust severity, contributing to more realistic applications beyond controlled experimental environments.

2. Materials and Methods

2.1. Study Area and Experimental Field Characteristics

The investigated field is in the Republic of Bulgaria, Dobrich region, General Toshevo municipality, in the land of the village of Petleshkovo on a laboratory experimental field sown with common winter wheat (Triticum aestivum) with coordinates 43.658628N, 28.025217E. The study area is shown in Figure 1, with the experimental area enclosed in a red rectangle.

It is located in the geographical region of Dobrudja. In terms of climate, the experimental area belongs to the eastern part of the Temperate Continental climate region of the Republic of Bulgaria. Its specificity is determined by the plateau-like nature of the relief, which is a low plateau with an altitude of about 200 m, and the constantly blowing winds due to the proximity of the Black Sea. The regional climate is described based on long-term climatic norms. The coldest month is January, with an average temperature of −0.5 °C. While winters are relatively harsh, extremely low temperatures are uncommon. For illustration, the absolute minimum air temperature of −22.7 °C was recorded in 1954, marking one of the coldest years.

Summer is cooler than other regions of the same climatic zone. The warmest month is July, with a temperature norm of +20.9 °C. The highest recorded temperature in the region, +38.5 °C, occurred in 1985 and is cited here as an example of a historic extreme. The annual precipitation rate at the institute is 527.5 mm, with a maximum in June (67.3 mm). Most precipitation (an average of 56.4%) falls during the vegetation months. The annual minimum is in January (30.5 mm). The relative humidity of the air varies from 69% to 88%. Its average value of 78% is significantly higher than the average for the country [29].

The soils in the experimental field represent the conditions of weakly leached chernozems and are comprehensively considered to be among the most fertile soils in Bulgaria, with the potential for obtaining high yields. They are characterized by a relatively thick humus horizon (60–80 cm), and by the content of humus in the arable layer, they are classified as medium humus. The characteristic qualities of the soils are as follows: neutral reaction, low supply of available nitrogen and phosphorus, and medium-to-good supply of exchangeable potassium. Their mechanical composition determines a favorable combination of water–physical properties with a very good structure of the subsoil horizons [30].

This study focuses on “Enola”, a widely cultivated winter common wheat variety. It is an awned variety of the II/b/strength group of wheat. The genotype exhibits a high adaptive potential and has the following characteristics [31]:

• High stable yields, complex disease resistance;
• Stem height: 80–90 cm;
• High lodging resistance;
• High productive tillering;
• Class: spiny, highly resistant to lodging;
• Mass of 1000 seeds: 41–44 g;
• Hectoliter weight: 81–83 kg;
• Early variety with a long grain filling period;
• Very good cold and winter hardiness;
• High drought tolerance;
• Good resistance to powdery mildew, brown, yellow, and black rust, and septoria; high resistance to fusarium;
• Productivity: 8.5 t/ha;
• Quality: Group B—medium strength wheat;
• Widely ecologically plastic variety for Bulgaria;
• Suitable for intensive cultivation.

According to data from the Ministry of Agriculture and Forestry of Bulgaria, Enola quickly gained popularity following its introduction into commercial production, accounting for up to 60% of the wheat cultivated in the country.

2.2. UAV Data Collection

The experimental data were acquired using an unmanned aerial vehicle (UAV) from DJI, specifically the Mavic 2 Pro model [32]. It has a Hasselblad L1D-20 camera, featuring a 20.1 MP 1” CMOS sensor that captures images in the RGB range. Since the NDVI is derived from the red and near-infrared light spectra, the built-in camera of the UAV cannot be used to capture these wavelengths. To address this, a second external Mapir camera [33] was mounted on the UAV, capturing images in the RGN (red–green–near-infrared) spectra. The Mapir camera’s specifications include R (red) at 660 nm ± 10 nm, G (green) at 550 nm ± 10 nm, and N (near-infrared) at 850 nm ± 20 nm, with a Sony Exmor R IMX117 sensor (12 MP, 4000 × 3000 px), featuring a Bayer matrix where the blue channel is repositioned in the near-infrared spectrum. The UAV is shown in Figure 2, which also illustrates the Mapir calibrator used to correct for variable sunlight intensity. The multispectral survey was conducted at a height of 100 m above the ground with 80% overlap in horizontal and vertical directions. The flight speed was limited to 10 m/s. The camera settings included a focal length of 3.0 mm, a shutter speed of 1/500, an aperture of f/2.8, and an ISO of 100, all set with fixed parameters.

Imaging was conducted on 5 June 2021 between 12:00 and 13:00 h to minimize the effect of shadows. In total, 300 photographs were captured in a single session, from which an orthomosaic was generated.

The Pix4Dcapture software [34] was used to plan the missions. This application is designed to plan and execute photogrammetry missions for 2D mapping and 3D modeling. It supports a variety of aerial mapping mission types, including polygon, grid, double-grid, circular, and free-flight missions, and is compatible with a wide range of aircraft.

The experimental field area is 68 × 360 m. An orthomosaic of the field in RGB is shown in Figure 3a. In Figure 3b, an orthomosaic in the RGN spectrum is shown, and in Figure 3c, an NDVI map.

The NDVI is a ratio that takes into account the amount of infrared rays reflected by healthy plants [35] and is calculated using the formula:

$$\mathrm{NDVI}={\displaystyle \frac{N-R}{N+R}}$$

(1)

where: $N$ is the near-infrared reflectance, R is the red reflectance.

It characterizes parameters such as vegetation density, growth, and the presence of weeds or diseases, and it is used to predict yields. These indices are generated by capturing images of green vegetation, which absorb electromagnetic waves in the visible red range and reflect them in the near-infrared range due to the reflective properties of chlorophyll.

The field is divided as shown in Figure 4, with Figure 4a presenting its NDVI map. The field is divided into three zones (A, B, and C) along its width and into 16 zones (marked 1–16) along its length. This division results in plots of size 22.5 × 22.6 m. The camera used for imaging has a resolution of 12 MP and a Sony Exmor sensor with a 2:3 aspect ratio. The images were captured from a height of 100 m above the terrain. The resulting NDVI map has a resolution of 150 × 800 pixels, with a ground resolution of 4 cm per pixel. Applying the corresponding division from Figure 4a, squares of 50 × 50 pixels are obtained. Figure 4b shows an RGN orthomosaic of the studied field, divided into three zones (D, E, and F) along its width and 16 zones (marked 1–16) along its height, forming zones of 50 × 50 pixels each.

This study examines the relationships between the NDVI and the red and near-infrared spectra, comparing these with an estimated value for the degree of infection. The infected zone, diagnosed with yellow rust, is assigned a value of 1, while zones without infection at this stage are assigned a value of 0. Based on the obtained NDVI map, the maximum numerical value of the vegetation index is reduced to 0, and the minimum value is set to 1. The subsequent development of rust spread is monitored using the estimated values and by analyzing the generalized relationships between the color data and the vegetation index. The results are analyzed using IBM SPSS software v. 26 [36].

A quantitative assessment of the values in each square of the NDVI map is conducted using ImageJ software v. 1.54p [37]. The number of pixels corresponding to each of the three-color categories is analyzed in each square. These values are used to calculate a contagion coefficient, denoted as K, using the following formula:

$$K=({\displaystyle \frac{Red}{Green}})/8$$

(2)

where: Red is the number of pixels with red color, and Green is the number of pixels with green color.

2.3. Statistical Analysis

The results were analyzed using multiple regression analysis to examine the dependence of K on the parameters R, N, and NDVI. The data were processed using IBM SPSS Statistics v. 26 software [38]. The equation for the Multiple Linear Regression Model [39] is as follows:

Y_i = β₁X_i1 + β₂X_i2 + … + β_KX_iK + εi,

(3)

where typically X_i1 = 1 is the intercept.

3. Results

The area of low NDVI in and around zone B5 was further investigated through field sampling to determine the cause of the degraded vegetation. The sampling results confirmed the presence of yellow rust (Puccinia striiformis), as shown in Figure 5.

The results for column A are presented in Figure 6, where Figure 6a shows the number of pixels from each square in column D of Figure 4. Figure 6b displays the generated average NDVI value and the contamination coefficient. The NDVI is derived from the average values of the colors in the divided areas shown in Figure 4b. It can be observed that a decrease in the NDVI, combined with an increase in the amount of red, is indicative of the presence of yellow rust.

The correlation analysis for column A reveals a Multiple R value of 0.93 and an R² value of 0.86, suggesting a strong relationship between the contamination coefficient and reflectance in the red and near-infrared spectrum. A more specific analysis of areas with the best vegetation, namely, A12, A13, and A14, demonstrates an even stronger correlation, with a Multiple R value of 0.96 and an R² value of 0.92.

Figure 7a presents the number of pixels from each square in column E of Figure 4. Figure 7b displays the generated average NDVI value and the infection coefficient for column B. Notably, plot B5 exhibits a significantly high value, corresponding to an early diagnosis of yellow rust, with an infection rate nearing 100%.

The correlation analysis for column B reveals a Multiple R value of 0.76 and an R² value of 0.58, indicating a moderate relationship between the contagion coefficient and reflectance in the red and near-infrared spectrum. A more specific analysis of plots B4, B5, and B6 demonstrates a stronger correlation, with a Multiple R value of 0.92 and an R² value of 0.84.

An increase in the K coefficient values is observed in plots C3 to C7; however, the values do not exceed 0.3, indicating that this area is at an increased risk of pathogen spread.

The analysis shows a Multiple R value of 0.87 and an R² value of 0.76 for column C, demonstrating a strong correlation between the estimated contamination coefficient and reflectance in the red and near-infrared spectrum.

The analysis of the relationships between cells with confirmed yellow rust development reveals a Multiple R value of 0.999 and an R² value of 0.999, indicating an exceptionally strong correlation. In contrast, for cells without good vegetative development, the Multiple R value is 0.979, with an R² value of 0.958, still demonstrating a strong but slightly lower correlation (Figure 8).

Figure 9 presents the atmospheric conditions in early June, which contributed to the onset of yellow rust development. The prevailing climatic conditions—cool temperatures during the first ten days of June (11–20 °C)—combined with regular precipitation and high atmospheric humidity, created a favorable environment for the disease to spread.

Figure 10a illustrates the infection coefficient for June 14, highlighting a notably high value in plot B4, the area where the pathogen was detected. Figure 10b presents the corresponding NDVI map, which reveals a deterioration in the overall vegetation index. The absence of green areas on the map further confirms the impact of the infection.

The boundary line for the K coefficient is plotted in Figure 10a, indicating the threshold above which yellow rust presence is confirmed. The critical boundary value is determined to be K = 0.12. The figure also shows that on the day of observation, certain areas remain unaffected by the pathogen’s spread. These include plots A11 to A14, and several plots in columns B and C, which are situated near the boundary line and exhibit values around 0.13. While these areas demonstrate early signs of pathogen development, the disease is not yet detectable in the visible spectrum.

Previously, our NDVI studies on wheat suggested that deviations in the index can be registered up to 10 days before visible spectral changes occur, reinforcing the potential of the NDVI for early disease detection.

Figure 11 illustrates the values of the infection coefficient on June 30, 25 days after the initial diagnosis of early-stage infection. Figure 11a shows the variation in the coefficient across different plots for each column of the study plan. The graph indicates that all plots exhibit an infection degree within the range of 0.2–0.4, exceeding the threshold value for pathogen presence. Field observations confirmed the widespread spread of yellow rust, followed by the subsequent wilting of the affected plants. The decline in NDVI values corresponds to the wheat entering the maturity stage and the cessation of vegetation.

Figure 11b presents the NDVI map of the study field for the same date, revealing a complete absence of green areas and a significant increase in red areas, indicating severe vegetation stress.

The summarized results of the descriptive statistics and multiple regression analysis are presented in

Table 1. Descriptive statistics.

	Mean	Std. Deviation	N	CV (%)
K	0.1749	0.13935	48	79.67
R	65.5325	30.02424	48	45.82
N	80.1404	33.16865	48	41.39
NDVI	0.1124	0.04125	48	36.70

. The table includes the mean, standard deviation, and coefficient of variation (CV, %) for each parameter. The CV values provide insight into the degree of variability, classified as low (<20%), moderate (20–40%), or high (>40%) variability.

The coefficient of variation indicated high variability for parameter K (CV = 79.67%) and moderate variability for the NDVI (36.70%), N (41.39%), and R (45.83%). These levels of variation highlight the heterogeneity of the measured parameters under field conditions.

The degree of linear relationship (correlation) between the variables is presented in

Table 2. A summary of the correlations.

		K	R	N	NDVI
Pearson Correlation	K	1.000	0.332	0.234	−0.597
	R	0.332	1.000	0.989	−0.746
	N	0.234	0.989	1.000	−0.646
	NDVI	−0.597	−0.746	−0.646	1.000
Sig. (1-tailed)	K		0.011	0.054	0.000
	R	0.011		0.000	0.000
	N	0.054	0.000		0.000
	NDVI	0.000	0.000	0.000
N	K	48	48	48	48
	R	48	48	48	48
	N	48	48	48	48
	NDVI	48	48	48	48

. It can be observed that K has a strong dependence on N.

Table 3. Statistical measures of the regression model with dependent variable K and predictors NDVI, N, and R.

Model	R	R Square	Adjusted R Square	Std. Error of the Estimate	Change Statistics					Durbin-Watson
					R Square Change	F Change	df1	df2	Sig. F Change
1	0.717	0.514	0.481	0.10041	0.514	15.507	3	44	0.000	1.682

summarizes the regression model, presenting the values of R, R², adjusted R², and the standard error of the estimate, which collectively assess the model’s goodness of fit. The multiple correlation coefficient (R) is 0.717, indicating a strong correlation between the dependent and independent variables. The coefficient of determination (R²) shows that 51.4% of the variance in K is explained by the independent variables R, N, and NDVI. Additionally, the adjusted R² accounts for 48.1% of the variance, providing a more refined measure of the model’s explanatory power relative to the variance in the predictor variables.

Table 4. Results from the ANOVA model with dependent variable K and predictors NDVI, N, and R.

Model		Sum of Squares	df	Mean Square	F	Sig.
1	Regression	0.469	3	0.156	15.507	0.000
	Residual	0.444	44	0.010
	Total	0.913	47

shows that the independent variables statistically significantly predicted the dependent variable, F (3, 44) = 15.507, p < 0.0005 (i.e., the regression model was a good fit to the data).

The results of the overall estimated coefficients of the K prediction model are shown in

Table 5. Coefficients of the K prediction model.

Model		Unstandardized Coefficients		Standardized Coefficients Beta	t	Sig.	95.0% Confidence Interval for B		Correlations			Collinearity Statistics
		B	Std. Error				Lower Bound	Upper Bound	Zero-order	Partial	Part	Tolerance	VIF
1	(Constant)	0.004	0.190		0.019	0.985	−0.380	0.387
	R	0.031	0.009	6.703	3.280	0.002	0.012	0.050	0.33	0.443	0.35	0.003	378.05
	N	−0.026	0.007	−6.084	−3.415	0.001	−0.041	−0.01	0.23	−0.458	−0.4	0.003	287.28
	NDVI	1.611	1.360	0.477	1.185	0.242	−1.129	4.352	−0.6	0.18	0.13	0.068	14.668

.

The degree of dependence between the variables R, N, and NDVI is shown in

Table 6. Coefficient correlations with dependent variable K.

Model			NDVI	N	R
1	Correlations	NDVI	1.000	−0.920	0.940
		N	−0.920	1.000	−0.997
		R	0.940	−0.997	1.000
	Covariances	NDVI	1.849	−0.009	0.012
		N	−0.009	5.601 × 10−5	−7.077 × 10−5
		R	0.012	−7.077 × 10−5	8.996 × 10−5

.

A probability density plot illustrating the effects of the parameters on the dependent variable K is presented in Figure 12. The distribution of K across the entire observed range of R is graphically depicted in Figure 12a, revealing a clear trend in how variations in R influence the dependent variable. Figure 12b demonstrates the relationship between K and N, indicating a statistically significant correlation that suggests the strong predictive capacity of N in determining K. Meanwhile, Figure 12c visualizes the probability density between K and NDVI, obtained through ANOVA, highlighting a strong concentration of probability values within a specific range. This suggests that changes in the NDVI have a pronounced effect on K, reinforcing its reliability as an indicator of system variability. The clustering patterns (for spatial data discovery based on the contagion coefficient) observed in these plots indicate a well-defined structure in the data, further supporting the hypothesis that variations in spectral indices and environmental parameters can be effectively captured through NDVI-based assessments. These findings validate the use of the NDVI as a tool for monitoring and predicting variations in KK, with potential applications in early disease detection.

4. Discussion

By combining field data and the methodology for assessing the presence of yellow rust, our results show that the proposed method for diagnosing yellow rust in a crop of common winter wheat gives good results. Similar to our study, some researchers use different approaches for diagnosing yellow rust in wheat. For example, in [40], the machine learning model Ir-UNet, designed to improve the disease identification of yellow rust segmentation accuracy, was combined with multispectral features from a RedEdge-equipped DJI M100 UAV. On the other hand, [41] provides hyperspectral image analysis for identifying yellow rust in wheat, and explores spectral reflectance, vegetation indices, and texture features with machine learning (SVM) for yellow rust identification. In this research, hyperspectral data were obtained from the Headwall VNIR imaging sensor system, not UAV-based acquisition, which makes it difficult to apply to large-scale agricultural areas.

In this study, the resulting map of the NDVI was divided into zones of three columns with 16 rows of 50 × 50 pixels, which correspond to a terrain size of 22.5 × 22.6 m. The columns are assigned the letters A, B, and C, and the rows are numbered from 1 to 16. In the plots thus formed, an assessment of the quantitative ratio between the red and green colors was made. Similar to our study but without the NDVI/vegetation indices, ref. [42] proposed a methodology for detecting and quantifying yellow rust based on hyperspectral images from a UAV. The same researchers assessed the severity and progression of the disease, using optimized sets of features and prediction algorithms with cross-information transfer (on the ground and UAV). Other authors [5] proposed machine learning models for predicting wheat yield under yellow rust stress. Thermal and RGB indices with biophysical parameters were integrated, showing the NDVI declines as rust severity increases, leading to yield loss. This study was limited to yield prediction and did not try to diagnose yellow rust.

In our research, we used a similar approach, and an orthomosaic consisting of 300 separate photos taken with an RGN camera was also studied. It was divided into three columns labeled as D, E, and F, and rows numbered from 1 to 16. The average value of the NDVI was calculated from the values of the red and near-infrared channels. Compared with the results of sampling from the laboratory experimental field with winter wheat, where an early diagnosis of yellow rust was made, a methodology was developed for quantitative assessment of the number of pixels in each plot with red color related to the number of infrared pixels. Such techniques [43] have been successfully implemented in previous studies using UAV-acquired hyperspectral imagery and machine learning for wheat rust diagnosis, through optimal band combinations using regression models for accurate disease index estimation across 960 wheat varieties. In [44], hyperspectral-based monitoring of wheat yellow rust using UAV data and machine learning was used to evaluate wavelet features and vegetation indices with SVM classification at different infestation stages (26 DAI and 42 DAI). Our approach refines these methods by offering an enhanced quantitative analysis that improves the reliability of early diagnosis, using more accessible equipment and achieving sufficiently good detection accuracy and quantification of the degree of yellow rust infection.

In our research, the results were compared with the NDVI, and an infection coefficient was formed. They were confirmed by observation in the following weeks until the wheat reached maturity. An assessment of the amount of infection in each plot was made, from which the dependence between the cells with confirmed development of yellow rust was obtained, showing Multiple R = 0.999 and R² = 0.999. Previous studies have indicated that UAV-based assessments can achieve high accuracy in yellow rust detection [45,46]. In plots without good vegetative development, we obtain Multiple R = 0.979 and R² = 0.958. Our study further solidifies these findings by demonstrating exceptionally high accuracy in detecting and quantifying yellow rust presence across various vegetative conditions.

To validate the results, a multiple regression analysis was conducted using the IBM SPSS Statistics v. 26 software product, in which the dependencies of the individual parameters on K were obtained. The values R = 0.717, R² = 0.514, adjusted R² = 0.481, and the independent variables statistically significantly predict the dependent variable, F (3, 44) = 15.507, p < 0.0005. These findings align with previous research emphasizing statistical validation of remote sensing methods for crop disease detection [47,48]. Our contribution enhances the statistical robustness of these methodologies by offering improved parameter estimation and validation techniques, reinforcing the practical application of remote sensing in disease monitoring.

5. Practical Applications and Limitations

The proposed methodology demonstrates significant potential for practical application in modern crop protection. Based on UAV-acquired NDVI data, the model enables rapid and cost-effective monitoring of yellow rust (Puccinia striiformis) infection over large agricultural areas. This approach allows for the timely identification of infected zones and facilitates spatial analysis of disease intensity and progression, which is difficult to achieve through traditional field sampling. It could be the base for a valuable decision-support tool for farmers and agronomists, enabling them to assess the dynamics of infection development and to optimize intervention strategies, such as targeted pesticide application or the selection of resistant cultivars. Furthermore, the approach is scalable and can be adapted for affordable drone equipment and standard multispectral cameras, making it accessible for a wide range of users.

Beyond yellow rust, the methodology can be extended to other crop diseases, provided that sufficient spectral data and field validation are available. This flexibility makes it a promising tool for integrated disease management in precision agriculture.

However, some limitations must be acknowledged. Currently, the system cannot independently identify the exact pathogen type using remote sensing data alone. Ground-truth validation by a trained agronomist remains essential for accurate diagnosis. This limitation may be addressed by developing a more comprehensive spectral database encompassing multiple pathogens, alongside studies on the specific reflectance characteristics of different diseases. As such data accumulate, advanced classification techniques, including machine learning and artificial intelligence, could enhance the model’s diagnostic capabilities by matching observed patterns to known spectral signatures.

Despite this limitation, the current model provides a powerful and practical framework for early detection and spatial monitoring of disease outbreaks, offering significant benefits for proactive crop health management.

6. Conclusions

This study developed a robust methodology for assessing infection levels and their spatial distribution using the Normalized Difference Vegetation Index (NDVI). By analyzing trends in NDVI fluctuations, we established expected boundaries for its variation in the presence of the yellow rust pathogen. Our findings confirm that the NDVI is an indicator for detecting pathogen-induced stress, which reduces vegetation index values, making it a valuable tool for estimating infection severity in specific areas. The results obtained through remote sensing were validated by direct sampling from an experimental wheat field, ensuring the reliability of the approach.

The multiple regression analysis demonstrated a significant correlation, with an R-value of 71.7% and an R² of 51.4%, indicating a moderate predictive capability of NDVI-based assessments. This suggests that the NDVI can be effectively utilized to monitor infection spread and assess the severity of plant stress caused by various pathogens. Importantly, this study provides a solid foundation for using the NDVI in large-scale disease monitoring, particularly in precision agriculture, where early and efficient detection of pathogens is critical for managing crop health.

The proposed methodology is highly applicable for large-scale plant health monitoring, offering a non-invasive and efficient means of early disease detection. It is particularly useful for precision agriculture, allowing farmers and researchers to identify affected areas promptly and optimize disease management strategies. However, while the NDVI method can effectively detect stress-induced vegetation decline, it does not distinguish between different stress factors. Thus, ground-truthing through field sampling remains essential to confirm the specific cause of vegetation deterioration.

Our future research will focus on refining the methodology by integrating additional spectral indices (e.g., EVI, SAVI) and machine learning techniques to enhance accuracy in pathogen differentiation. Expanding this study to include various crops and environmental conditions would further validate the approach’s generalizability. Furthermore, developing automated systems for real-time NDVI monitoring could significantly improve early disease detection and intervention strategies, contributing to more sustainable and efficient agricultural practices. This research opens the door for future innovations in disease management and monitoring, offering promising perspectives for more resilient agricultural systems.