Unmanned Aerial Vehicle-Based Hyperspectral Imaging for Potato Virus Y Detection: Machine Learning Insights

Abstract

The potato is the third most important crop in the world, and more than 375 million metric tonnes of potatoes are produced globally on an annual basis. Potato Virus Y (PVY) poses a significant threat to the production of seed potatoes, resulting in economic losses and risks to food security. Current detection methods for PVY typically rely on serological assays for leaves and PCR for tubers; however, these processes are labor-intensive, time-consuming, and not scalable. In this proof-of-concept study, we propose the use of unmanned aerial vehicles (UAVs) integrated with hyperspectral cameras, including a downwelling irradiance sensor, to detect the PVY in commercial growers’ fields. We used a 400–1000 nm visible and near-infrared (Vis-NIR) hyperspectral camera and trained several standard machine learning and deep learning models with optimized hyperparameters on a curated dataset. The performance of the models is promising, with the convolutional neural network (CNN) achieving a recall of 0.831, reliably identifying the PVY-infected plants. Notably, UAV-based imaging maintained performance levels comparable to ground-based methods, supporting its practical viability. The hyperspectral camera captures a wide range of spectral bands, many of which are redundant in identifying the PVY. Our analysis identified five key spectral regions that are informative in identifying the PVY. Two of them are in the visible spectrum, two are in the near-infrared spectrum, and one is in the red-edge spectrum. This research shows that early-season PVY detection is feasible using UAV hyperspectral imaging, offering the potential to minimize economic and yield losses. It also highlights the most relevant spectral regions that carry the distinctive signatures of PVY. This research demonstrates the feasibility of early-season PVY detection using UAV hyperspectral imaging and provides guidance for developing cost-effective multispectral sensors tailored to this task.

1. Introduction

Potato Virus Y (PVY) is a pathogenic virus within the Potyvirus genus and Potyviridae family [1] that is responsible for substantial losses in quality and yield in seed potatoes globally. The potato leaves may have mosaic patterns, deformations, leaf curling, and tuber necrosis depending on the strain and environmental conditions when infected by the virus [2]. PVY is one of the most destructive pathogens in potatoes and causes economic yield losses of up to 80% [3]. Although the global economic loss due to PVY is currently unknown at the moment, over USD 190 million has been reported as an annual loss in the European Union alone [4]. Thus, the early and accurate detection of PVY is crucial for implementing timely management strategies to reduce economic loss and ensure food security.

The traditional approach to detecting PVY includes serological assays and molecular techniques [5], such as enzyme-linked immunosorbent assay (ELISA) and reverse transcription polymerase chain reaction (RT-PCR) [6]. While these methods are highly sensitive and specific, the process entails extensive labor, time, significant resource requirements, and susceptibility to human errors. The procedure involves picking leaves manually from the fields, and because the virus is highly contagious [7], there is a risk of disease spread from vehicles or people walking through the field, potentially increasing pathogen spread. As agriculture moves towards precision farming, there is a growing demand for advanced, non-invasive, and high-throughput diagnostic technologies to address these limitations. Due to the increasing economic impact and the growing need for non-invasive methods, remote sensing technologies have acquired significant attention in detecting the PVY virus. Remote sensing technologies offer efficient and non-invasive approaches for monitoring crop health and various diseases, including the PVY-induced mosaic caused by PVY [8,9].

Remote sensing technologies utilizing multispectral and hyperspectral imaging (HSI) are pivotal in identifying physiological changes in plants that can indicate viral infections such as PVY [9]. These systems produce reflectance patterns that are correlated with the presence of PVY [10] and may allow for early detection before visual symptoms manifest. Vegetation indices such as the normalized difference vegetation index (NDVI) are derived from satellite imagery, and researchers have found it useful to monitor crop health and growth [11,12]. The spectral signatures change with the growth stages of the crop, and thus are useful for pre-symptomatic detection, significantly improving the timeliness of interventions and reducing the economic impact on yield losses.

HSI combines images and spectroscopy to provide spatial and detailed spectral information across a wide range of wavelengths, spanning from visible and near-infrared (Vis–NIR, 400–1000 nm) as well as the shortwave infrared (SWIR, 1000–2500 nm) regions [13]. The detailed HSI spectral information is capable of capturing subtle changes in plant physiology caused by stress or disease. These changes may not be visible to the naked eye; however, they can manifest unique spectral signatures that can distinguish between healthy and infected plants. The advent of HSI in precision agriculture has a wide variety of applications, ranging from disease detection [14,15,16,17,18], weed identification [19,20,21], nutrient assessment [22,23,24,25,26], yield predictions [27,28,29,30], and soil characterization [31,32,33,34]. HSI can be integrated with UAVs for field-scale non-destructive sampling and the real-time monitoring of PVY status [14]. The two primary challenges with HSI processing are the calibration of the data using downwelling irradiance to obtain the reflectance data and the high dimensionality of HSI data, as they typically consist of hundreds of spectral bands, resulting in a ‘data cube’ that requires advanced computational techniques to extract the meaningful information.

Advancements in machine learning (ML) and artificial intelligence have revolutionized the interpretation and capabilities of remote sensing for PVY detection. Deep learning (DL) techniques are applied to various data types such as hyperspectral, multispectral, and satellite imagery, achieving high classification accuracy in distinguishing healthy plants from PVY-infected potato plants [35,36]. ML and DL techniques excel at identifying patterns, reducing dimensionality, and building task-specific predictive models that make them suitable for precision agriculture applications. Commonly used ML-DL algorithms for HSI analysis include support vector machines (SVMs) [37,38,39,40], random forests (RFs) [41,42,43,44], and neural networks [38,41,45,46,47]. These algorithms can effectively handle high-dimensional HSI data while retaining most of the original information. Convolutional neural networks (CNNs) have particularly gained popularity in HSI analysis for their ability to extract hierarchical features, as found in HSI data [45,46,47]. A labeled dataset is required for plant disease detections, and because of the expense and scarcity of such datasets, unsupervised and semi-supervised learning approaches are gaining traction as viable alternatives.

While the integration of HSI and ML-DL has demonstrated considerable potential for PVY detections, several challenges remain. First, the spectral variability caused by environmental factors such as weather conditions, atmospheric illuminations, soil types, and plant canopy structures will not assist in generalizing the learned models. Pre-processing techniques, as discussed in Section 2.2.1, are essential for minimizing these effects and enhancing the quality of HSI data. Second, the high dimensionality of the HSI data is computationally expensive, particularly for large-scale applications. Commonly used dimensionality reduction techniques to address such issues are principal component analysis (PCA), linear discriminant analysis (LDA), and singular value decomposition (SVD) [48]. The choice of dimensionality reduction technique can significantly impact the performance of the training model. After careful evaluation, we decided to perform spatial downsampling on the HSI data and use all the spectral bands, thus reducing the pixel resolution while retaining the full set of hyperspectral bands. Third, the interpretability of ML-DL models remains a concern, especially for HSI data. It is important to understand the most relevant spectral bands carrying the most valuable information for the specific tasks. We investigated various techniques and found that only five small spectral regions of the HSI data carry the most relevant information, as reported in Section 3.5. Nevertheless, the challenges in processing HSI data and model interpretation can be conquered; PVY detection using HSI remains relatively underexplored.

The integration of unmanned aerial vehicles (UAVs) with mounted remote sensing tools has been instrumental in capturing high-resolution images and enabling the real-time monitoring of the potato fields [9]. Images collected with UAVs have high spatial resolutions and offer close-field assessments of crop health. The use of UAVs for PVY detection or any other precision agriculture management is ecological, enhances the reliability and efficiency of crop monitoring, and significantly reduces labor costs for large-acre fields. The fusion of UAV-acquired images with machine learning and deep learning techniques has enhanced the scalability of these technologies and offers the potential to build customized systems for PVY detection using configurable integrated circuits. Producers can implement quick, targeted management strategies in the infected areas of the field to reduce the spread of the virus and minimize economic loss.

Although there have been promising advancements in remote sensing techniques to detect PVY, challenges remain in the practical implementation of these systems in real producer field settings. Factors such as atmospheric conditions, lighting, variations in soil nutrients, moisture, crop canopy heterogeneity, and other confounding factors can affect remote sensing data. Furthermore, most research on PVY detection has collected datasets in a greenhouse or lab environment by taking images of leaves or individual plants [10,49,50,51]. Thus, they do not simulate the real environmental conditions of an open field. The integration of UAVs with hyperspectral cameras distinguishes our work from the other relevant research, as we collected the dataset from an actual potato field out in the open, where environmental factors pose significant challenges. We hypothesize that the presence of PVY in producer fields can be detected using hyperspectral cameras integrated with UAVs. Moving to a UAV platform limits our ability to control the environment and lighting. This work aimed to establish a proof of concept to demonstrate that we could account for these uncertainties and still enable the detection of PVY. Trials were conducted in 2023 on a limited number of diseased and healthy potato plants to establish these capabilities. While ongoing work is increasing the size of this dataset, here we present the promising results from our preliminary proof-of-concept study. We also analyze and identify the most relevant spectral regions that carry valuable information toward detecting PVY. This work is an extended report of our preliminary work based on hyperspectral remote sensing for PVY detection [14].

This research employs machine learning and deep learning techniques to detect the PVY virus in potato fields using hyperspectral cameras mounted on a UAV and identify the most relevant spectral bands in the visible and near-infrared spectral region. The paper is organized into several sections. Section 2.1 details the experimental field and the hyperspectral dataset acquired using a UAV. Section 2.2 discusses the data processing and labeling procedures. Section 2.3 provides background on the machine learning and deep learning techniques used in this work and describes the hyperparameter optimization approaches. The metrics used to evaluate the performance of the training models are explained in Section 2.5, followed by feature selection methods in Section 2.6 and the outline of the research in Section 2.7. The performance evaluation and comparisons are presented in Section 3. Finally, the paper concludes with a discussion.

2. Materials and Methods

2.1. Field and Data

The experimental field was curated to meet the needs of this research, and the statuses of individual plants were closely monitored. Hyperspectral data were collected using a visible and near-infrared (Vis-NIR) camera mounted on a drone suitable for remote sensing applications.

2.1.1. Experimental Field

For research purposes, the experimental field, as shown in Figure 1, comprises multiple uniformly arranged plots for the PVY monitoring. Note that cloud shadows are visible in the image, indicating that it is not corrected for downwelling irradiance. We collected this mapped field using Micasense Altum-PT [52], a multi-spectral camera, and used it as a reference image for the field layout. The plots were planted at the Bozeman Agriculture Research and Teaching (BART) Farm of Montana State University, Bozeman, MT. There are five blocks of potatoes, each with four rows. Each block contains four 20-foot plots with three feet of space between them and 20 plants spaced one foot apart. The plots were arranged to simulate family groups in first-generation seed potatoes. To reduce the spread of the virus across plots, the potato cultivars in each row were assigned as follows: one plot is experimental (susceptible to PVY), and the other three plots are Payette Russet potatoes (not susceptible to PVY). The Payette Russet variety is highly resistant to different strains and genetic variants of PVY [53], which significantly helps reduce the spread of the virus and contains it within the desired plots.

Some of the susceptible plots were planted with a mixture of healthy and PVY-infected seed potatoes, utilizing four different treatments to align with our research goals. In Treatment One, tubers that tested positive by immunochromatography (IC) [54] were mixed 50:50 with Generation Two Umatilla cultivar tubers from Premiere Seed Potatoes—both summer and winter lots. IC is a rapid diagnostic technique that utilizes antibodies to bind the target molecule and detect the presence of PVY in the tubers. In Treatment Two, tubers that tested negative by IC but came from plants whose sister tubers tested positive were planted. In Treatment Three, Generation Two Umatilla (UMA) cultivars from Premiere Seed Potatoes were planted. Finally, in Treatment Four, Generation Three Dark Red Norland (DRN) cultivars from Spring Creek Farms were planted.

Table 1. Treatment descriptions across the field for all five blocks of the experimental field. This is important to help facilitate the tracking of the virus spread. T# represents the treatment used in the specific plots as mentioned in each of the rows.

Treatment	Block 100	Block 200	Block 300	Block 400	Block 500
T1—50% PVY	104 D	201 B	302 C	401 A	504 C
T2—PVY + plants, PVY− Tubers	101 B	203 C	301 A	402 C	501 D
T3—Uma control	103 A	204 D	303 B	404 D	503 B
T4—DRN control	102 C	202 A	304 D	404 B	502 A

presents the various treatments applied across the different experimental plots. The plots were tracked and later analyzed in the lab using the enzyme-linked immunosorbent assay (ELISA) test [55] to evaluate the spread of the virus and establish the ground truth for the dataset. Section 2.2.2 provides details about the ELISA test procedure. The Montana Seed Potato Lab was responsible for the experimental setup and for validating the in-season PVY status of each plant.

2.1.2. UAV and Hyperspectral Camera

Unmanned aerial vehicles (UAVs), or drones in general, provide the ability to acquire high-spatial-resolution spectral data, which is a preferred tool in remote sensing. UAV systems enable disease surveys to be conducted without physically entering the field, thereby reducing the risk of spreading the virus. For this project, we used the Vector Hexacopter [56] from Vision Aerial, equipped with the Resonon Pika L hyperspectral imaging camera [57], to capture high-resolution spectral data. The drone was flown at an altitude of 15 m above ground level (AGL) at a speed of 0.5 m/s to acquire the imagery. Figure 2 shows the complete integration of the drone and the camera. The imaging setup provided a cross-track resolution of 0.5 cm, an along-track resolution of 0.2 cm, a ground swath width of 4.64 m, and a spectral resolution of 2 nm.

Vector Hexacopter UAV:The Vector Hexacopter is an American-made industrial drone designed for long-range missions and heavy lifting, with a maximum payload capacity of 11 lbs. The drone can operate for approximately 30 min with the camera attached, and its batteries can be hot-swapped if needed. The drone has six rotors and is powered by two 22,000 mAh solid-state Li-ion batteries. The GPS accuracy of the drone is typically within $\pm 1$ m, though it can vary up to $\pm 2.5$ m under certain conditions. The drone’s software includes a loitering function that holds its position in manual mode when no flight instructions are given by the pilot. Mission planning is supported for autonomous flights, and additional features include brake mode, return-to-launch, low-battery protection, and lost-link protection.

Resonon Pika L Hyperspectral Camera: The Resonon Pika L is a compact, lightweight Vis-NIR hyperspectral camera that captures 281 spectral channels in the 400–1000 nm range. The camera offers a spectral resolution of 2.7 nm and 900 spatial pixels. In this project, we used a 17 mm focal length lens with a 17.6 degree field of view. To improve GPS accuracy, an Ellipse 3D dual-antenna GPS/IMU was integrated. Additionally, to correct for light reflection in the images, a downwelling irradiance sensor, customized and calibrated by Resonon for the Pika-L camera, was employed. Spectronon ver. 3.5.4 [58], a free software provided by Resonon, supports hyperspectral data acquisition and analysis. The hyperspectral camera was calibrated prior to data collection, following the guidelines provided by Resonon [59]. Subsequently, the acquired hyperspectral images were radiometrically and reflectance-calibrated, as outlined in Section 2.2.1.

The calibration of hyperspectral sensors is critical for ensuring the validity and reproducibility of reflectance measurements, which directly influence the performance of machine learning models. In this study, we used the Resonon Pika L hyperspectral camera, which was calibrated prior to data collection following the manufacturer’s standard guidelines. Reflectance calibration was performed using a customized downwelling irradiance sensor mounted alongside the camera (see Figure 2 for the setup, and Section 2.2.1 for the calibration methods). According to personal communication with the manufacturer, the final reflectance uncertainties for this system are ±7.5–9.5% across the primary visible spectrum (400–700 nm) and ±9.5% across the near-infrared spectrum (700–900 nm). At the spectral edges, below 400 nm and above 900 nm, the accuracy degrades to ±13% and ±12%, respectively. These were calculated by combining the inverse signal-to-noise of the camera, the stated uncertainties of the radiance calibration source at the factory, and the specified uncertainty of the downwelling radiometer. These are all added in quadrature, which assumes that they are uncorrelated with each other. These tolerances are consistent with field-grade hyperspectral systems and validate the integrity of the hypercube data used in our analysis. While these uncertainties may introduce minor variability in pixel-level classification, the promising performance of the trained models indicates that the overall system accuracy is sufficient for reliable detection of PVY symptoms in field conditions.

2.1.3. Hyperspectral Dataset

The data used in this experiment were collected on 12 July 2023, within one hour of solar noon at the Montana State University BART farm. The Resonon Pika L hyperspectral camera generated datacubes consisting of 2000 line scans with 900 spatial pixels and 300 spectral channels. Note that some of the spectral channels were very noisy and did not contain any meaningful information; therefore, they were discarded. Unfortunately, the magnetometer in the HSI camera failed, and the data could not be geo-rectified or stitched together. Although the UAV’s independent guidance system ensured a properly gridded flight path, we had to manually arrange the images using their orientation and the ground control points with the checkerboard patterns to create the field layout, as seen in Figure 3. While the radiometric calibration and overall data quality remained unaffected, the primary limitation was the inability to accurately geo-locate individual plants within the field and produce a fully accurate stitched map. A Micasense Altum-PT multispectral camera [52] mounted on a DJI Matrice 600 UAV [60] was flown five days later over the same field to create the reference image used for the manual rectification and arrangement of the hyperspectral images. The geo-rectified and orthomosaic image of the field, which served as the reference, is shown in Figure 1.

Some of the images in Figure 3 appear wobbly, as the drone was unstable during lower-altitude flights at slower speeds, and the issue with the IMU magnetometer prevented correction through geo-rectification. We flew the drone at 15 m AGL and a speed of 0.5 m/s, as the field was relatively small and our focus was on capturing high-resolution images. However, a flight altitude of ≥50 m AGL and a speed of 3–5 m/s would likely have resulted in greater stability and steadier images. Upon reviewing the collected hyperspectral images, we identified 19 images containing susceptible plots that held meaningful data, which were used in this experiment. Each image contains 2000 × 900 = 1,800,000 pixels. Of these, 15 images (approximately 80%) were used for the training and validation of the machine learning models, while the remaining four images were reserved for testing the trained models.

2.2. Data Processing

The hyperspectral data captured using the Resonon Pika L camera were preprocessed using the hyperspectral analysis software, Spectronon [58]. These preprocessing steps are essential to ensure accurate and meaningful analysis.

2.2.1. Preprocessing

Figure 4 illustrates a flowchart of the preprocessing steps required to prepare the raw data for ML-DL analysis. The steps include radiometric calibration, reflectance calibration, smoothing, bad band removal, normalization, NDVI calculation, and downsampling.

Raw data: The raw data are uncalibrated and represented as digital numbers. These data cannot be compared with data from different instruments or under varying illumination conditions until they are corrected to account for those differences. Figure 5a shows an RGB (red, green, blue) representation of a sample raw image. The sample image includes healthy and infected plants, resistant varieties, and background soil, and plants from the susceptible plots that could not be isolated due to GPS uncertainty. These non-isolated plants are labeled as unknown, as shown in Figure 5c. The spectral signatures of this image vary significantly among the classes, as illustrated in Figure 6a. Plant foliage and dry soil exhibit high reflectance in the visible spectrum, while healthy plants display strong peaks in the red-edge region (680–750 nm), which is typical for vegetation. Infected plants also show a similar pattern, although the peaks are less pronounced in the near-infrared (NIR) region (>700 nm). Shadows appear as almost a flat line due to the absence of illumination, and wet soil reflects similarly to dry soil but with a lower intensity due to water absorption.

Radiance data: Radiance refers to the amount of electromagnetic energy emitted or reflected from a surface per unit area and solid angle [61]. It depends on both the direction and intensity of illumination. Radiance values also vary due to atmospheric effects such as scattering, absorption, and other interferences. Radiance data for any wavelength, $\lambda$, can be derived from the raw data (digital numbers, $DN\left(\lambda \right)$) using Equation (1).

$$L\left(\lambda \right)={\displaystyle \frac{DN\left(\lambda \right)-D{N}_{\mathrm{dark}}}{C\left(\lambda \right)\times \Delta e}}$$

(1)

where $L\left(\lambda \right)$ is the radiance, $D{N}_{\mathrm{dark}}$ is the dark correction signal of the hyperspectral sensor, and $C\left(\lambda \right)$ is the radiometric calibration coefficient of the hyperspectral sensor, and $\Delta e$ is the exposure time of the sensor. Figure 6b shows the radiance calibration across all categories. After radiance calibration, plants and soils become more distinguishable. Wet soil remains consistent with dry soil but exhibits lower intensity, while shadows appear flat with minimal reflectance across all wavelengths. Healthy plants show strong intensity in the red-edge bands and high NIR reflectance due to chlorophyll content. Infected plants exhibit a similar pattern but with reduced reflectance, indicating physiological stress.

Reflectance data: Reflectance is the ratio of the light that leaves a target to the amount of light that strikes the target [62]. It is dimensionless and has no units. To calculate reflectance, we need the downwelling spectral irradiance or simply, downwelling irradiance, which is the amount of solar illumination reaching the ground. Figure 7a shows the amount of illumination that reached the ground during the time of flight, as captured by a downwelling sensor mounted on the UAV alongside the hyperspectral camera. Let $I\left(\lambda \right)$ be the downwelling irradiance, then reflectance $R\left(\lambda \right)$ is:

$$R\left(\lambda \right)={\displaystyle \frac{\pi \times L\left(\lambda \right)}{I\left(\lambda \right)}}$$

(2)

We treat the reflecting surface as a Lambertian reflector [63], and assume that the angles of observation and illumination remain constant during data collection. These are acceptable assumptions when the observation and illumination angles are minimized. The hyperspectral camera is oriented nadir and has a relatively narrow field of view, which helps limit the variation in observation angles. Additionally, flights are conducted near solar noon to reduce the changes in illumination angle. These conditions justify the assumptions, as most non-smooth natural surfaces behave as diffuse reflectors within the minimized range of angles described above.

Reflectance calibration is crucial because it enables a comparison of the data captured at different times or with different sensors. Figure 7b displays the calibrated reflectance measurements for various categories on a scale of $R\left(\lambda \right)\times 10,000$. Oxygen A-band absorption occurs around 747–770 nm [64], which results in noticeable changes in reflectance. A diminished downwelling signal due to oxygen absorption is also observed in the same wavelength region in Figure 7a. As expected, shadows appear as flat lines with almost no reflectance. Distinct contrast is noticeable between dry soil and wet soil due to water absorption differences. While both healthy and infected plants show similar spectra in the visible range, they differ in the NIR region, where reduced reflectance in infected plants indicates physiological stress or infection.

Savitzky–Golay smoothing: To reduce the noise in the reflectance data, a Savitzky–Golay filter was fitted through the data with a third-degree polynomial and 13-point sliding window using the method of linear least squares [65]. This is a widely used filter to smooth the hyperspectral data while preserving the vital spectral features. Smoothed reflectance data using the Savitzky–Golay filter can be obtained as follows:

$$R_{\mathrm{smoothed}}\left(\lambda \right)=\sum _{j=-w}^w{c}_j·R_{\lambda +j}$$

(3)

where $R_{\mathrm{smoothed}}\left(\lambda \right)$ is the smoothed reflectance, $c_j$ are the polynomial coefficients, and w is half of the sliding window size. Figure 8a shows the smoothed reflectance data for all the categories. The noise is reduced in the data and it looks smoother. However, the key spectral features are preserved.

Bad band removal: The hyperspectral sensor provides 300 spectral channels from 387.12 nm to 1023.5 nm. Some of the spectra at the beginning and the end are extremely noisy and do not contain any meaningful information. If the bands are represented by band numbers 0–299, we cropped the wavelengths and kept 10–243 to remove the noisy bands at the beginning and the end. Due to the oxygen A-band absorption [64], we also removed a few bands (747–770 nm) from the red edge region. Finally, we ended up with 223 spectral channels for the experiment.

Normalization: We normalized the data using the maximum method, where each pixel in the image is divided by the maximum intensity across all the wavelengths in that pixel. The values are then scaled to a range from 0 to 1, effectively reducing the variability in measurement conditions and improving the machine learning analysis. Figure 8b shows the normalized reflectance spectra for a few selected pixels of the sample image. The spectral features that arise in the normalized shadow data are due to the diffuse illumination reflecting from the plant leaves, leading to spectra that are similar to those of a plant but with much lower intensity. Figure 5b shows the true color representation of the sample image after completing all the preprocessing steps, from raw data to normalized data.

NDVI and Downsampling: We calculated the normalized difference vegetation index (NDVI) of the images, which primarily measures the photosynthetic activity of plants to identify vegetation [66] and potentially remove any unwanted objects from the images. The NDVI is calculated from the red band and the near-infrared (NIR) band.

$$\mathrm{NDVI}={\displaystyle \frac{\mathrm{NIR}-\mathrm{Red}}{\mathrm{NIR}+\mathrm{Red}}}$$

(4)

We also downsampled the images to simulate a higher altitude above AGL and reduce the computational complexity during machine learning model training. If an image is $I(x,y)$, we first apply a convolution with a kernel of size $k\times k$, then subsample by taking every k-th element to downsample. For the hyperspectral cube, the third dimension (spectral bands) was not downsampled.

$$\begin{array}{c}\hfill I_{\mathrm{conv}}(x,y)=\sum _{i=0}^{k-1}\sum _{j=0}^{k-1}I(x+I,y+j)\\ \hfill I_{\mathrm{downsampled}}(x^{\prime },y^{\prime })=I_{\mathrm{conv}}(k{x}^{\prime },k{y}^{\prime })\end{array}$$

(5)

2.2.2. ELISA Test

ELISA (enzyme-linked immunosorbent assay) was performed on each plant individually to track their PVY status. ELISA is a qualitative serological assay used to detect PVY [55,67,68] in potato leaves. The leaf samples were collected from the field, and the location of the plants in the plot was recorded. The leaf samples were then extracted using a drill press in pressing trays at 4000 psi in a blotto buffer solution containing sodium azide as a preservative. This solution is used to dilute the enzyme conjugate, which is then added to the immunocapture plates that had been prepared earlier. These plates are coated with IgG antibodies and blocked to prevent non-specific binding. The samples added to the plates are then incubated overnight at 4 °C. The plates are washed four times with PBS (phosphate-buffered saline) Tween, and enzyme-conjugated antibodies are added. The plates are then incubated again at 37 °C for 1–4 h, followed by another three washes in PBS Tween. P-nitrophenyl phosphate substrate is added, and the plates are read when the controls develop a suitable color (after 30–60 min). This method ensures the detection of PVY and is crucial to creating the ground truth labels of our dataset.

2.2.3. Labeling

We labeled the hyperspectral images at the pixel level using the ELISA test results as the ground truth reference. The dataset was labeled semi-manually using the computer vision annotation tool (CVAT) [69], assigning pixel-level annotations to the background, infected, not infected, resistant, and unknown classes. A sample of the labeled data is shown in Figure 5c. Note that the images in Figure 3 were labeled using Figure 1 as the reference layout of the field.

2.3. Machine Learning and Deep Learning

The hyperspectral dataset was used to train and evaluate several standard machine learning (ML) methods, including traditional and deep learning (DL) approaches. The methods used in this work include support vector machine (SVM) [70], decision tree [71], k-nearest neighbors (KNN) [72], logistic regression [73], neural network (NN) [74], and convolutional neural network (CNN) [75]. The primary goal of this study is not to develop new ML-DL methods, but rather to rigorously evaluate the performance of well-established models in a challenging and practical application setting—the UAV-based hyperspectral detection of PVY in open-field conditions.

2.4. Hyperparameter Optimizations

Hyperparameter optimization is vital for improving the performance and generalization of machine learning and deep learning models. Model parameters are learned during the training; however, hyperparameters are set prior to the training phase to achieve optimal model performance [76]. The process of testing multiple hyperparameter settings to minimize the loss function is known as hyperparameter optimization [77]. For a set of hyperparameters $\gamma$, the loss function $L\left(f\right(x,\gamma ),y)$ must be minimized to obtain the optimized set of hyperparameters ${\gamma }^{\prime }$. Here, $f(x,\gamma )$ is the objective function, and y is the true class to be predicted.

$${\gamma }^{\prime }=\underset{\gamma }{\mathrm{argmin}}L(f(x,\gamma ),y)$$

(6)

Hyperparameters can vary depending on the model used, such as the kernel type, number of learners, learning rate, batch size, number of layers, and so on. Several optimization techniques exist for hyperparameter tuning: random search, grid search, and Bayesian optimization are among the widely used methods. Random search samples from the hyperparameter space randomly, while grid search evaluates all possible combinations of hyperparameters, which can be computationally expensive [78]. On the other hand, Bayesian optimization is a probabilistic approach based on Bayes’ theorem [79], allowing for efficient searching through the hyperparameter space. This method outperforms traditional techniques, especially in high-dimensional settings.

2.5. Evaluation Metrics

We used four metrics to evaluate the performance of all the models: overall accuracy, precision, recall, and F1-score. These metrics provide insights into the model’s predictive capability and are useful in classification tasks [80] such as predicting the presence of PVY in potatoes. Accuracy provides the ratio of correctly predicted observations to total observations. Accuracy is an intuitive metric; however, it may not be a reliable one, especially in imbalanced datasets like ours, where infected observations are much fewer than healthy ones. The metrics precision and recall provide valuable insights in such cases. Precision is the proportion of true positive observations among all predicted positive observations, and recall is the proportion of true positive observations among all actual positive observations. Precision is crucial when false positives are costly, and recall is crucial when failing to identify all the positive observations would be costly [81]. Precision and recall are also known as confidence and sensitivity, respectively. In our case, high recall is desirable as we aim to correctly identify the PVY-infected plants. The fourth metric, the F1 score, is the harmonic mean of precision and recall, which is a balanced metric that considers both false positives and false negatives. All four metrics range from 0 to 1, and a high score across all metrics indicates strong model performance.

$$\begin{array}{cc}\hfill \mathrm{accuracy}& ={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}\hfill \\ \hfill \mathrm{precision}& ={\displaystyle \frac{TP}{TP+FP}}\hfill \\ \hfill \mathrm{recall}& ={\displaystyle \frac{TP}{TP+FN}}\hfill \\ \hfill \mathrm{F}1\mathrm{score}& =2\times {\displaystyle \frac{\mathrm{precision}·\mathrm{recall}}{\mathrm{precision}+\mathrm{recall}}}\hfill \end{array}$$

(7)

where $TP$ refers to the correctly predicted positive observations, $TN$ refers to the correctly predicted negative observations, $FP$ refers to the incorrectly predicted positive observations, and $FN$ refers to the incorrectly predicted negative observations.

2.6. Band Selections

Training models with high-dimensional hyperspectral images is computationally very expensive. However, identifying the most relevant spectral bands can contribute to effective classification and reduce the computational cost of model training. A new multi-spectral camera could be developed that would be cheaper than a hyperspectral camera if the most relevant bands were identified. Training a model with hundreds of hyperspectral bands may suffer from the curse of dimensionality [82] and degrade model performance. Band selection, or feature selection techniques in general, typically fall into three main categories: filter type, wrapper type, and embedded type [83]. The filter-type selection method relies on statistical measures such as feature relevance and variance to evaluate the importance of features. This approach is model-agnostic and considered part of the data preprocessing phase [78]. The wrapper-type selection method is model-dependent and uses subsets of features to evaluate model performance. A stopping criterion, such as mean squared error or accuracy, is used to terminate the model training. The embedded-type method integrates feature selection into the model training phase, allowing the model to learn feature importance during training.

2.7. Workflow

The raw data from the hyperspectral camera need to be calibrated and cleaned to remove the atmospheric effects and make the data independent of specific camera and light conditions. First, the radiance calibration is performed using Equation (1), followed by Equation (2) for the reflectance calibration. A Savitzky–Golay smoothing filter is then applied using Equation (3). Some wavelength bands are very noisy and are removed to clean the dataset for training the ML/DL algorithms. We then normalize the dataset. NDVI is calculated for each image, using Equation (4), which is used to create a potato mask based on a threshold. We use this mask to obtain only the potato foliage, which is then downsampled using Equation (5) along with the ground-truth labels. Our data were collected at a very low AGL and over a small field, so we downsampled to simulate a higher AGL, which is more realistic for larger fields. The data are now prepared for ML/DL analyses, and we experiment with the algorithms mentioned in Section 2.3. Figure 9 shows the overall workflow of the steps performed to prepare the dataset for analysis. The preprocessing steps mentioned here refer back to Figure 4.

3. Results

3.1. Dataset

As mentioned in Section 2.1.3, we have 19 usable hyperspectral images and used 15 for model training and four for testing. Some of the images did not contain any infected plants, so four images were handpicked for testing the models. We are doing pixel-level detection, and after doing the preprocessing, cleaning, and downsampling, we ended up with 23,844 hyperspectral pixels in total for all the images. Of those, 1794 pixels were labeled as infected, and 177 infected pixels belonged to three test images. The other test image did not contain any infected pixels, which was useful for testing the model’s confidence. That leaves us with 1617 infected pixels for training and validation. To create a balanced dataset for training, we randomly picked 1617 healthy pixels from the pool of healthy pixels in the training images. A detailed breakdown of the dataset composition, including the number of pixels per class and per dataset split, is provided in

Table 2. Summary of the training and testing datasets used in this study, including the number of images, downsampled hyperspectral pixels, number of infected pixels, and class distribution. Pixel counts are grouped into resistant and susceptible plots (total pixels), with the number of infected pixels reported from the susceptible subset. The ratio of infected pixels is calculated with respect to the total number of pixels.

Dataset	Images	Total Pixels	Susceptible	Infected	Infected Ratio (%)
Training	15	72,819	19,868	1617	0.081
Testing	4	23,364	3976	177	0.045
Total	19	96,183	23,844	1794	0.075

. Figure 10 shows that the mean of all the healthy and infected pixels in the dataset. We notice some spectral differences in the visible region; however, the NIR region appears quite similar. This could be due to smaller differences in water content between the plants. Note that the spectra are normalized in this figure.

3.2. Model Architectures and Optimizations

Bayes optimization [79] was employed with a maximum of 30 objective function evaluations to fine-tune the hyperparameters of various models. For the SVM, the optimal hyperparameters identified were a box constraint of 860.726, a kernel scale of 55.033, and a Gaussian kernel function. For the decision tree model, the optimal hyperparameters included a minimum leaf size of 28, a maximum number of splits set to 3081, and a split criterion of “GDI”. In the KNN model, the optimal parameters were 13 for the number of neighbors and “correlation” as the distance metric. Logistic regression yielded optimal values with a lambda of 0.0058 and a ridge (L2) regularization term.

A feedforward neural network (FNN) was developed with an input layer matching the dataset’s 223 features. The architecture included two hidden layers containing 128 and 64 units, each followed by a ReLU activation function, and an output layer with a softmax activation function for classification. A one-dimensional convolutional neural network (1D-CNN) was also constructed. The architecture began with a convolutional layer with a kernel size of 3, followed by batch normalization and a ReLU activation function. A max-pooling layer with a stride of 2 was used to reduce dimensionality while preserving critical patterns. This was followed by a fully connected layer with 64 units and another ReLU activation. Finally, a fully connected layer with 2 units (equal to the number of classes) was added, followed by a softmax layer for predictions. Both the FNN and CNN models were trained using the Adam optimizer for 50 epochs with a mini-batch size of 32. A 5-fold cross-validation strategy was implemented for all models, where applicable, to ensure their robust performance evaluation.

Training runtimes were recorded with and without data normalization to evaluate each model’s computational efficiency. Let n denote the number of training samples, d denote the number of input features (223 in this case), and i denote the number of training iterations or epochs, where applicable. The support vector machine (SVM) exhibited the highest computational cost, requiring approximately 821.17 s with normalization and 1129.09 s without normalization, consistent with its theoretical training complexity of $\mathcal{O}(n^{2}d+n^3)$ for Gaussian kernels. The decision tree model, though inherently less complex ($\mathcal{O}(n·d·logn)$), required 24.62 s with normalization and 53.98 s without. K-Nearest Neighbors (KNN), which has negligible training complexity but incurs high prediction costs, completed training in 136.06 s with normalization and 121.06 s without. Logistic regression, characterized by linear training complexity ($\mathcal{O}(n·d·i)$), achieved a moderate runtime of 79.29 s with normalization and 120.41 s without. The feedforward neural network (FNN) and the one-dimensional convolutional neural network (1D-CNN) both leveraged the Adam optimizer and mini-batch training for efficiency. The FNN completed training in 48.39 s with normalization and 45.12 s without, while the 1D-CNN required 59.29 s and 57.45 s, respectively. These results highlight that normalization generally improved training efficiency for most models, particularly those sensitive to feature scaling, and that neural models, despite their parameter count, offered competitive runtimes due to efficient gradient-based optimization. The experiments were conducted on a machine running Ubuntu 24.02.LTS, with an Intel Core i7-9700K CPU @ 3.60 GHz, 64 GB RAM, and an NVIDIA GeForce RTX 2070 GPU.

3.3. Model Performances

Table 3. Classification results of support vector machine (SVM), decision tree (DT), K-nearest neighbors (KNN), logistic regression (LR), feedforward neural network (FNN), and convolutional neural network (CNN) on the test set. The model was trained and tested both with and without normalization on the susceptible data. The performance metrics evaluated are accuracy (Acc), precision (Prec), recall (Rec), and F1 score (F1). The best result for each of the metrics is shown in bold.

	With Normalization				Without Normalization
Model	Acc	Prec	Rec	F1	Acc	Prec	Rec	F1
SVM	0.953	0.478	0.729	0.577	0.950	0.462	0.746	0.570
DT	0.836	0.178	0.746	0.288	0.743	0.116	0.718	0.199
KNN	0.868	0.216	0.751	0.335	0.898	0.268	0.746	0.395
LR	0.951	0.470	0.746	0.576	0.936	0.391	0.768	0.518
FNN	0.860	0.214	0.797	0.337	0.958	0.547	0.294	0.382
CNN	0.872	0.234	0.831	0.366	0.955	0.495	0.599	0.542

,

Table 4. Classification results of support vector machine (SVM), decision tree (DT), K-nearest neighbors (KNN), logistic regression (LR), feedforward neural network (FNN), and convolutional neural network (CNN) on the test set. The model was trained on the susceptible data with and without normalization, and tested on the combined set of resistant and susceptible plants. The performance metrics evaluated are accuracy (Acc), precision (Prec), recall (Rec), and F1 score (F1). The best result for each of the metrics is shown in bold.

	With Normalization				Without Normalization
Model	Acc	Prec	Rec	F1	Acc	Prec	Rec	F1
SVM	0.882	0.047	0.746	0.088	0.822	0.031	0.746	0.060
DT	0.668	0.017	0.746	0.033	0.640	0.015	0.718	0.030
KNN	0.707	0.019	0.757	0.038	0.775	0.025	0.746	0.048
LR	0.857	0.037	0.723	0.071	0.749	0.023	0.768	0.044
FNN	0.768	0.026	0.808	0.050	0.969	0.078	0.294	0.124
CNN	0.715	0.022	0.831	0.042	0.932	0.065	0.599	0.118

and

Table 5. Classification results of support vector machine (SVM), decision tree (DT), K-nearest neighbors (KNN), logistic regression (LR), feedforward neural network (FNN), and convolutional neural network (CNN), evaluated on a test set solely consisting of resistant varieties with no infected samples. Models were trained on susceptible data, both with and without normalization. Since there are no infected samples in the test set, the ideal behavior is to classify all pixels as healthy. Accuracy (Acc) is reported. Precision (Prec), recall (Rec), and F1 score are omitted here, as all models yield 0 true positives (TP), making recall and F1 undefined, and precision equal to 0.

Model	Acc with Normalization	Acc Without Normalization
SVM	0.867	0.795
DT	0.633	0.618
KNN	0.673	0.750
LR	0.836	0.710
FNN	0.749	0.971
CNN	0.679	0.927

, along with Figure 11, summarize the performance of the support vector machine (SVM), decision tree (DT), K-nearest neighbors (KNN), logistic regression (LR), feedforward neural network (FNN), and convolutional neural network (CNNs), evaluated under different test conditions to assess their effectiveness in detecting PVY-infected plants from hyperspectral UAV imagery. All models were trained on pixel-level data from susceptible potato varieties and tested under different scenarios to evaluate classification accuracy, sensitivity, and generalizability. The results are reported both with and without spectral normalization.

3.3.1. Performance on Susceptible-Only Test Data

Table 3 presents the classification results when both training and testing were performed using data from susceptible varieties only. This controlled setting allows for evaluating model performance under relatively uniform conditions. Among all models, SVM achieved the best overall performance with normalization, reaching an accuracy of 0.953, a precision of 0.478, and the highest F1 score of 0.577. Logistic regression also performed well, with balanced metrics across the board. CNN stood out with the highest recall (0.831), indicating its strong sensitivity to infected pixels, a critical factor for early disease detection. Interestingly, FNN and CNN showed improved performance without normalization, achieving accuracies of 0.958 and 0.955, respectively. FNN attained the highest precision in the unnormalized case (0.547), while CNN produced its best F1 score (0.542). These results suggest that normalization has varying impacts across model types: while SVM and LR benefit from normalization, deeper models like FNN and CNN can effectively learn from the raw spectral signatures without normalization.

Figure 11 illustrates the confusion matrices for all models on the susceptible-only test set with normalization, providing further insight into the results presented in Table 3. SVM (Figure 11a) achieved a strong balance between detection and specificity, with 129 true positives and 141 false positives, aligning with its high precision (0.478) and F1 score (0.577). Logistic regression (Figure 11d) showed similar performance with 132 true positives and 149 false positives, confirming its overall reliability.

CNN (Figure 11f) identified the highest number of infected pixels (147 true positives), achieving the best recall (0.831). However, this came with a high false positive count (480), lowering its precision (0.234). FNN (Figure 11e) similarly had high sensitivity (141 TPs), but produced the highest number of false positives (519), resulting in low precision (0.214). These results suggest that, while CNN and FNN are aggressive in detecting infections, they compromise on specificity.

DT (Figure 11b) and KNN (Figure 11c) both had moderate TP counts (132–133) but much higher FP counts (609 and 483, respectively), explaining their poor precision and lower overall effectiveness. In summary, while CNN and FNN excel in recall, SVM and LR offer a more balanced approach with fewer false positives, making them preferable in scenarios where precision is critical.

3.3.2. Generalization to Mixed (Resistant + Susceptible) Test Data

To evaluate model robustness in more realistic field scenarios, the models were tested on a mixed dataset comprising both susceptible and resistant varieties (Table 4). This introduced a higher degree of class imbalance, as resistant plots only contain healthy samples. As expected, this led to a significant drop in precision across all models, reflecting increased false positive rates. However, recall remained consistently high, suggesting that models could still successfully detect infected pixels within the susceptible regions. For instance, CNN achieved the highest recall (0.831) with normalization, and LR and SVM also retained strong recall values around 0.746. FNN and CNN again performed the best overall without normalization, achieving the highest accuracies (0.969 and 0.932) and F1 scores (0.124 and 0.118, respectively). This suggests that the models are effective at detecting infection, even in the presence of distributional shifts and extreme class imbalance.

This trade-off between precision and recall highlights a key insight: the inclusion of resistant varieties, such as Payette Russet, which are not susceptible to PVY [53] and were absent from the training set, introduces spectral variability that the models were not exposed to during training. These resistant plants naturally exhibit different spectral signatures from the susceptible varieties, leading to increased false positives when the model misinterprets these unfamiliar healthy signals as signs of infection. This explains the notable drop in precision across all models. However, despite this domain shift, the models maintained strong recall, correctly identifying most of the infected pixels within the susceptible plots. This high sensitivity is encouraging for early-stage PVY detection, where the cost of missing infected plants (false negatives) is typically higher than the cost of false positives. Nevertheless, improving model generalization through training on more diverse and representative datasets, including resistant varieties, is essential for reducing false positives in real-world deployment.

3.3.3. Performance on Resistant-Only Test Data

Table 5 presents model performance on a test set solely consisting of resistant plants with no infections. In this ideal negative scenario, models should classify all pixels as healthy. Since no true positives are possible, only accuracy is reported. FNN and CNN again demonstrated strong performance, particularly without normalization, achieving accuracies of 0.971 and 0.927, respectively. SVM showed the highest accuracy with normalization (0.867), reinforcing its stability when spectral features are scaled. These results confirm that FNN and CNN are not only effective in detecting infections but also capable of minimizing false positives under healthy-only conditions, an essential characteristic for practical deployment.

3.3.4. Summary

In summary, the results across all three test scenarios demonstrate that the CNN and FNN models are strong candidates for UAV-based PVY detection. Polder et al. reported a recall of 0.88 using deep learning on ground-based imagery and full lighting control [36]. In contrast, the best practice in the industry involves visual inspectors walking the field to identify PVY, achieving a recall rate of approximately 0.70 [84].

Our best recall of 0.831 indicates that integrating UAVs and lighting compensation from downwelling sensors did not compromise model performance. Instead, it enabled remote, early detection without the risk of spreading the virus through field contact. The CNN consistently achieved high recall, making it well suited for early detection tasks, while the FNN shows high precision and accuracy, particularly in unnormalized settings. SVM and LR performed reliably under normalization but faced challenges in precision under varying test conditions. The analysis of confusion matrices in Figure 11 further confirms these trends: SVM and LR maintained a favorable balance between true positives and false positives, while CNN demonstrated strong sensitivity with a tendency to over-predict infection. The inclusion of resistant samples in testing highlighted limitations in generalizability, emphasizing the need for more diverse training data to reduce false positives and improve performance across different plant varieties and field conditions.

3.4. Prediction Analysis

The four images in the test set are images 26, 31, 39, and 43 from the dataset. The models were trained and tested only on the susceptible potato plants; thus, in our prediction analysis, we only used the susceptible potatoes from the images. Note that several PVY-resistant varieties were also used to reduce the spread of the PVY virus. The predictions of the ML-DL models are shown for test image 39 in Figure 12 and for test image 43 in Figure 13. All prediction statistics are provided in

Table 6. Prediction analysis of different ML-DL models on the test sets. The table shows the accuracy for each of the models on the images, along with the number of known infected pixels (# true infected), the number of predicted infected pixels by the trained model (# predicted infected), and the number of correctly identified infected pixels (# correct infected).

Image	Model	Accuracy	# True Infected	# Predicted Infected	# Correct Infected
26	SVM	0.942	54	42	38
	DT	0.820		80	36
	KNN	0.887		61	38
	LR	0.928		51	40
	FNN	0.901		56	38
	CNN	0.910		69	46
31	SVM	0.935	57	49	26
	DT	0.875		111	32
	KNN	0.893		92	30
	LR	0.928		55	26
	FNN	0.899		101	37
	CNN	0.892		103	35
39	SVM	0.963	66	119	65
	DT	0.885		233	64
	KNN	0.902		209	65
	LR	0.968		113	66
	FNN	0.890		230	66
	CNN	0.887		234	66
43	SVM	0.954	0	60	0
	DT	0.759		317	0
	KNN	0.807		254	0
	LR	0.953		62	0
	FNN	0.792		273	0
	CNN	0.832		221	0

.

From Figure 12b, the infections are primarily located along the lower edges of the right plot. The blank areas in this plot correspond to non-isolated plants whose PVY status is unknown due to GPS limitations. However, the entire plot is known to be susceptible to PVY. The prediction maps in subplots (c)–(h) show that all models performed well overall. Upon closer inspection, logistic regression (LR), FNN, and CNN slightly over-predict infection, labeling 113, 230, and 234 pixels as infected, respectively—despite only 66 pixels being truly infected. Notably, all three models correctly identified all 66 infected pixels, reflecting strong recall.

In contrast, Figure 13b represents a region with no known infections. Aside from the one plot shown, the surrounding plots are resistant to PVY and were excluded. Here, decision tree (DT) produced the highest number of false positives (317), followed by FNN (273), despite the absence of infection. SVM, consistent with its performance across other tests, predicted the fewest infected pixels and maintained a balanced response. Given the scattered nature of the false positives, applying a post-processing technique such as a low-pass spatial filter could help smooth predictions and reduce noise, potentially improving overall accuracy.

Table 6 provides a per-image evaluation of all the ML/DL models across four test images, reporting classification accuracy, the number of ground truth infected pixels, predicted infected pixels, and correctly identified infected pixels. In Image 26 (54 infected pixels), CNN identified the most true positives (46), while SVM achieved the highest accuracy (0.942) with fewer false positives (42 predicted infected, 38 correct). Image 31, which had 57 infected pixels, showed similar trends: FNN correctly identified the most infected pixels (37), though SVM again had the highest accuracy (0.935), with fewer over-predictions.

In Image 39, which had the most infections (66), CNN, FNN, and LR all correctly identified all 66 infected pixels. However, CNN and FNN predicted significantly more infected pixels overall (234 and 230), indicating over-prediction. LR achieved the highest accuracy (0.968) with better precision (113 predicted infected). Image 43 contained no infections; ideally, models should predict zero infected pixels. SVM and LR closely approximated this ideal, predicting only 60 and 62 infected pixels, respectively, and achieving the highest accuracies (0.954 and 0.953). In contrast, DT, KNN, FNN, and CNN produced hundreds of false positives, with DT predicting 317 infected pixels in a completely healthy image.

3.5. Feature Importance

Hyperspectral imaging has the potential to identify the PVY virus in potato fields. However, hyperspectral cameras provide densely sampled spectral data, which include redundant and less significant spectral bands. By pinpointing the most relevant bands, we can design cost-effective multispectral cameras tailored to the task of classifying the PVY virus. To identify the most relevant bands, we utilized six different feature ranking methods. These are the following: the chi-square test [85]—which evaluates the independence between each feature and target variable; minimum redundancy maximum relevance (MRMR) [86]—which balances relevance and redundancy by selecting highly correlated features with the target variable but minimally correlated with each other; the F-test [87]—which evaluates the variance ratio between classes for each feature; rank features [88]—which rank key features based on independent evaluation criteria; neighborhood component analysis (NCA) [89]—which learns feature weights by optimizing a distance-based classification objective; and the reliefF algorithm [89]—which ranks features by considering their ability to differentiate between instances that are near each other. First, we ranked the features using all six methods independently, then normalized the scores to a scale of 0 to 1. Finally, we calculated the cumulative importance of the features, as shown in bar graphs in Figure 14.

The cumulative importance highlights five key regions across the spectrum centered at 411 nm, 533 nm, 709 nm, 782 nm, and 892 nm. The first two bands (411 nm and 533 nm) correspond to the spectral features influenced by chlorophyll content in the plants. The middle band (709 nm) identifies the critical red-edge region of the spectrum. The last two bands (782 nm and 892 nm) are the NIR range and are associated with water content and plant cell structure. For reference, the mean-normalized spectra of healthy and infected plants are also illustrated in the figure.

4. Discussion

This research is a proof-of-concept artificial-intelligence-based system for the efficient detection of PVY. It is well established that a large and diverse dataset is essential for effectively generalizing any trained ML-DL models. For this initial study, only a limited number of varieties from a single growing season were available, resulting in a relatively small dataset. Consequently, the test set was also small, and while the trained models demonstrated promising performance, further testing on larger and more diverse datasets is critical to confirm the reliability and generalizability of these results. Although the models have shown strong performance, we acknowledge that their effectiveness may vary across different potato varieties and environmental conditions. To address this, we are actively collaborating with multiple seed potato producers to expand our dataset. However, labeling the data remains a significant challenge; it is time-consuming and requires expert visual inspection along with ELISA testing to confirm PVY presence. This step is crucial for establishing a reliable ground truth necessary to train and validate ML-DL models. Despite these limitations, the promising results from this preliminary study strongly indicate that the UAV-based hyperspectral detection of PVY is feasible and worth pursuing further. A more diverse and comprehensive dataset will be essential for validating our findings and enhancing the generalizability of the system.

Unfortunately, we faced a magnetometer/IMU issue with our camera, and by the time we realized the problem, it was too late in the season to conduct another flight and capture new data. This resulted in a GPS error, and we were not able to perform geo-rectification on the hyperspectral images or create an orthomosaic map. An orthomosaic map is a large, detailed, georeferenced image of the target location, created by carefully stitching the smaller images. Fortunately, after a few days, we flew another UAV with a multispectral camera, which helped us manually stitch an orthomosaic map from the distorted HSI data.

In this study, our ML/DL models were trained on a pixel-based classification task. As part of future work, it would be valuable to explore training models on small image patches that capture more spatial context, potentially leading to improved robustness. Our current dataset is highly imbalanced, with relatively few infected plants compared to healthy ones. Thus, it would be helpful to have a dataset with a greater number of infected samples. It would also be insightful to examine how model performance varies with the timing of data collection during the growing season. While time-series analysis could offer additional insights, it may be less practical for producers who prioritize early-season detection. Frequent HSI data collection throughout the growing season would benefit the research, providing deeper insights into how early PVY can be detected. The rate of the PVY spread across the field could also be estimated through regular data collection. Looking ahead, we plan to experiment with more advanced architectures as additional data become available. In particular, 3D CNNs or hybrid (2D + spectral) approaches could better exploit the spectral-spatial structure of hyperspectral data. Likewise, spectral-spatial transformer models offer promising potential for capturing complex dependencies across wavelengths and spatial dimensions. These advanced techniques may further enhance detection accuracy and generalizability in practical, field-based PVY detection scenarios.

Refer back to Figure 11 for the confusion matrices of all experimental models on the susceptible-only test set. Even in the worst case, the models were able to detect approximately 75–83% of infected pixels (recall), indicating decent sensitivity. However, the confidence in a positive prediction, measured by precision, remains relatively low, ranging from about 20% to 55% depending on the model. This means that, when a model predicts infection, it is correct only about half the time at best. In the context of seed potato production, where the acceptable disease incidence threshold is typically below 0.5%, such a performance is not yet sufficient to serve as a standalone decision-making tool for disease management. The current models lack strong predictive power, particularly in minimizing false positives, which limits their direct application in certification or roguing. That said, these results represent a promising starting point. With further development, particularly through expanded and more diverse training data, the system may reach the level of accuracy required for practical use in UAV-based field certification and early disease detection workflows.

All the models trained on this work are based on supervised learning, i.e., the model is trained with inputs that have corresponding outputs. However, it is quite expensive to create such a dataset; thus, we would be interested in exploring semi-supervised and unsupervised learning approaches. Semi-supervised learning involves training a model where some inputs are labeled and most are unlabeled. An unsupervised learning method involves training a model using only unlabeled data. Although we observed high precision and recall in our hyperparameter-optimized supervised models, it would be compelling to explore complex deep neural networks. Various dimensionality reduction techniques also need to be employed to address the large data volume characteristic of hyperspectral imagery.

5. Conclusions

Early detection of PVY in seed potatoes is crucial to minimize economic yield losses and ensure food security around the globe. In this paper, we present a proof-of-concept, machine learning-based approach using UAV-acquired hyperspectral imaging (HSI) data for detecting PVY. We collected and published a new hyperspectral PVY dataset containing ground truth labels for healthy and infected plants, contributing a valuable resource to the community. The HSI data originally contained 300 spectral bands in the range of 400–1000 nm; however, we curated the spectral bands to address noise and oxygen absorption effects, resulting in 223 usable spectral bands. While the dataset is highly imbalanced, we used a subset to create a balanced training set for the supervised ML-DL models.

We trained several ML-DL algorithms and tested them on susceptible plants only. We obtained an accuracy of 0.953 with the SVM model when the data were normalized, and 0.958 with FNN when the data were not normalized. The CNN model performed best in predicting infected plants, achieving a recall of 0.831. We also trained the models on susceptible data and tested them on a mixed dataset (resistant and susceptible combined) to evaluate the robustness of our approach. Due to the high class imbalance, we observed low precision across model performances; however, all models showed high recall. CNN was able to achieve a recall of 0.831 with normalized data. We also tested our approach on the resistant variety only (no infections), while the models were trained on the susceptible variety to introduce diversity. SVM achieved an accuracy of 0.867 with normalized data, and FNN achieved 0.971 without normalization.

Our feature importance analysis revealed that only a small subset of spectral bands carry the most relevant information for PVY classification, pointing to the feasibility of designing lightweight, cost-effective multispectral sensors. While there is room for improvement and generalization with a larger and more diverse dataset, the results are promising and show the potential of building a customized multispectral camera integrated with a UAV, focused on early PVY detection. Despite certain limitations, such as a small test set, lack of geo-rectification, and challenges in labeling via ELISA, our study highlights the viability of remote sensing and AI-based approaches in field conditions.