Hyperspectral Imaging and Machine Learning Detect Seed-Borne Virus in Faba Beans Without Damage

A team of researchers at the University of Saskatchewan has demonstrated that hyperspectral imaging combined with machine learning can detect pea seed-borne mosaic virus (PSbMV) in individual faba bean seeds without destroying them, achieving accuracy levels approaching 98 percent. The work, published in the journal Plant Methods, offers a potential alternative to the slow, labor-intensive, and destructive laboratory methods currently used to screen pulse crop seed lots for one of their most economically damaging seed-transmitted pathogens. Because PSbMV is carried within the seed itself and often produces no visible symptoms on the seed coat, infected seeds can silently enter the planting chain, establishing virus in fields and spreading through aphid vectors before growers realize anything is wrong.

PSbMV is a potyvirus that infects a wide range of legume species, including pea, faba bean, lentil, and chickpea. When it is transmitted through seed, the virus can reduce both yield and seed quality, and infected plants serve as inoculum sources for aphid-mediated spread within and between fields. Seed health certification programs rely on laboratory assays such as enzyme-linked immunosorbent assays, reverse transcription polymerase chain reaction, and grow-out tests to detect the pathogen. These methods are accurate but each seed tested is consumed in the process, and screening the large numbers of individual seeds needed to certify a commercial seed lot is expensive and time-consuming. The search for a rapid, non-destructive, and scalable screening technology has therefore been a long-standing goal in seed pathology.

The Saskatchewan research group, led by Simin Sabaghian of the Department of Plant Sciences with collaborators in Mechanical Engineering, hypothesized that viral infection produces measurable changes in the optical properties of seeds. Virus replication alters host metabolism, affecting moisture content, the concentration of phenolic compounds, starch composition, and tissue structure. Each of these biochemical and physical changes can modify how seed tissue reflects and absorbs light across the visible, near-infrared, and shortwave-infrared portions of the electromagnetic spectrum. Hyperspectral imaging captures this information in detail: rather than recording three broad color bands like a conventional camera, a hyperspectral system records a continuous reflectance spectrum for every pixel in an image, allowing researchers to detect subtle spectral fingerprints invisible to the human eye.

In the study, individual faba bean seeds were imaged using two hyperspectral systems covering the visible and near-infrared range together with the shortwave-infrared range, spanning approximately 400 to 1700 nanometers. Directional reflectance spectra were extracted from the calibrated images after preprocessing. The researchers used principal component analysis to explore the data and found that, although infected and healthy seeds are visually indistinguishable, they separate subtly but consistently along principal component axes. This confirmed that infection does leave a detectable optical signature, providing the statistical foundation for building classification models.

The team then trained and compared several supervised machine learning classifiers on the spectral data, testing different preprocessing strategies to maximize discrimination. The best-performing combination was a support vector machine coupled with standard normal variate preprocessing, a technique that normalizes each spectrum to correct for scattering effects and baseline shifts. This model achieved a mean fivefold cross-validation accuracy of 97.2 percent and, critically, an independent holdout accuracy of 98.3 percent, with a receiver operating characteristic area under the curve of 0.994. The near-perfect ROC-AUC indicates that the model separates infected from healthy seeds almost flawlessly across all decision thresholds, a level of performance that suggests the spectral signature of infection is robust rather than an artifact of the training data.

Recognizing that full-spectrum hyperspectral systems are expensive and data-heavy, the researchers also sought to reduce the dimensionality of the problem. Competitive adaptive reweighted sampling, a variable selection algorithm, was used to identify a compact subset of the most informative wavelengths. The selected bands were distributed across the visible, near-infrared, and shortwave-infrared regions, consistent with the idea that multiple biochemical features, including pigmentation, water absorption features, and phenolic-related absorptions, contribute to the classification signal. When the reduced set of wavelengths was used to retrain the model, discrimination remained strong, with cross-validation and holdout accuracies of 96.2 percent and 90.8 percent respectively and a holdout ROC-AUC of 0.974. The modest drop in holdout performance is the expected trade-off of model simplification, but the result demonstrates that a much smaller and cheaper sensor could plausibly deliver near-comparable screening performance in a commercial setting.

To ground the spectral findings in plant biology, the researchers quantified seed moisture content and performed proton nuclear magnetic resonance metabolite profiling on the seeds. The metabolomic analysis revealed elevated phenolic signals in virus-infected seeds, indicating that PSbMV infection triggers a measurable shift in seed secondary metabolism. Phenolic compounds are common components of plant defense responses, and their accumulation likely alters reflectance in both the visible range, where they influence coloration, and the near-infrared and shortwave-infrared ranges, where they modify absorption features. This mechanistic link between infection-induced biochemistry and optical response strengthens the case that the classifier is genuinely detecting disease physiology rather than confounding factors such as seed size, shape, or surface texture.

The practical implications extend across the pulse seed industry. A hyperspectral screening line could potentially evaluate thousands of individual seeds per hour without consuming any of them, allowing certified seed producers to identify and remove infected seeds before planting or sale. Because the seeds remain intact, positive findings could be verified with molecular assays while non-infected seed is preserved. The single-seed resolution is particularly important for seed-borne pathogens, where infection levels are often low and bulked sampling can dilute the signal; a system that scores every seed individually provides a direct estimate of infection incidence within a lot. The Saskatchewan group suggests that the approach could form the foundation for rapid, non-destructive seed health assessment tools and integrated virus management strategies in pulse crops.

Challenges remain before the technology reaches commercial deployment. The models were developed and validated on faba bean seeds under controlled laboratory conditions, and performance will need to be confirmed across different faba bean varieties, growing environments, infection severities, and seed lots to ensure the spectral models generalize beyond the training population. Instrument calibration transfer between hyperspectral cameras, the integration of imaging hardware into high-throughput sorting equipment, and regulatory acceptance of optical screening in certification programs are additional hurdles. Nevertheless, the study provides a rigorous proof of concept, supported by independent validation, wavelength reduction, and metabolomic corroboration, that a virus hiding inside a seed can be exposed by the light it reflects. As pulse production expands globally to meet growing demand for plant protein, tools that protect seed health without sacrificing seed will become increasingly valuable, and this work marks a significant step toward that goal.

Subject of Research: Non-destructive detection of pea seed-borne mosaic virus in faba bean seeds using hyperspectral imaging and machine learning

Article Title: Non-destructive hyperspectral imaging and machine learning for detection of pea seed-borne mosaic virus in faba bean seeds

Article References: Sabaghian, S., Jaliliantabar, F., Noble, S. D., Onu, G., & Prager, S. M. (2026). Non-destructive hyperspectral imaging and machine learning for detection of pea seed-borne mosaic virus in faba bean seeds. Plant Methods. https://doi.org/10.1186/s13007-026-01588-5

Image Credits: AI Generated

DOI: 10.1186/s13007-026-01588-5

Keywords: hyperspectral imaging, machine learning, PSbMV, faba bean, seed health, plant virus detection, support vector machine, non-destructive testing, seed-borne pathogens, pulse crops, NMR metabolomics, spectral classification