All-in-one machine learning framework for early detection and characterization of sugar beet diseases using hyperspectral imaging

Authors: ABDALLA, ALWASEELA - Colorado State University; Todd, Olivia; PENNAM, SAI - Colorado State University; HOOPES, EMMA - Colorado State University; NGUYEN, NGA - Colorado State University; Dorn, Kevin; DAO, PHUONG - Colorado State University

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/12/2025
Publication Date: 11/13/2025
Citation: Abdalla, A., Todd, O.E., Pennam, S.V., Hoopes, E., Nguyen, N., Dorn, K.M., Dao, P. 2025. All-in-one machine learning framework for early detection and characterization of sugar beet diseases using hyperspectral imaging. Smart Agricultural Technology. 12. Article e101633. https://doi.org/10.1016/j.atech.2025.101633.
DOI: https://doi.org/10.1016/j.atech.2025.101633

Interpretive Summary: Sugar beets provide over half of the domestically produced sugar in the US annually. Many pests, pathogens, and abiotic stresses negatively impact crop quality, and result in $2.3 billion in economic losses per year. To combat these losses, plant breeders aim to improve crop resistance to these stresses. USDA-ARS scientists and collaborators at Colorado State University developed a new computational toolkit to utilize advanced imaging systems to quickly identify the genetic sources of stress resistance from crop gene banks like the USDA-ARS National Plant Germplasm System. These tools will enable plant breeders to more quickly develop improved varieties for growers and improve the resilience of US sugar beet production. This computational toolkit can also be adapted to other important US crops like corn, wheat, soybean, cotton, potato, tree fruits, and citrus.

Technical Abstract: Soilborne diseases like Fusarium oxysporum and Rhizoctonia solani significantly impact sugar beet production, causing major yield losses. Accurate disease rating and characterization enhance disease management and breeding by tracking progression, assessing resistance, and guiding control strategies. Existing diagnostic approaches often focus on limited aspects of disease assessment, addressing only one or two ICQP objectives—identification, classification, quantification, and prediction—leaving gaps in comprehensive disease management. This study proposes an all-in-one framework that integrates hyperspectral imaging and machine learning to address all ICQP objectives. Hyperspectral data were collected from 122 plants inoculated with F. oxysporum and R. solani over 30 days using a Specim IQ hyperspectral sensor (400–1000 nm, 204 bands). To ensure accurate spectral data extraction, image segmentation was performed using a trained Deeplabv3+ model. Optimal wavelengths for each ICQP task were identified using the ANOVA algorithm and fed into three machine learning classifiers, including random forest (RF), multilayer perceptron (MLP), and support vector machine (SVM). The study revealed that no single spectral region or machine learning model was universally optimal across all ICQP objectives. Chlorophyll-sensitive wavelengths (670–700 nm) were optimal for disease identification, while the near-infrared range (830–1000 nm) provided critical insights for disease type classification. RF achieved the highest accuracy (96%) in identifying healthy and infected plants and demonstrated strong performance in disease type classification. For disease quantification, MLP achieved superior results with 94% accuracy and an IoU of 88%, enabling detailed pixel-level mapping of disease severity with high confidence. This study demonstrates the importance of task-specific optimization in spectral analysis and machine learning, linking spectral features to ICQP objectives—identification, classification, quantification, and prediction—while ensuring explainable analysis and capturing subtle physiological changes in disease infection over time.