Semester of Graduation

Fall 2025

Degree

Master of Science (MS)

Department

Plant Pathology

Document Type

Thesis

Abstract

In sweetpotato, global production exceeds 90 million tons annually, yet yields are threatened by numerous pathogens, including the potyvirus sweet potato feathery mottle virus (SPFMV). Virus infections can reduce yields by 25–40%, but surveillance is hampered by asymptomatic infections and the labor-intensive nature of detection. Hyperspectral imaging (HSI), combined with machine learning (ML), offers a promising pathway to scalable, non-destructive detection; however, studies focused on virus infection in sweetpotato remain limited.

This study demonstrates the feasibility of ML-enabled HSI for SPFMV detection under greenhouse conditions via a curated spectral library of two commercial cultivars, 'Beauregard' and 'Orleans', and the biological indicator Ipomoea setosa across early infection stages with both symptomatic and asymptomatic plants [3–20 days post inoculation (DPI)]. The infection status of each imaged plant was confirmed via reverse transcriptase quantitative polymerase chain reaction confirmed SPFMV-positive and SPFMV-negative plants. Supervised classifiers trained on standardized and transformed reflectance data (400–1000 nm) achieved study-wide accuracies ≥ 76%. Evaluation of principal components analysis and vegetation index-based transformations revealed the feasibility and effectiveness of these methods to both reduce computational load and improve model performance.

Results showed relative viral load increased over time in all three hosts, and we identified a viral-load threshold above which classification accuracy increased sharply. Importantly, successful detection was achievable as early as 3 DPI, indicating that spectral changes preceded prominent symptom expression. Feature-level analyses implicated red-edge and near-infrared regions, as well as chlorophyll-related indices, as informative markers from 3 to 20 DPI.

This work contributes a reusable, labeled spectral library tailored to SPFMV, a transparent imaging-to-model pipeline suited for greenhouse screening and pre-field studies, and quantitative evidence that HSI-ML can detect SPFMV regardless of symptomology. Collectively, these data establish a practical foundation for hyperspectral, ML-based pathogen surveillance in sweetpotato and point toward scalable remote-sensing systems for early, non-destructive virus monitoring.

Date

10-31-2025

Committee Chair

Imana Power

Available for download on Thursday, January 01, 2026

Share

COinS