Author ORCID Identifier:
Date of Graduation
12-2025
Document Type
Thesis
Degree Name
Master of Science in Crop, Soil & Environmental Sciences (MS)
Degree Level
Graduate
Department
Crop, Soil & Environmental Sciences
Advisor/Mentor
Brye, Kristofor R.
Committee Member
Ashworth, Amanda J.
Second Committee Member
Poncet, Aurelie M.
Third Committee Member
Koparan, Cengiz
Fourth Committee Member
Burgos, Nilda R.
Keywords
Dimensionality reduction; Herbicide stress prediction; Hyperspectral sensing; Plant physiological processes; Random Forest; Vegetation indices
Abstract
Hyperspectral sensing enables detection of herbicide stress in weeds, but practical use requires strategies for redundancy control, rating variability management, and evaluator bias mitigation. This thesis evaluates common lambsquarters (Chenopodium album) under glyphosate exposure, aiming to identify key wavelengths, construct vegetation indices, and improve injury modeling. Greenhouse experiments used dimensionality reduction methods (principal component analysis, Relief F, Bayesian discriminant analysis) to isolate 34 unique wavelengths from about 3,200 bands. Calibration data were collected 14 days after treatment (DAT) to support wavelength selection and model development, while data from 27 DAT were used for independent validation. Class comparisons across six statistically defined injury levels highlighted pigment-sensitive regions in the green and red and structure-sensitive regions in the near infrared, consistent with physiological responses to glyphosate. Using selected bands, 45,732 two- and three-band vegetation indices were generated and evaluated with linear, quadratic, and generalized beta regressions, producing over 137,000 models. Regression approaches captured moderate to severe injury but struggled with untreated or mildly affected plants. Random Forest regression reduced residual mean absolute error by 45% and described nonlinear responses more faithfully, though generalization remained limited. To address evaluator subjectivity, a Bayesian transformation converted visual ratings into variance-adjusted values, stabilizing uncertainty at intermediate injury while preserving confidence at 0% and 100%. A three-band index at 447, 502, and 967 nm achieved strong calibration performance (R² = 0.83, MAE = 6.77, and RMSE = 8.70). However, validation at 27 DAT showed weak generalization due to temporal and scale mismatches. The key impact of this work lies in demonstrating clear, physiologically grounded spectral signatures of glyphosate injury and establishing a foundation for future leaf-level, variance-aware approaches that can enable objective, scalable, and more reliable herbicide-injury detection in precision agriculture.
Citation
Soto Valencia, M. A. (2025). Hyperspectral Indicators and Characterization of Glyphosate-Induced Stress in Common Lambsquarters (Chenopodium album L.). Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/6019
