Drone hyperspectral imaging accurately predicts wheat stem rust disease severity

Authors: ABDULRIDHA, JAAFAR - University Of Minnesota; MIN, AN - University Of Minnesota; Rouse, Matthew; Kianian, Shahryar; ISLER, VOLKAN - University Of Minnesota; YANG, CE - University Of Minnesota

Submitted to: Meeting Abstract
Publication Type: Abstract Only
Publication Acceptance Date: 1/27/2023
Publication Date: 1/27/2023
Citation: Abdulridha, J., Min, A., Rouse, M.N., Kianian, S., Isler, V., Yang, C. 2023. Drone hyperspectral imaging accurately predicts wheat stem rust disease severity [abstract]. Unpublished.

Interpretive Summary: Wheat is grown on over 37 million acres in the United States. The rust diseases of wheat can cause devastating yield losses. Mitigation of wheat rusts includes both genetic resistance and fungicides. Accurate and early detection of wheat rust diseases using new technologies could facilitate timely application of fungicides and accurate forecasting of subsequent disease spread. In this study, we tested whether a hyperspectral camera mounted on an unmanned aerial vehicle (UAV) could accurately detect the (1) presence and (2) classes of severity of wheat stem rust in 960 small experimental field plots. We applied analyses to select wavelengths and spectral vegetation indices (SVIs): quadratic data analysis, random forest classifier (RFC), decision tree classification, and support vector machine. The wheat plots were divided into four categories based on disease severity levels validated by visual ratings: healthy (severity 0%), mildly diseases (severity 1-15%), moderately diseases (severity 16-34%), and highly diseased (severity 35% or greater). The highest overall classification accuracy (85%) was obtained through the RFC method. Using RFC, the accuracy of the SVIs ranged between 82%-96%. Binary classification of mildly diseased vs. non-diseased resulted in a 96% classification accuracy. These results suggested that hyperspectral imaging is sensitive enough to discriminate low levels of stem rust severity from the absence of disease. The SVIs selected from this study could be used to build a new inexpensive multi-spectral camera to diagnose wheat stem rust disease accurately and remotely. Accurate remote detection of disease can inform management decisions to protect U.S. wheat production from disease losses.

Technical Abstract: The rust diseases of wheat are the most consistent biotic yield constraints to wheat production. Mitigation of wheat rusts includes both genetic resistance and fungicides. Accurate and early detection of wheat rust diseases using new technologies could facilitate timely application of fungicides and accurate forecasting of subsequent disease spread. In this study, we tested whether a hyperspectral camera mounted on an unmanned aerial vehicle (UAV) could accurately detect the (1) presence and (2) classes of severity of wheat stem rust in 960 small experimental field plots. We applied analyses to select wavelengths and spectral vegetation indices (SVIs): quadratic data analysis, random forest classifier (RFC), decision tree classification, and support vector machine. The wheat plots were divided into four categories based on disease severity levels validated by visual ratings: healthy (severity 0%), mildly diseases (severity 1-15%), moderately diseases (severity 16-34%), and highly diseased (severity 35% or greater). The highest overall classification accuracy (85%) was obtained through the RFC method. Using RFC, the accuracy of the SVIs ranged between 82%-96%. The best performing SVIs out of 14 tested included Green NDVI, Photochemical Reflectance Index, Red-Edge Vegetation Stress Index, and Chlorophyll Green. Binary classification of mildly diseased vs. non-diseased resulted in a 96% classification accuracy. These results suggested that hyperspectral imaging is sensitive enough to discriminate low levels of stem rust severity from the absence of disease. The SVIs selected from this study could be used to build a new inexpensive multispectral camera to diagnose wheat stem rust disease accurately and remotely.