Phenotyping virus-infected grapevine leaves through hyperspectral imaging and machine learning

E. Sawyer, M. Fuchs, M.L. Cooper, B. Corrales, K. Singh, T. Frnzyan, K. Vasquez, E. Laroche-Pinel, L. Brillante
The use of hyperspectral imaging and machine learning techniques was explored for the detection and classification of two major groups of grapevine viruses, namely grapevine leafroll-associated viruses (GLRaVs) and grapevine red blotch virus (GRBV). Due to the incurable nature of viral diseases, accurate detection and subsequent removal of infected grapevines is vital to maintaining a healthy vineyard. Because many current methods of virus detection are costly, time-consuming, and unable to scale, this work seeks to develop machine learning methods to detect virus infection and classify the causal virus(es) from hyperspectral images within the near-visible light spectrum (500-700 nm). Hyperspectral imaging was used to capture about 440 leaf images from both healthy and infected red wine cultivars. Leaves were sampled twice, one month apart, in two different vineyards during fruit ripening. Leaves were separated from the petioles and imaged in a dark environment under controlled led lighting that did not emit in the infrared region. Images were preprocessed by accessing raw image radiance data, segmented from the background, and converted to reflectance values using a Spectralon standard included within each frame. Concurrently, viral infection was assessed in petioles by PCR-based assays. Reflectance data were used as input in two machine learning methods, random forest and convolutional neural networks (CNNs) and were applied to image data for prediction of virus infection status. As assessed with a 5-fold cross-validation scheme, the highest performing random forest model produced an accuracy of 68% while the CNN model achieved an accuracy of 70% (averaged across both healthy and infected leaf sample classes). While differentiation between plants infected with GLRaVs and GRBV proved to be relatively challenging, both models showed promising accuracies across classes. Further investigation is needed to improve virus prediction of infected vines in the vineyard.
Sawyer, E., Fuchs, M., Cooper, M.L., Corrales, B., Singh, K., Frnzyan, T., Vasquez, K., Laroche-Pinel, E. and Brillante, L. (2024). Phenotyping virus-infected grapevine leaves through hyperspectral imaging and machine learning. Acta Hortic. 1390, 267-272
DOI: 10.17660/ActaHortic.2024.1390.32
grapevine, virus, leaf infection, hyperspectral imaging, machine learning

Acta Horticulturae