Toward an automatic way to identify red blotch infected vines from hyperspectral images acquired in the field
Context and purpose of the study – Vineyards are affected by different virus diseases, which can lower yield and affect the quality of grapes. Grapevine red blotch disease is one of them, and no curative solution exists. Once infected, a vine must be removed and replaced with a virus-free vine (aka roguing). Screening vineyards to look for symptoms can be time-consuming and needs well-trained experts. To improve this process, we conducted an experiment identifying infected vines using a hyperspectral camera in the field.
Material and methods – We monitored one vineyard in Rutherford, California, at the symptomatic stage in September and October 2020 and in August (pre-symptomatic stage), September and October 2021. More than 700 vines were sampled and analyzed through Polymerase Chain Reaction (PCR). We imaged the same vine canopies using a Senop HSC hyperspectral camera mounted on a tripod and captured 230 bands from the visible (510 nm) to the near-infrared (900 nm). We segmented leaves from the background through a U-Net neural network model and extracted the canopy signal. We tested different machine learning algorithms, Random Forest (RF), Partial Least Square (PLS), Support Vector Machine (SVM), and their multiple-model ensembles, to predict the PCR results (Infected vs. Non-infected). We evaluated and interpreted each model using mean accuracies, confusion matrices, and feature importance computation. We also computed a spectral binning and used recursive feature elimination (RFE) selection.
Results – The stacking ensemble of PLS and SVM models had the highest overall (cross-validated) accuracy of 69.5% for the entire dataset, 61% for the pre-symptomatic, and 74.5% for the symptomatic dataset. In this dataset, the model correctly classified non-infected vines with 83% accuracy and infected vines with 65% accuracy. Absolute values of PLS coefficients were the most important for reflectance at wavelengths between 550-600 nm and 750-800 nm. Concerning the permutation importance of the SVM model, the greatest values were obtained for reflectance around 600 nm, 710 nm, and 830 nm. These wavelengths are related to pigments known to be affected by red blotch. Using the RFE on the binning dataset, the overall accuracy reached 73.3% using 23 bands for the entire dataset and 76% using 30 bands for the symptomatic dataset. This study proves that hyperspectral imaging can help reduce the spread of red blotch by identifying vines that may be infected and could be rogued or molecularly analyzed if higher certainty is desired.
Issue: GiESCO 2023
1Department of Viticulture & Enology, California State University Fresno, Fresno, CA, USA
2Viticulture and Enology Research Center, California State University Fresno, Fresno, CA, USA
3Department of Mathematics, California State University Fresno, Fresno, CA, USA
4University of California, Agriculture & Natural Resources, Napa, CA, USA
5Plant Pathology and Plant-Microbe Biology, Cornell University, Geneva, NY, USA
Contact the author*
disease detection, Grapevine Red Blotch virus, hyperspectral imaging, machine learning, imaging spectroscopy