Figure 10.

Classification performance of 2D-CNN (a) and 3D-CNN (b) models on a hold-out test set (n = 16 HSIs). The RF approach was the best classifier, with 71% validation accuracy for the 2D-CNN model and 75% validation accuracy for the 3D-CNN model. Both correctly classified 50% of 16 HSIs in the test set. The image subtitles indicate true and predicted classes (healthy versus GVCV-infected vines). Green image subtitles indicated a correct classification, while red subtitles indicated an incorrect classification. The row order corresponds to the measuring dates on August 7th, August 29th, September 19th, and October 8th, 2019.