EPIDEMIOLOGICAL CRITERIA AND DEEP CONVOLUTIONAL NEURAL NETWORKS FOR EFFICIENT SELECTION OF CULTIVARS AGAINST THE WHEAT BLAST DISEASE
Wheat blast, a disease caused by the fungal pathogen Magnaporthe oryzae pathotype Triticum, threatens global wheat production. Limited epidemiological information makes the wheat blast disease hard to contain and control, and without data, recommendations about the selection and deployment of resistant cultivars remain a challenge. Besides, cultivar selection relies on human visual disease evaluations, which can be time-consuming, labor-intensive, and subjective. We hypothesized that epidemiological parameters could be relevant to support wheat blast breeding tactics, and reliable visual estimates paired with images of wheat spike blast could be used to train deep convolutional neural networks (DCNN) models for disease severity classification. To test these hypotheses, we focused on the following objectives: 1) to evaluate ten cultivars for wheat blast resistance under field conditions using epidemiological parameters, and 2) develop accurate and reliable DCNN models to classify wheat spike blast severity under controlled conditions. For objective 1, we evaluated wheat leaf blast and wheat spike blast severity and estimated the total area under the disease progress curve (tAUDPC), final disease severity, and epidemic type. Disease progress curves of ten cultivars were fitted by the logistic (R2=0.70-0.96) and Gompertz (R2=0.64-0.94) models, pointing out to polycyclic epidemics. We concluded that tAUDPC, disease progression rate, and final disease severity could support cultivar selection for wheat blast resistance. For objective 2, wheat spike blast severity was visually estimated, and Red Green Blue images were acquired from six cultivars with various resistance levels under controlled conditions. Severity estimations were paired with each wheat spike image and created two datasets. Dataset 1 (n=5,123) included maturing and non-matured wheat spikes, and Dataset 2 (n=4,509) had only non-matured spikes. Each dataset was analyzed for inter-rater agreement between disease severity estimation of two pathologists and disease measurements of Image J, then classified by severity categories to train and test the DCNN model. The model trained with only non-matured spikes had higher precision (0.90-0.95), F-1 (0.87-0.95), and recall (0.84-0.96) than the model trained with maturing and non-matured spikes (0.75-0.95, 0.79-0.95, and 0.74-0.96, respectively). We concluded that the trained DCNN model could be used as the basis of a phenotyping tool for wheat spike blast severity classification.