SPECIAL SESSION: Genomics and Phenomics to Characterize Host Resistance
Deep Learning for Image-Based Detection of Northern Leaf Blight in Maize
Michael Gore - Plant Breeding and Genetics Section, Cornell University.
In the United States and parts of Canada, northern leaf blight (NLB), a fungal foliar disease of maize caused by Setosphaeria turcica, has progressively become more severe in the past 5 years. In maize, breeding for genetic resistance is the most effective and economical approach for control of NLB. When breeding for NLB resistance, the visual scoring of gray-brown necrotic NLB lesions at multiple time points throughout the growing season is essential, but this effort is very time-consuming and prone to discrepancies between different raters. In that light, efforts are ongoing to develop a non-destructive, image-based phenotyping system that allows for rapid and accurate detection of NLB lesions (presence/absence) under field conditions. In the first iteration of this system, an image analysis pipeline that integrated convolutional neural networks (CNNs)—deep learning models used for image recognition and classification—was developed through training on a large manually annotated image data set, then deployed to detect the presence/absence of NLB lesions in diverse field images. Overall, the system attained an accuracy of 95% and greater on test set images held out from training.