Optimizing corn tar spot measurement: a deep learning approach using red-green-blue imaging and Stromata Contour Detection Algorithm for leaf-level disease severity analysis

Authors: LEE, DA-YOUNG - Pohang University Of Science & Technology; NA, DONG-YEOP - Pohang University Of Science & Technology; GONGORA-CANUL, CARLOS - Technical Institute Of Mexico; JIMENEZ-BEITA, FIDEL - Purdue University; Goodwin, Stephen - Steve; CRUZ, ANDRES - Purdue University; DELP, EDWARD - Purdue University; ACOSTA, ALEX - Purdue University; LEE, JEONG-SOO - Pohang University Of Science & Technology; FALCONI, CESAR - Armed Forces University; CRUZ, CHRISTIAN - Purdue University

Submitted to: Plant Disease
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/14/2024
Publication Date: 8/19/2024
Citation: Lee, D., Na, D., Gongora-Canul, C., Jimenez-Beita, F.E., Goodwin, S.B., Cruz, A., Delp, E.J., Acosta, A.G., Lee, J., Falconi, C.E., Cruz, C.D. 2024. Optimizing corn tar spot measurement: a deep learning approach using red-green-blue imaging and Stromata Contour Detection Algorithm for leaf-level disease severity analysis. Plant Disease. https://doi.org/10.1094/pdis-12-23-2702-re.
DOI: https://doi.org/10.1094/pdis-12-23-2702-re

Interpretive Summary: Septoria tritici blotch is one of the most important diseases of wheat worldwide, yet the exact steps taken by the pathogen to cause disease are not known. To compare gene expression of the causal fungus during compatible and resistant responses, the pathogen was inoculated onto a highly susceptible wheat cultivar, two wheat cultivars with different specific resistance genes and the non-host cereal barley and samples were collected for RNA sequencing at six time points from 1-23 days after inoculation. The results showed large differences in gene expression between infection of wheat versus barley at 3 days after inoculation, and between the susceptible and two resistant interactions on wheat at ten days after inoculation. Expression of 31 possible pathogenicity-related genes occurred early during the time course. Proteins from two of these genes interacted at specific locations in tobacco that could indicate a role in gene signaling or regulation of host gene expression. These results will help plant pathologists better understand the infection process and facilitate identification of genes that are involved in interactions with wheat versus barley. This understanding may help lead to improved resistance in the future.

Technical Abstract: Detecting and quantifying tar spot of corn primarily relies on visual assessments of signs of stromata (black, raised fungal fruiting bodies) and symptoms. Here, we present Stromata Contour Detection Algorithm version 2 (SCDA v2), the improved and enhanced version of the previously developed SCDA version 1 (SCDA v1) through the integration of deep learning. Initially, SCDA v1 was utilized to detect regions of the leaf where stromata can be observed and these ‘region proposals’ (stromata and non-stromata) were then used to train a binary Convolutional Neural Network (CNN). Through the utilization of a diverse Red-Green-Blue (RGB) image dataset consisting of corn leaves at different developmental stages, tar spot severities, and canopies, our approach eliminates the need to determine the optimal decision-making input parameters for stromata detection for various datasets, which was the primary constraint of SCDA v1. The developed binary CNN classifier for tar spot stromata demonstrated accuracies of 95.3 %, 96.8 %, and 96.9 % after training, validation, and testing, respectively. Using metrics, including Lin’s concordance correlation coefficient, the performance of the SCDA v2 using images acquired under controlled and field environments also rigorously examined precision, recall, and Dice coefficients. Our findings provide evidence on the promising potential of SCDA v2 to be applicable for tar spot surveillance and phenotyping projects.