Combining Convolutional Neural Network (CNN) with a Stromata Contour Detection Algorithm (SCDA v1.0) to Detect and Quantify Tar Spot of Corn

Authors: LEE, DA-YOUNG - POHANG UNIVERSITY OF SCIENCE & TECHNOLOGY; NA, DONG-NEOP - POHANG UNIVERSITY OF SCIENCE & TECHNOLOGY; Goodwin, Stephen - Steve; LEE, JEONG-SOO - POHANG UNIVERSITY OF SCIENCE & TECHNOLOGY; CRUZ, CHRISTIAN - PURDUE UNIVERSITY

Submitted to: International Congress of Plant Pathology Abstracts and Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 5/10/2023
Publication Date: 8/20/2023
Citation: Lee, D., Na, D., Goodwin, S.B., Lee, J., Cruz, C. 2023. Combining Convolutional Neural Network (CNN) with a Stromata Contour Detection Algorithm (SCDA v1.0) to Detect and Quantify Tar Spot of Corn. International Congress of Plant Pathology Abstracts and Proceedings.

Interpretive Summary: N/A

Technical Abstract: Plant disease detection and quantification depend on accurate visual observations by human experts. However, human rater subjectivity along with labor- and time-intensive disease ratings reduce the reliability and accuracy of disease detection and the throughput needed for the surveillance of emerging diseases. Tar spot of corn, originally endemic to Mexico and Latin America, is an emerging fungal disease in the United States. To accurately detect and track this disease, the Stromata Contour Detection Algorithm (SCDA) v1.0 was developed in 2021, which detects and quantifies tar spot stromata using Red-Green-Blue (RGB) images of infected maize leaves. SCDA provides accurate measurements of tar spot intensity based on optimal input parameters that require empirical analyses of numerous images within a specific dataset. Here, we combined the capabilities of the SCDA and Convolutional Neural Network (CNN) to eliminate the empirical search for optimal input parameters while automating the stromata detection process. A preliminary dataset of 9000 tar spot images generated by the SCDA was annotated by a human rater and then used as testing and validation datasets (8:2 rule) to train the binary CNN classifier. The trained CNN model achieved high accuracy (> 93 %) and minimal loss (> 0.15 %) for both testing and validation sets. Our approach will be critical for building an effective, high-throughput, and efficient detection and surveillance platform for tar spot of corn.