Related Experiment Video
Updated: May 27, 2025
Author Spotlight: Quantification of Aflatoxins and Phytoalexins in Peanut Seeds to Identify Genetic Resistance Against Aspergillus
Published on: April 19, 2024
Automated pipeline for leaf spot severity scoring in peanuts using segmentation neural networks
Joshua Larsen1,2, Jeffrey Dunne3,4, Robert Austin3
1Department of Electrical and Computer Engineering, NC State University, 890 Oval Dr, Raleigh, NC, 27606, USA. jclarse2@ncsu.edu.
Background:
Late and early leaf spot in peanuts is a foliar disease contributing to a significant amount of lost yield globally. Peanut breeding programs frequently focus on developing disease-resistant peanut genotypes. However, existing phenotyping protocols employ subjective rating scales, performed by human raters, who determine the severity of leaf spot infection. The objective of this study was to develop an objective end-to-end pipeline that can serve to replace an expert human scorer in the field. This was accomplished using image capture protocols and segmentation neural networks that extracted lesion areas from plot-level images to determine an appropriate rating for infection severity.
Results:
The pipeline incorporated a neural network that accurately determined the infected leaf surface area and identified dead leaves from plot-level cellphone imagery. Image processing algorithms then convert these labels into quality metrics that can efficiently score these images based on infected versus non-infected area. The pipeline was evaluated using field data from plots with varying leaf spot severity, creating a dataset of thousands of images that spanned conventional visual severity scores ranging from 1-9. These predictions were based on the amount of infected leaf area and the presence of defoliated leaves in the surrounding area. We were able to demonstrate automated scoring, as compared to expert visual scoring, with a root mean square error of 0.996 visual scores, on individual images (one image per plot), and 0.800 visual scores when three images were captured of each plot.
Conclusion:
Results indicated that the model and image processing pipeline can serve as an alternative to human scoring. Eliminating human subjectivity for the scoring protocols will allow non-experts to collect scores and may enable drone-based data collection. This could reduce the time needed to obtain new lines or identify new genes responsible for leaf spot resistance in peanut.
