Skip to main content
ARS Home » Plains Area » Fargo, North Dakota » Edward T. Schafer Agricultural Research Center » Cereal Crops Improvement Research » Research » Publications at this Location » Publication #420899
Research Project: Improvement of Disease and Pest Resistance in Barley, Durum, Oat, and Wheat Using Genetics and Genomics

Location: Cereal Crops Improvement Research

Title: A flexible deep learning model to quantify signs and symptoms of cereal rusts (Puccinia spp.) from field-collected images

Author
item Carlson, Craig
item WELLER, THEODORE - St Edward'S University
item VAUGHN, HESTON - Mississippi State University
item LARSON, KARSTEN - North Dakota State University
item RAHMAN, AFRINA - North Dakota State University
item Sapkota, Suraj
item GHARAKHANI, HUSSEIN - Mississippi State University
item BABAR, ALI - University Of Florida
item HARRISON, STEPHEN - Louisiana State University
item Esvelt Klos, Kathy

Submitted to: Meeting Abstract
Publication Type: Abstract Only
Publication Acceptance Date: 12/1/2024
Publication Date: 12/1/2024
Citation: Carlson, C.H., Weller, T.E., Vaughn, H., Larson, K., Rahman, A., Sapkota, S., Gharakhani, H., Babar, A., Harrison, S., Esvelt Klos, K.L. 2024. A flexible deep learning model to quantify signs and symptoms of cereal rusts (Puccinia spp.) from field-collected images. Meeting Abstract. Poster No. PE57854.

Interpretive Summary:

Technical Abstract: The most devastating disease of wild and cultivated oat (Avena spp.) is oat crown rust, which is caused by Puccinia coronata f. sp. avenae (Pca), a highly volatile and rapidly-evolving fungal pathogen. Within the last decade, Pca was responsible for a total yield loss of 40% in the Upper Midwest. To develop stable, disease resistant cultivars, quantitative resistance is needed, as well as a standardized method to assess disease. Here, various computer vision models were evaluated and compared with human-based assessments of oat crown rust, including statistical power of genome-wide association studies (GWAS). First, a dataset of 60 field-collected oat leaf images was utilized to train disease detection and segmentation models. Each image was annotated into six classes: background, healthy-leaf, chlorosis, necrosis, white-streak, and pustule. The dataset was divided into 40 (training) : 20 (validation) sets. To enhance the robustness of the models and mitigate overfitting due to the limited dataset size, each image in the training set was augmented 50 times using a combination of techniques. State-of-the-art deep learning architectures were employed for object segmentation (U-Net, Fast R-CNN, Feature Pyramid Network, and LinkNet) and tested with four backbone networks (ResNet34, ResNet101, VGG16, and EfficientNetB0) that were selected for their proven efficacy in feature extraction in semantic segmentation tasks. The models were trained using four loss functions (Dice Loss, Jaccard Loss, Binary Focal Loss, and Categorical Cross-Entropy) and evaluated using several metrics to assess the overall and class-specific segmentation performance, including: Mean Intersection over Union, class-wise IoU, Mean Pixel Accuracy, overall Pixel Accuracy, and class-wise Pixel Accuracy. The best performing segmentation model for rust disease was U-Net + ResNet34 + Jaccard Loss. Finally, we compare human (visual) rust scores from the 2024 field season versus image-based U-Net model predictions via GWAS in an oat diversity panel.